diff --git a/.nojekyll b/.nojekyll
index 8eb89b354..247870d21 100644
--- a/.nojekyll
+++ b/.nojekyll
@@ -1 +1 @@
-8c1dbc37
\ No newline at end of file
+8418d357
\ No newline at end of file
diff --git a/FAQS.html b/FAQS.html
index 06dc5b422..09747ca86 100644
--- a/FAQS.html
+++ b/FAQS.html
@@ -206,6 +206,12 @@ ul.task-list li input[type="checkbox"] {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="./docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="./docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/TODO.html b/TODO.html
index fdfac5172..11bb9b17b 100644
--- a/TODO.html
+++ b/TODO.html
@@ -206,6 +206,12 @@ ul.task-list li input[type="checkbox"] {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="./docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="./docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/amd_hpc.html b/docs/amd_hpc.html
index 17d9d76e1..fe5d21f32 100644
--- a/docs/amd_hpc.html
+++ b/docs/amd_hpc.html
@@ -130,7 +130,7 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <button type="button" class="quarto-btn-toggle btn" data-bs-toggle="collapse" role="button" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">
         <i class="bi bi-layout-text-sidebar-reverse"></i>
       </button>
-        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/amd_hpc.html">AMD GPUs on HPC Systems</a></li></ol></nav>
+        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/amd_hpc.html">AMD GPUs on HPC Systems</a></li></ol></nav>
         <a class="flex-grow-1" role="navigation" data-bs-toggle="collapse" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">      
         </a>
     </div>
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
@@ -453,7 +459,7 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 <!-- main -->
 <main class="content" id="quarto-document-content">
 
-<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/amd_hpc.html">AMD GPUs on HPC Systems</a></li></ol></nav>
+<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/amd_hpc.html">AMD GPUs on HPC Systems</a></li></ol></nav>
 <div class="quarto-title">
 <h1 class="title">AMD GPUs on HPC Systems</h1>
 </div>
diff --git a/docs/batch_vs_grad.html b/docs/batch_vs_grad.html
index b03bf96eb..23617dfad 100644
--- a/docs/batch_vs_grad.html
+++ b/docs/batch_vs_grad.html
@@ -207,6 +207,12 @@ ul.task-list li input[type="checkbox"] {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/cli.html b/docs/cli.html
index 21adfdda0..c950d7058 100644
--- a/docs/cli.html
+++ b/docs/cli.html
@@ -240,6 +240,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/config.html b/docs/config.html
index e73b3d4c9..bdaef15c1 100644
--- a/docs/config.html
+++ b/docs/config.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
@@ -616,417 +622,425 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 <span id="cb1-157"><a href="#cb1-157" aria-hidden="true" tabindex="-1"></a><span class="at">      </span><span class="fu">system</span><span class="kw">:</span><span class="at"> </span><span class="kw">[</span><span class="st">"system"</span><span class="kw">]</span></span>
 <span id="cb1-158"><a href="#cb1-158" aria-hidden="true" tabindex="-1"></a><span class="at">      </span><span class="fu">tool</span><span class="kw">:</span><span class="at"> </span><span class="kw">[</span><span class="st">"tool"</span><span class="kw">]</span></span>
 <span id="cb1-159"><a href="#cb1-159" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-160"><a href="#cb1-160" aria-hidden="true" tabindex="-1"></a><span class="co">    # IMPORTANT: The following fields determine which parts of the conversation to train on.</span></span>
-<span id="cb1-161"><a href="#cb1-161" aria-hidden="true" tabindex="-1"></a><span class="co">    # Priority order: message_field_training &gt; message_field_training_detail &gt; train_on_inputs or role in roles_to_train</span></span>
-<span id="cb1-162"><a href="#cb1-162" aria-hidden="true" tabindex="-1"></a><span class="co">    # See examples at `docs/dataset-formats/conversation.qmd`</span></span>
-<span id="cb1-163"><a href="#cb1-163" aria-hidden="true" tabindex="-1"></a><span class="co">    # Note: If the below 4 fields are set to empty, defaults to training only on the last message.</span></span>
-<span id="cb1-164"><a href="#cb1-164" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-165"><a href="#cb1-165" aria-hidden="true" tabindex="-1"></a><span class="co">    # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.</span></span>
-<span id="cb1-166"><a href="#cb1-166" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">roles_to_train</span><span class="kw">:</span><span class="at"> </span><span class="kw">[</span><span class="st">"assistant"</span><span class="kw">]</span><span class="co">  # default</span></span>
-<span id="cb1-167"><a href="#cb1-167" aria-hidden="true" tabindex="-1"></a><span class="co">    # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:</span></span>
-<span id="cb1-168"><a href="#cb1-168" aria-hidden="true" tabindex="-1"></a><span class="co">    # - all: train on all EOS tokens</span></span>
-<span id="cb1-169"><a href="#cb1-169" aria-hidden="true" tabindex="-1"></a><span class="co">    # - turn (default): train on the EOS token at the end of each trainable turn</span></span>
-<span id="cb1-170"><a href="#cb1-170" aria-hidden="true" tabindex="-1"></a><span class="co">    # - last: train on the last EOS token in the conversation</span></span>
-<span id="cb1-171"><a href="#cb1-171" aria-hidden="true" tabindex="-1"></a><span class="co">    # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.</span></span>
-<span id="cb1-172"><a href="#cb1-172" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">train_on_eos</span><span class="kw">:</span><span class="at"> last</span></span>
-<span id="cb1-173"><a href="#cb1-173" aria-hidden="true" tabindex="-1"></a><span class="co">    # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.</span></span>
-<span id="cb1-174"><a href="#cb1-174" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">message_field_training</span><span class="kw">:</span><span class="at"> training</span></span>
-<span id="cb1-175"><a href="#cb1-175" aria-hidden="true" tabindex="-1"></a><span class="co">    # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.</span></span>
-<span id="cb1-176"><a href="#cb1-176" aria-hidden="true" tabindex="-1"></a><span class="co">    # The value of the key is a List[Dict] containing `begin_offset` (start character index in content), `end_offset` (end character index in content), and `train` (boolean whether to train).</span></span>
-<span id="cb1-177"><a href="#cb1-177" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">message_field_training_detail</span><span class="kw">:</span><span class="at"> train_detail</span></span>
-<span id="cb1-178"><a href="#cb1-178" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-179"><a href="#cb1-179" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-180"><a href="#cb1-180" aria-hidden="true" tabindex="-1"></a><span class="co"># If false, the datasets will not be shuffled and will keep their original order in `datasets`.</span></span>
-<span id="cb1-181"><a href="#cb1-181" aria-hidden="true" tabindex="-1"></a><span class="co"># The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.</span></span>
-<span id="cb1-182"><a href="#cb1-182" aria-hidden="true" tabindex="-1"></a><span class="fu">shuffle_merged_datasets</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
-<span id="cb1-183"><a href="#cb1-183" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-184"><a href="#cb1-184" aria-hidden="true" tabindex="-1"></a><span class="at">Deduplicates datasets and test_datasets with identical entries.</span></span>
-<span id="cb1-185"><a href="#cb1-185" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_exact_deduplication</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
-<span id="cb1-186"><a href="#cb1-186" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-187"><a href="#cb1-187" aria-hidden="true" tabindex="-1"></a><span class="co"># A list of one or more datasets to eval the model with.</span></span>
-<span id="cb1-188"><a href="#cb1-188" aria-hidden="true" tabindex="-1"></a><span class="co"># You can use either test_datasets, or val_set_size, but not both.</span></span>
-<span id="cb1-189"><a href="#cb1-189" aria-hidden="true" tabindex="-1"></a><span class="fu">test_datasets</span><span class="kw">:</span></span>
-<span id="cb1-190"><a href="#cb1-190" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> </span><span class="fu">path</span><span class="kw">:</span><span class="at"> /workspace/data/eval.jsonl</span></span>
-<span id="cb1-191"><a href="#cb1-191" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">ds_type</span><span class="kw">:</span><span class="at"> json</span></span>
-<span id="cb1-192"><a href="#cb1-192" aria-hidden="true" tabindex="-1"></a><span class="co">    # You need to specify a split. For "json" datasets the default split is called "train".</span></span>
-<span id="cb1-193"><a href="#cb1-193" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">split</span><span class="kw">:</span><span class="at"> train</span></span>
-<span id="cb1-194"><a href="#cb1-194" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">type</span><span class="kw">:</span><span class="at"> completion</span></span>
-<span id="cb1-195"><a href="#cb1-195" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">data_files</span><span class="kw">:</span></span>
-<span id="cb1-196"><a href="#cb1-196" aria-hidden="true" tabindex="-1"></a><span class="at">      </span><span class="kw">-</span><span class="at"> /workspace/data/eval.jsonl</span></span>
-<span id="cb1-197"><a href="#cb1-197" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-198"><a href="#cb1-198" aria-hidden="true" tabindex="-1"></a><span class="co"># use RL training: 'dpo', 'ipo', 'kto'</span></span>
-<span id="cb1-199"><a href="#cb1-199" aria-hidden="true" tabindex="-1"></a><span class="fu">rl</span><span class="kw">:</span></span>
-<span id="cb1-200"><a href="#cb1-200" aria-hidden="true" tabindex="-1"></a><span class="co"># whether to perform weighting if doing DPO training. Boolean.</span></span>
-<span id="cb1-201"><a href="#cb1-201" aria-hidden="true" tabindex="-1"></a><span class="fu">dpo_use_weighting</span><span class="kw">:</span></span>
-<span id="cb1-202"><a href="#cb1-202" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-203"><a href="#cb1-203" aria-hidden="true" tabindex="-1"></a><span class="co"># reward modelling: `True` or `False`</span></span>
-<span id="cb1-204"><a href="#cb1-204" aria-hidden="true" tabindex="-1"></a><span class="fu">reward_model</span><span class="kw">:</span></span>
-<span id="cb1-205"><a href="#cb1-205" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-206"><a href="#cb1-206" aria-hidden="true" tabindex="-1"></a><span class="co"># process reward modelling: `True` or `False`</span></span>
-<span id="cb1-207"><a href="#cb1-207" aria-hidden="true" tabindex="-1"></a><span class="fu">process_reward_model</span><span class="kw">:</span></span>
+<span id="cb1-160"><a href="#cb1-160" aria-hidden="true" tabindex="-1"></a><span class="co">    # Optional[bool]. Whether to drop the system turn from the dataset. Only works with chat_template.</span></span>
+<span id="cb1-161"><a href="#cb1-161" aria-hidden="true" tabindex="-1"></a><span class="co">    # This does not drop the default system message from chat_template if it exists. If you wish to,</span></span>
+<span id="cb1-162"><a href="#cb1-162" aria-hidden="true" tabindex="-1"></a><span class="co">    # we recommend using a custom jinja template with the default system message removed or</span></span>
+<span id="cb1-163"><a href="#cb1-163" aria-hidden="true" tabindex="-1"></a><span class="co">    # adding a system turn with empty content.</span></span>
+<span id="cb1-164"><a href="#cb1-164" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">drop_system_message</span><span class="kw">:</span></span>
+<span id="cb1-165"><a href="#cb1-165" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-166"><a href="#cb1-166" aria-hidden="true" tabindex="-1"></a><span class="co">    # IMPORTANT: The following fields determine which parts of the conversation to train on.</span></span>
+<span id="cb1-167"><a href="#cb1-167" aria-hidden="true" tabindex="-1"></a><span class="co">    # Priority order: message_field_training &gt; message_field_training_detail &gt; train_on_inputs or role in roles_to_train</span></span>
+<span id="cb1-168"><a href="#cb1-168" aria-hidden="true" tabindex="-1"></a><span class="co">    # See examples at `docs/dataset-formats/conversation.qmd`</span></span>
+<span id="cb1-169"><a href="#cb1-169" aria-hidden="true" tabindex="-1"></a><span class="co">    # Note: If the below 4 fields are set to empty, defaults to training only on the last message.</span></span>
+<span id="cb1-170"><a href="#cb1-170" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-171"><a href="#cb1-171" aria-hidden="true" tabindex="-1"></a><span class="co">    # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.</span></span>
+<span id="cb1-172"><a href="#cb1-172" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">roles_to_train</span><span class="kw">:</span><span class="at"> </span><span class="kw">[</span><span class="st">"assistant"</span><span class="kw">]</span><span class="co">  # default</span></span>
+<span id="cb1-173"><a href="#cb1-173" aria-hidden="true" tabindex="-1"></a><span class="co">    # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:</span></span>
+<span id="cb1-174"><a href="#cb1-174" aria-hidden="true" tabindex="-1"></a><span class="co">    # - all: train on all EOS tokens</span></span>
+<span id="cb1-175"><a href="#cb1-175" aria-hidden="true" tabindex="-1"></a><span class="co">    # - turn (default): train on the EOS token at the end of each trainable turn</span></span>
+<span id="cb1-176"><a href="#cb1-176" aria-hidden="true" tabindex="-1"></a><span class="co">    # - last: train on the last EOS token in the conversation</span></span>
+<span id="cb1-177"><a href="#cb1-177" aria-hidden="true" tabindex="-1"></a><span class="co">    # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.</span></span>
+<span id="cb1-178"><a href="#cb1-178" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">train_on_eos</span><span class="kw">:</span><span class="at"> last</span></span>
+<span id="cb1-179"><a href="#cb1-179" aria-hidden="true" tabindex="-1"></a><span class="co">    # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.</span></span>
+<span id="cb1-180"><a href="#cb1-180" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">message_field_training</span><span class="kw">:</span><span class="at"> training</span></span>
+<span id="cb1-181"><a href="#cb1-181" aria-hidden="true" tabindex="-1"></a><span class="co">    # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.</span></span>
+<span id="cb1-182"><a href="#cb1-182" aria-hidden="true" tabindex="-1"></a><span class="co">    # The value of the key is a List[Dict] containing `begin_offset` (start character index in content), `end_offset` (end character index in content), and `train` (boolean whether to train).</span></span>
+<span id="cb1-183"><a href="#cb1-183" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">message_field_training_detail</span><span class="kw">:</span><span class="at"> train_detail</span></span>
+<span id="cb1-184"><a href="#cb1-184" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-185"><a href="#cb1-185" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-186"><a href="#cb1-186" aria-hidden="true" tabindex="-1"></a><span class="co"># If false, the datasets will not be shuffled and will keep their original order in `datasets`.</span></span>
+<span id="cb1-187"><a href="#cb1-187" aria-hidden="true" tabindex="-1"></a><span class="co"># The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.</span></span>
+<span id="cb1-188"><a href="#cb1-188" aria-hidden="true" tabindex="-1"></a><span class="fu">shuffle_merged_datasets</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
+<span id="cb1-189"><a href="#cb1-189" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-190"><a href="#cb1-190" aria-hidden="true" tabindex="-1"></a><span class="at">Deduplicates datasets and test_datasets with identical entries.</span></span>
+<span id="cb1-191"><a href="#cb1-191" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_exact_deduplication</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
+<span id="cb1-192"><a href="#cb1-192" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-193"><a href="#cb1-193" aria-hidden="true" tabindex="-1"></a><span class="co"># A list of one or more datasets to eval the model with.</span></span>
+<span id="cb1-194"><a href="#cb1-194" aria-hidden="true" tabindex="-1"></a><span class="co"># You can use either test_datasets, or val_set_size, but not both.</span></span>
+<span id="cb1-195"><a href="#cb1-195" aria-hidden="true" tabindex="-1"></a><span class="fu">test_datasets</span><span class="kw">:</span></span>
+<span id="cb1-196"><a href="#cb1-196" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> </span><span class="fu">path</span><span class="kw">:</span><span class="at"> /workspace/data/eval.jsonl</span></span>
+<span id="cb1-197"><a href="#cb1-197" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">ds_type</span><span class="kw">:</span><span class="at"> json</span></span>
+<span id="cb1-198"><a href="#cb1-198" aria-hidden="true" tabindex="-1"></a><span class="co">    # You need to specify a split. For "json" datasets the default split is called "train".</span></span>
+<span id="cb1-199"><a href="#cb1-199" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">split</span><span class="kw">:</span><span class="at"> train</span></span>
+<span id="cb1-200"><a href="#cb1-200" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">type</span><span class="kw">:</span><span class="at"> completion</span></span>
+<span id="cb1-201"><a href="#cb1-201" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">data_files</span><span class="kw">:</span></span>
+<span id="cb1-202"><a href="#cb1-202" aria-hidden="true" tabindex="-1"></a><span class="at">      </span><span class="kw">-</span><span class="at"> /workspace/data/eval.jsonl</span></span>
+<span id="cb1-203"><a href="#cb1-203" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-204"><a href="#cb1-204" aria-hidden="true" tabindex="-1"></a><span class="co"># use RL training: 'dpo', 'ipo', 'kto'</span></span>
+<span id="cb1-205"><a href="#cb1-205" aria-hidden="true" tabindex="-1"></a><span class="fu">rl</span><span class="kw">:</span></span>
+<span id="cb1-206"><a href="#cb1-206" aria-hidden="true" tabindex="-1"></a><span class="co"># whether to perform weighting if doing DPO training. Boolean.</span></span>
+<span id="cb1-207"><a href="#cb1-207" aria-hidden="true" tabindex="-1"></a><span class="fu">dpo_use_weighting</span><span class="kw">:</span></span>
 <span id="cb1-208"><a href="#cb1-208" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-209"><a href="#cb1-209" aria-hidden="true" tabindex="-1"></a><span class="co"># The name of the chat template to use for training, following values are supported:</span></span>
-<span id="cb1-210"><a href="#cb1-210" aria-hidden="true" tabindex="-1"></a><span class="co"># - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.</span></span>
-<span id="cb1-211"><a href="#cb1-211" aria-hidden="true" tabindex="-1"></a><span class="co"># - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py</span></span>
-<span id="cb1-212"><a href="#cb1-212" aria-hidden="true" tabindex="-1"></a><span class="co"># - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to. E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not available in the tokenizer.</span></span>
-<span id="cb1-213"><a href="#cb1-213" aria-hidden="true" tabindex="-1"></a><span class="co"># - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.</span></span>
-<span id="cb1-214"><a href="#cb1-214" aria-hidden="true" tabindex="-1"></a><span class="co"># The selected chat template will be saved to the tokenizer_config.json for easier inferencing</span></span>
-<span id="cb1-215"><a href="#cb1-215" aria-hidden="true" tabindex="-1"></a><span class="co"># Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template.</span></span>
-<span id="cb1-216"><a href="#cb1-216" aria-hidden="true" tabindex="-1"></a><span class="fu">chat_template</span><span class="kw">:</span><span class="at"> tokenizer_default</span></span>
-<span id="cb1-217"><a href="#cb1-217" aria-hidden="true" tabindex="-1"></a><span class="co"># custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null.</span></span>
-<span id="cb1-218"><a href="#cb1-218" aria-hidden="true" tabindex="-1"></a><span class="fu">chat_template_jinja</span><span class="kw">:</span><span class="at"> </span><span class="ch">null</span></span>
-<span id="cb1-219"><a href="#cb1-219" aria-hidden="true" tabindex="-1"></a><span class="co"># Changes the default system message</span></span>
-<span id="cb1-220"><a href="#cb1-220" aria-hidden="true" tabindex="-1"></a><span class="fu">default_system_message</span><span class="kw">:</span><span class="at"> You are a helpful assistant. Please give a long and detailed answer.</span><span class="co"> # Currently only supports chatml.</span></span>
-<span id="cb1-221"><a href="#cb1-221" aria-hidden="true" tabindex="-1"></a><span class="co"># Axolotl attempts to save the dataset as an arrow after packing the data together so</span></span>
-<span id="cb1-222"><a href="#cb1-222" aria-hidden="true" tabindex="-1"></a><span class="co"># subsequent training attempts load faster, relative path</span></span>
-<span id="cb1-223"><a href="#cb1-223" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_prepared_path</span><span class="kw">:</span><span class="at"> data/last_run_prepared</span></span>
-<span id="cb1-224"><a href="#cb1-224" aria-hidden="true" tabindex="-1"></a><span class="co"># Push prepared dataset to hub</span></span>
-<span id="cb1-225"><a href="#cb1-225" aria-hidden="true" tabindex="-1"></a><span class="fu">push_dataset_to_hub</span><span class="kw">:</span><span class="co"> # repo path</span></span>
-<span id="cb1-226"><a href="#cb1-226" aria-hidden="true" tabindex="-1"></a><span class="co"># The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`</span></span>
-<span id="cb1-227"><a href="#cb1-227" aria-hidden="true" tabindex="-1"></a><span class="co"># if not set.</span></span>
-<span id="cb1-228"><a href="#cb1-228" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_processes</span><span class="kw">:</span><span class="co"> # defaults to os.cpu_count() if not set</span></span>
-<span id="cb1-229"><a href="#cb1-229" aria-hidden="true" tabindex="-1"></a><span class="co"># Keep dataset in memory while preprocessing</span></span>
-<span id="cb1-230"><a href="#cb1-230" aria-hidden="true" tabindex="-1"></a><span class="co"># Only needed if cached dataset is taking too much storage</span></span>
-<span id="cb1-231"><a href="#cb1-231" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_keep_in_memory</span><span class="kw">:</span></span>
-<span id="cb1-232"><a href="#cb1-232" aria-hidden="true" tabindex="-1"></a><span class="co"># push checkpoints to hub</span></span>
-<span id="cb1-233"><a href="#cb1-233" aria-hidden="true" tabindex="-1"></a><span class="fu">hub_model_id</span><span class="kw">:</span><span class="co"> # private repo path to push finetuned model</span></span>
-<span id="cb1-234"><a href="#cb1-234" aria-hidden="true" tabindex="-1"></a><span class="co"># how to push checkpoints to hub</span></span>
-<span id="cb1-235"><a href="#cb1-235" aria-hidden="true" tabindex="-1"></a><span class="co"># https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy</span></span>
-<span id="cb1-236"><a href="#cb1-236" aria-hidden="true" tabindex="-1"></a><span class="fu">hub_strategy</span><span class="kw">:</span></span>
-<span id="cb1-237"><a href="#cb1-237" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets</span></span>
-<span id="cb1-238"><a href="#cb1-238" aria-hidden="true" tabindex="-1"></a><span class="co"># Required to be true when used in combination with `push_dataset_to_hub`</span></span>
-<span id="cb1-239"><a href="#cb1-239" aria-hidden="true" tabindex="-1"></a><span class="fu">hf_use_auth_token</span><span class="kw">:</span><span class="co"> # boolean</span></span>
-<span id="cb1-240"><a href="#cb1-240" aria-hidden="true" tabindex="-1"></a><span class="co"># How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.</span></span>
-<span id="cb1-241"><a href="#cb1-241" aria-hidden="true" tabindex="-1"></a><span class="fu">val_set_size</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.04</span></span>
-<span id="cb1-242"><a href="#cb1-242" aria-hidden="true" tabindex="-1"></a><span class="co"># Num shards for whole dataset</span></span>
-<span id="cb1-243"><a href="#cb1-243" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_shard_num</span><span class="kw">:</span></span>
-<span id="cb1-244"><a href="#cb1-244" aria-hidden="true" tabindex="-1"></a><span class="co"># Index of shard to use for whole dataset</span></span>
-<span id="cb1-245"><a href="#cb1-245" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_shard_idx</span><span class="kw">:</span></span>
-<span id="cb1-246"><a href="#cb1-246" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-247"><a href="#cb1-247" aria-hidden="true" tabindex="-1"></a><span class="co"># The maximum length of an input to train with, this should typically be less than 2048</span></span>
-<span id="cb1-248"><a href="#cb1-248" aria-hidden="true" tabindex="-1"></a><span class="co"># as most models have a token/context limit of 2048</span></span>
-<span id="cb1-249"><a href="#cb1-249" aria-hidden="true" tabindex="-1"></a><span class="fu">sequence_len</span><span class="kw">:</span><span class="at"> </span><span class="dv">2048</span></span>
-<span id="cb1-250"><a href="#cb1-250" aria-hidden="true" tabindex="-1"></a><span class="co"># Pad inputs so each step uses constant sized buffers</span></span>
-<span id="cb1-251"><a href="#cb1-251" aria-hidden="true" tabindex="-1"></a><span class="co"># This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently</span></span>
-<span id="cb1-252"><a href="#cb1-252" aria-hidden="true" tabindex="-1"></a><span class="fu">pad_to_sequence_len</span><span class="kw">:</span></span>
-<span id="cb1-253"><a href="#cb1-253" aria-hidden="true" tabindex="-1"></a><span class="co"># Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'</span></span>
-<span id="cb1-254"><a href="#cb1-254" aria-hidden="true" tabindex="-1"></a><span class="fu">sample_packing</span><span class="kw">:</span></span>
-<span id="cb1-255"><a href="#cb1-255" aria-hidden="true" tabindex="-1"></a><span class="co"># Set to 'false' if getting errors during eval with sample_packing on.</span></span>
-<span id="cb1-256"><a href="#cb1-256" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_sample_packing</span><span class="kw">:</span></span>
-<span id="cb1-257"><a href="#cb1-257" aria-hidden="true" tabindex="-1"></a><span class="co"># You can set these packing optimizations AFTER starting a training at least once.</span></span>
-<span id="cb1-258"><a href="#cb1-258" aria-hidden="true" tabindex="-1"></a><span class="co"># The trainer will provide recommended values for these values.</span></span>
-<span id="cb1-259"><a href="#cb1-259" aria-hidden="true" tabindex="-1"></a><span class="fu">sample_packing_eff_est</span><span class="kw">:</span></span>
-<span id="cb1-260"><a href="#cb1-260" aria-hidden="true" tabindex="-1"></a><span class="fu">total_num_tokens</span><span class="kw">:</span></span>
-<span id="cb1-261"><a href="#cb1-261" aria-hidden="true" tabindex="-1"></a><span class="co"># Increasing the following values helps with packing, but usually only slightly (&lt;%1.)</span></span>
-<span id="cb1-262"><a href="#cb1-262" aria-hidden="true" tabindex="-1"></a><span class="co"># The number of samples packed at a time.</span></span>
-<span id="cb1-263"><a href="#cb1-263" aria-hidden="true" tabindex="-1"></a><span class="fu">sample_packing_group_size</span><span class="kw">:</span><span class="at"> </span><span class="dv">100000</span></span>
-<span id="cb1-264"><a href="#cb1-264" aria-hidden="true" tabindex="-1"></a><span class="co"># The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.</span></span>
-<span id="cb1-265"><a href="#cb1-265" aria-hidden="true" tabindex="-1"></a><span class="fu">sample_packing_bin_size</span><span class="kw">:</span><span class="at"> </span><span class="dv">200</span></span>
-<span id="cb1-266"><a href="#cb1-266" aria-hidden="true" tabindex="-1"></a><span class="co"># whether to concatenate samples during pretraining</span></span>
-<span id="cb1-267"><a href="#cb1-267" aria-hidden="true" tabindex="-1"></a><span class="fu">pretraining_sample_concatenation</span><span class="kw">:</span></span>
-<span id="cb1-268"><a href="#cb1-268" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-269"><a href="#cb1-269" aria-hidden="true" tabindex="-1"></a><span class="co"># Use batch flattening for speedups when not using sample_packing</span></span>
-<span id="cb1-270"><a href="#cb1-270" aria-hidden="true" tabindex="-1"></a><span class="fu">batch_flattening</span><span class="kw">:</span></span>
-<span id="cb1-271"><a href="#cb1-271" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-272"><a href="#cb1-272" aria-hidden="true" tabindex="-1"></a><span class="co"># Passed through to transformers when loading the model when launched without accelerate</span></span>
-<span id="cb1-273"><a href="#cb1-273" aria-hidden="true" tabindex="-1"></a><span class="co"># Use `sequential` when training w/ model parallelism to limit memory</span></span>
-<span id="cb1-274"><a href="#cb1-274" aria-hidden="true" tabindex="-1"></a><span class="fu">device_map</span><span class="kw">:</span></span>
-<span id="cb1-275"><a href="#cb1-275" aria-hidden="true" tabindex="-1"></a><span class="co"># Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model.</span></span>
-<span id="cb1-276"><a href="#cb1-276" aria-hidden="true" tabindex="-1"></a><span class="fu">max_memory</span><span class="kw">:</span></span>
+<span id="cb1-209"><a href="#cb1-209" aria-hidden="true" tabindex="-1"></a><span class="co"># reward modelling: `True` or `False`</span></span>
+<span id="cb1-210"><a href="#cb1-210" aria-hidden="true" tabindex="-1"></a><span class="fu">reward_model</span><span class="kw">:</span></span>
+<span id="cb1-211"><a href="#cb1-211" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-212"><a href="#cb1-212" aria-hidden="true" tabindex="-1"></a><span class="co"># process reward modelling: `True` or `False`</span></span>
+<span id="cb1-213"><a href="#cb1-213" aria-hidden="true" tabindex="-1"></a><span class="fu">process_reward_model</span><span class="kw">:</span></span>
+<span id="cb1-214"><a href="#cb1-214" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-215"><a href="#cb1-215" aria-hidden="true" tabindex="-1"></a><span class="co"># The name of the chat template to use for training, following values are supported:</span></span>
+<span id="cb1-216"><a href="#cb1-216" aria-hidden="true" tabindex="-1"></a><span class="co"># - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.</span></span>
+<span id="cb1-217"><a href="#cb1-217" aria-hidden="true" tabindex="-1"></a><span class="co"># - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py</span></span>
+<span id="cb1-218"><a href="#cb1-218" aria-hidden="true" tabindex="-1"></a><span class="co"># - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to. E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not available in the tokenizer.</span></span>
+<span id="cb1-219"><a href="#cb1-219" aria-hidden="true" tabindex="-1"></a><span class="co"># - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.</span></span>
+<span id="cb1-220"><a href="#cb1-220" aria-hidden="true" tabindex="-1"></a><span class="co"># The selected chat template will be saved to the tokenizer_config.json for easier inferencing</span></span>
+<span id="cb1-221"><a href="#cb1-221" aria-hidden="true" tabindex="-1"></a><span class="co"># Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template.</span></span>
+<span id="cb1-222"><a href="#cb1-222" aria-hidden="true" tabindex="-1"></a><span class="fu">chat_template</span><span class="kw">:</span><span class="at"> tokenizer_default</span></span>
+<span id="cb1-223"><a href="#cb1-223" aria-hidden="true" tabindex="-1"></a><span class="co"># custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null.</span></span>
+<span id="cb1-224"><a href="#cb1-224" aria-hidden="true" tabindex="-1"></a><span class="fu">chat_template_jinja</span><span class="kw">:</span><span class="at"> </span><span class="ch">null</span></span>
+<span id="cb1-225"><a href="#cb1-225" aria-hidden="true" tabindex="-1"></a><span class="co"># Changes the default system message. Currently only supports chatml.</span></span>
+<span id="cb1-226"><a href="#cb1-226" aria-hidden="true" tabindex="-1"></a><span class="fu">default_system_message</span><span class="kw">:</span><span class="at"> You are a helpful assistant. Please give a long and detailed answer.</span></span>
+<span id="cb1-227"><a href="#cb1-227" aria-hidden="true" tabindex="-1"></a><span class="co"># Axolotl attempts to save the dataset as an arrow after packing the data together so</span></span>
+<span id="cb1-228"><a href="#cb1-228" aria-hidden="true" tabindex="-1"></a><span class="co"># subsequent training attempts load faster, relative path</span></span>
+<span id="cb1-229"><a href="#cb1-229" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_prepared_path</span><span class="kw">:</span><span class="at"> data/last_run_prepared</span></span>
+<span id="cb1-230"><a href="#cb1-230" aria-hidden="true" tabindex="-1"></a><span class="co"># Push prepared dataset to hub</span></span>
+<span id="cb1-231"><a href="#cb1-231" aria-hidden="true" tabindex="-1"></a><span class="fu">push_dataset_to_hub</span><span class="kw">:</span><span class="co"> # repo path</span></span>
+<span id="cb1-232"><a href="#cb1-232" aria-hidden="true" tabindex="-1"></a><span class="co"># The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`</span></span>
+<span id="cb1-233"><a href="#cb1-233" aria-hidden="true" tabindex="-1"></a><span class="co"># if not set.</span></span>
+<span id="cb1-234"><a href="#cb1-234" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_processes</span><span class="kw">:</span><span class="co"> # defaults to os.cpu_count() if not set</span></span>
+<span id="cb1-235"><a href="#cb1-235" aria-hidden="true" tabindex="-1"></a><span class="co"># Keep dataset in memory while preprocessing</span></span>
+<span id="cb1-236"><a href="#cb1-236" aria-hidden="true" tabindex="-1"></a><span class="co"># Only needed if cached dataset is taking too much storage</span></span>
+<span id="cb1-237"><a href="#cb1-237" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_keep_in_memory</span><span class="kw">:</span></span>
+<span id="cb1-238"><a href="#cb1-238" aria-hidden="true" tabindex="-1"></a><span class="co"># push checkpoints to hub</span></span>
+<span id="cb1-239"><a href="#cb1-239" aria-hidden="true" tabindex="-1"></a><span class="fu">hub_model_id</span><span class="kw">:</span><span class="co"> # private repo path to push finetuned model</span></span>
+<span id="cb1-240"><a href="#cb1-240" aria-hidden="true" tabindex="-1"></a><span class="co"># how to push checkpoints to hub</span></span>
+<span id="cb1-241"><a href="#cb1-241" aria-hidden="true" tabindex="-1"></a><span class="co"># https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy</span></span>
+<span id="cb1-242"><a href="#cb1-242" aria-hidden="true" tabindex="-1"></a><span class="fu">hub_strategy</span><span class="kw">:</span></span>
+<span id="cb1-243"><a href="#cb1-243" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets</span></span>
+<span id="cb1-244"><a href="#cb1-244" aria-hidden="true" tabindex="-1"></a><span class="co"># Required to be true when used in combination with `push_dataset_to_hub`</span></span>
+<span id="cb1-245"><a href="#cb1-245" aria-hidden="true" tabindex="-1"></a><span class="fu">hf_use_auth_token</span><span class="kw">:</span><span class="co"> # boolean</span></span>
+<span id="cb1-246"><a href="#cb1-246" aria-hidden="true" tabindex="-1"></a><span class="co"># How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.</span></span>
+<span id="cb1-247"><a href="#cb1-247" aria-hidden="true" tabindex="-1"></a><span class="fu">val_set_size</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.04</span></span>
+<span id="cb1-248"><a href="#cb1-248" aria-hidden="true" tabindex="-1"></a><span class="co"># Num shards for whole dataset</span></span>
+<span id="cb1-249"><a href="#cb1-249" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_shard_num</span><span class="kw">:</span></span>
+<span id="cb1-250"><a href="#cb1-250" aria-hidden="true" tabindex="-1"></a><span class="co"># Index of shard to use for whole dataset</span></span>
+<span id="cb1-251"><a href="#cb1-251" aria-hidden="true" tabindex="-1"></a><span class="fu">dataset_shard_idx</span><span class="kw">:</span></span>
+<span id="cb1-252"><a href="#cb1-252" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-253"><a href="#cb1-253" aria-hidden="true" tabindex="-1"></a><span class="co"># The maximum length of an input to train with, this should typically be less than 2048</span></span>
+<span id="cb1-254"><a href="#cb1-254" aria-hidden="true" tabindex="-1"></a><span class="co"># as most models have a token/context limit of 2048</span></span>
+<span id="cb1-255"><a href="#cb1-255" aria-hidden="true" tabindex="-1"></a><span class="fu">sequence_len</span><span class="kw">:</span><span class="at"> </span><span class="dv">2048</span></span>
+<span id="cb1-256"><a href="#cb1-256" aria-hidden="true" tabindex="-1"></a><span class="co"># Pad inputs so each step uses constant sized buffers</span></span>
+<span id="cb1-257"><a href="#cb1-257" aria-hidden="true" tabindex="-1"></a><span class="co"># This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently</span></span>
+<span id="cb1-258"><a href="#cb1-258" aria-hidden="true" tabindex="-1"></a><span class="fu">pad_to_sequence_len</span><span class="kw">:</span></span>
+<span id="cb1-259"><a href="#cb1-259" aria-hidden="true" tabindex="-1"></a><span class="co"># Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'</span></span>
+<span id="cb1-260"><a href="#cb1-260" aria-hidden="true" tabindex="-1"></a><span class="fu">sample_packing</span><span class="kw">:</span></span>
+<span id="cb1-261"><a href="#cb1-261" aria-hidden="true" tabindex="-1"></a><span class="co"># Set to 'false' if getting errors during eval with sample_packing on.</span></span>
+<span id="cb1-262"><a href="#cb1-262" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_sample_packing</span><span class="kw">:</span></span>
+<span id="cb1-263"><a href="#cb1-263" aria-hidden="true" tabindex="-1"></a><span class="co"># You can set these packing optimizations AFTER starting a training at least once.</span></span>
+<span id="cb1-264"><a href="#cb1-264" aria-hidden="true" tabindex="-1"></a><span class="co"># The trainer will provide recommended values for these values.</span></span>
+<span id="cb1-265"><a href="#cb1-265" aria-hidden="true" tabindex="-1"></a><span class="fu">sample_packing_eff_est</span><span class="kw">:</span></span>
+<span id="cb1-266"><a href="#cb1-266" aria-hidden="true" tabindex="-1"></a><span class="fu">total_num_tokens</span><span class="kw">:</span></span>
+<span id="cb1-267"><a href="#cb1-267" aria-hidden="true" tabindex="-1"></a><span class="co"># Increasing the following values helps with packing, but usually only slightly (&lt;%1.)</span></span>
+<span id="cb1-268"><a href="#cb1-268" aria-hidden="true" tabindex="-1"></a><span class="co"># The number of samples packed at a time.</span></span>
+<span id="cb1-269"><a href="#cb1-269" aria-hidden="true" tabindex="-1"></a><span class="fu">sample_packing_group_size</span><span class="kw">:</span><span class="at"> </span><span class="dv">100000</span></span>
+<span id="cb1-270"><a href="#cb1-270" aria-hidden="true" tabindex="-1"></a><span class="co"># The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.</span></span>
+<span id="cb1-271"><a href="#cb1-271" aria-hidden="true" tabindex="-1"></a><span class="fu">sample_packing_bin_size</span><span class="kw">:</span><span class="at"> </span><span class="dv">200</span></span>
+<span id="cb1-272"><a href="#cb1-272" aria-hidden="true" tabindex="-1"></a><span class="co"># whether to concatenate samples during pretraining</span></span>
+<span id="cb1-273"><a href="#cb1-273" aria-hidden="true" tabindex="-1"></a><span class="fu">pretraining_sample_concatenation</span><span class="kw">:</span></span>
+<span id="cb1-274"><a href="#cb1-274" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-275"><a href="#cb1-275" aria-hidden="true" tabindex="-1"></a><span class="co"># Use batch flattening for speedups when not using sample_packing</span></span>
+<span id="cb1-276"><a href="#cb1-276" aria-hidden="true" tabindex="-1"></a><span class="fu">batch_flattening</span><span class="kw">:</span></span>
 <span id="cb1-277"><a href="#cb1-277" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-278"><a href="#cb1-278" aria-hidden="true" tabindex="-1"></a><span class="co"># If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model</span></span>
-<span id="cb1-279"><a href="#cb1-279" aria-hidden="true" tabindex="-1"></a><span class="fu">adapter</span><span class="kw">:</span><span class="at"> lora</span></span>
-<span id="cb1-280"><a href="#cb1-280" aria-hidden="true" tabindex="-1"></a><span class="co"># If you already have a lora model trained that you want to load, put that here.</span></span>
-<span id="cb1-281"><a href="#cb1-281" aria-hidden="true" tabindex="-1"></a><span class="co"># This means after training, if you want to test the model, you should set this to the value of `output_dir`.</span></span>
-<span id="cb1-282"><a href="#cb1-282" aria-hidden="true" tabindex="-1"></a><span class="co"># Note that if you merge an adapter to the base model, a new subdirectory `merged` will be created under the `output_dir`.</span></span>
-<span id="cb1-283"><a href="#cb1-283" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_model_dir</span><span class="kw">:</span></span>
-<span id="cb1-284"><a href="#cb1-284" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-285"><a href="#cb1-285" aria-hidden="true" tabindex="-1"></a><span class="co"># LoRA hyperparameters</span></span>
-<span id="cb1-286"><a href="#cb1-286" aria-hidden="true" tabindex="-1"></a><span class="co"># For more details about the following options, see:</span></span>
-<span id="cb1-287"><a href="#cb1-287" aria-hidden="true" tabindex="-1"></a><span class="co"># https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2</span></span>
-<span id="cb1-288"><a href="#cb1-288" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_r</span><span class="kw">:</span><span class="at"> </span><span class="dv">8</span></span>
-<span id="cb1-289"><a href="#cb1-289" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_alpha</span><span class="kw">:</span><span class="at"> </span><span class="dv">16</span></span>
-<span id="cb1-290"><a href="#cb1-290" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_dropout</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.05</span></span>
-<span id="cb1-291"><a href="#cb1-291" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_target_modules</span><span class="kw">:</span></span>
-<span id="cb1-292"><a href="#cb1-292" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> q_proj</span></span>
-<span id="cb1-293"><a href="#cb1-293" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> v_proj</span></span>
-<span id="cb1-294"><a href="#cb1-294" aria-hidden="true" tabindex="-1"></a><span class="co">#  - k_proj</span></span>
-<span id="cb1-295"><a href="#cb1-295" aria-hidden="true" tabindex="-1"></a><span class="co">#  - o_proj</span></span>
-<span id="cb1-296"><a href="#cb1-296" aria-hidden="true" tabindex="-1"></a><span class="co">#  - gate_proj</span></span>
-<span id="cb1-297"><a href="#cb1-297" aria-hidden="true" tabindex="-1"></a><span class="co">#  - down_proj</span></span>
-<span id="cb1-298"><a href="#cb1-298" aria-hidden="true" tabindex="-1"></a><span class="co">#  - up_proj</span></span>
-<span id="cb1-299"><a href="#cb1-299" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_target_linear</span><span class="kw">:</span><span class="co"> # If true, will target all linear modules</span></span>
-<span id="cb1-300"><a href="#cb1-300" aria-hidden="true" tabindex="-1"></a><span class="fu">peft_layers_to_transform</span><span class="kw">:</span><span class="co"> # The layer indices to transform, otherwise, apply to all layers</span></span>
-<span id="cb1-301"><a href="#cb1-301" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-302"><a href="#cb1-302" aria-hidden="true" tabindex="-1"></a><span class="co"># If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.</span></span>
-<span id="cb1-303"><a href="#cb1-303" aria-hidden="true" tabindex="-1"></a><span class="co"># For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.</span></span>
-<span id="cb1-304"><a href="#cb1-304" aria-hidden="true" tabindex="-1"></a><span class="co"># `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.</span></span>
-<span id="cb1-305"><a href="#cb1-305" aria-hidden="true" tabindex="-1"></a><span class="co"># https://github.com/huggingface/peft/issues/334#issuecomment-1561727994</span></span>
-<span id="cb1-306"><a href="#cb1-306" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_modules_to_save</span><span class="kw">:</span></span>
-<span id="cb1-307"><a href="#cb1-307" aria-hidden="true" tabindex="-1"></a><span class="co">#  - embed_tokens</span></span>
-<span id="cb1-308"><a href="#cb1-308" aria-hidden="true" tabindex="-1"></a><span class="co">#  - lm_head</span></span>
-<span id="cb1-309"><a href="#cb1-309" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-310"><a href="#cb1-310" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_fan_in_fan_out</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
-<span id="cb1-311"><a href="#cb1-311" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-312"><a href="#cb1-312" aria-hidden="true" tabindex="-1"></a><span class="co"># Apply custom LoRA autograd functions and activation function Triton kernels for</span></span>
-<span id="cb1-313"><a href="#cb1-313" aria-hidden="true" tabindex="-1"></a><span class="co"># speed and memory savings</span></span>
-<span id="cb1-314"><a href="#cb1-314" aria-hidden="true" tabindex="-1"></a><span class="co"># See: https://axolotl-ai-cloud.github.io/axolotl/docs/lora_optims.html</span></span>
-<span id="cb1-315"><a href="#cb1-315" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_mlp_kernel</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
-<span id="cb1-316"><a href="#cb1-316" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_qkv_kernel</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
-<span id="cb1-317"><a href="#cb1-317" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_o_kernel</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
-<span id="cb1-318"><a href="#cb1-318" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-319"><a href="#cb1-319" aria-hidden="true" tabindex="-1"></a><span class="co"># LoRA+ hyperparameters</span></span>
-<span id="cb1-320"><a href="#cb1-320" aria-hidden="true" tabindex="-1"></a><span class="co"># For more details about the following options, see:</span></span>
-<span id="cb1-321"><a href="#cb1-321" aria-hidden="true" tabindex="-1"></a><span class="co"># https://arxiv.org/abs/2402.12354  and `src/axolotl/core/train_builder.py`</span></span>
-<span id="cb1-322"><a href="#cb1-322" aria-hidden="true" tabindex="-1"></a><span class="fu">loraplus_lr_ratio</span><span class="kw">:</span><span class="co"> # loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4.</span></span>
-<span id="cb1-323"><a href="#cb1-323" aria-hidden="true" tabindex="-1"></a><span class="fu">loraplus_lr_embedding</span><span class="kw">:</span><span class="co"> #  loraplus learning rate for lora embedding layers. Default value is 1e-6.</span></span>
+<span id="cb1-278"><a href="#cb1-278" aria-hidden="true" tabindex="-1"></a><span class="co"># Passed through to transformers when loading the model when launched without accelerate</span></span>
+<span id="cb1-279"><a href="#cb1-279" aria-hidden="true" tabindex="-1"></a><span class="co"># Use `sequential` when training w/ model parallelism to limit memory</span></span>
+<span id="cb1-280"><a href="#cb1-280" aria-hidden="true" tabindex="-1"></a><span class="fu">device_map</span><span class="kw">:</span></span>
+<span id="cb1-281"><a href="#cb1-281" aria-hidden="true" tabindex="-1"></a><span class="co"># Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model.</span></span>
+<span id="cb1-282"><a href="#cb1-282" aria-hidden="true" tabindex="-1"></a><span class="fu">max_memory</span><span class="kw">:</span></span>
+<span id="cb1-283"><a href="#cb1-283" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-284"><a href="#cb1-284" aria-hidden="true" tabindex="-1"></a><span class="co"># If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model</span></span>
+<span id="cb1-285"><a href="#cb1-285" aria-hidden="true" tabindex="-1"></a><span class="fu">adapter</span><span class="kw">:</span><span class="at"> lora</span></span>
+<span id="cb1-286"><a href="#cb1-286" aria-hidden="true" tabindex="-1"></a><span class="co"># If you already have a lora model trained that you want to load, put that here.</span></span>
+<span id="cb1-287"><a href="#cb1-287" aria-hidden="true" tabindex="-1"></a><span class="co"># This means after training, if you want to test the model, you should set this to the value of `output_dir`.</span></span>
+<span id="cb1-288"><a href="#cb1-288" aria-hidden="true" tabindex="-1"></a><span class="co"># Note that if you merge an adapter to the base model, a new subdirectory `merged` will be created under the `output_dir`.</span></span>
+<span id="cb1-289"><a href="#cb1-289" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_model_dir</span><span class="kw">:</span></span>
+<span id="cb1-290"><a href="#cb1-290" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-291"><a href="#cb1-291" aria-hidden="true" tabindex="-1"></a><span class="co"># LoRA hyperparameters</span></span>
+<span id="cb1-292"><a href="#cb1-292" aria-hidden="true" tabindex="-1"></a><span class="co"># For more details about the following options, see:</span></span>
+<span id="cb1-293"><a href="#cb1-293" aria-hidden="true" tabindex="-1"></a><span class="co"># https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2</span></span>
+<span id="cb1-294"><a href="#cb1-294" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_r</span><span class="kw">:</span><span class="at"> </span><span class="dv">8</span></span>
+<span id="cb1-295"><a href="#cb1-295" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_alpha</span><span class="kw">:</span><span class="at"> </span><span class="dv">16</span></span>
+<span id="cb1-296"><a href="#cb1-296" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_dropout</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.05</span></span>
+<span id="cb1-297"><a href="#cb1-297" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_target_modules</span><span class="kw">:</span></span>
+<span id="cb1-298"><a href="#cb1-298" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> q_proj</span></span>
+<span id="cb1-299"><a href="#cb1-299" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> v_proj</span></span>
+<span id="cb1-300"><a href="#cb1-300" aria-hidden="true" tabindex="-1"></a><span class="co">#  - k_proj</span></span>
+<span id="cb1-301"><a href="#cb1-301" aria-hidden="true" tabindex="-1"></a><span class="co">#  - o_proj</span></span>
+<span id="cb1-302"><a href="#cb1-302" aria-hidden="true" tabindex="-1"></a><span class="co">#  - gate_proj</span></span>
+<span id="cb1-303"><a href="#cb1-303" aria-hidden="true" tabindex="-1"></a><span class="co">#  - down_proj</span></span>
+<span id="cb1-304"><a href="#cb1-304" aria-hidden="true" tabindex="-1"></a><span class="co">#  - up_proj</span></span>
+<span id="cb1-305"><a href="#cb1-305" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_target_linear</span><span class="kw">:</span><span class="co"> # If true, will target all linear modules</span></span>
+<span id="cb1-306"><a href="#cb1-306" aria-hidden="true" tabindex="-1"></a><span class="fu">peft_layers_to_transform</span><span class="kw">:</span><span class="co"> # The layer indices to transform, otherwise, apply to all layers</span></span>
+<span id="cb1-307"><a href="#cb1-307" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-308"><a href="#cb1-308" aria-hidden="true" tabindex="-1"></a><span class="co"># If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.</span></span>
+<span id="cb1-309"><a href="#cb1-309" aria-hidden="true" tabindex="-1"></a><span class="co"># For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.</span></span>
+<span id="cb1-310"><a href="#cb1-310" aria-hidden="true" tabindex="-1"></a><span class="co"># `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.</span></span>
+<span id="cb1-311"><a href="#cb1-311" aria-hidden="true" tabindex="-1"></a><span class="co"># https://github.com/huggingface/peft/issues/334#issuecomment-1561727994</span></span>
+<span id="cb1-312"><a href="#cb1-312" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_modules_to_save</span><span class="kw">:</span></span>
+<span id="cb1-313"><a href="#cb1-313" aria-hidden="true" tabindex="-1"></a><span class="co">#  - embed_tokens</span></span>
+<span id="cb1-314"><a href="#cb1-314" aria-hidden="true" tabindex="-1"></a><span class="co">#  - lm_head</span></span>
+<span id="cb1-315"><a href="#cb1-315" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-316"><a href="#cb1-316" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_fan_in_fan_out</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
+<span id="cb1-317"><a href="#cb1-317" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-318"><a href="#cb1-318" aria-hidden="true" tabindex="-1"></a><span class="co"># Apply custom LoRA autograd functions and activation function Triton kernels for</span></span>
+<span id="cb1-319"><a href="#cb1-319" aria-hidden="true" tabindex="-1"></a><span class="co"># speed and memory savings</span></span>
+<span id="cb1-320"><a href="#cb1-320" aria-hidden="true" tabindex="-1"></a><span class="co"># See: https://axolotl-ai-cloud.github.io/axolotl/docs/lora_optims.html</span></span>
+<span id="cb1-321"><a href="#cb1-321" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_mlp_kernel</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
+<span id="cb1-322"><a href="#cb1-322" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_qkv_kernel</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
+<span id="cb1-323"><a href="#cb1-323" aria-hidden="true" tabindex="-1"></a><span class="fu">lora_o_kernel</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
 <span id="cb1-324"><a href="#cb1-324" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-325"><a href="#cb1-325" aria-hidden="true" tabindex="-1"></a><span class="fu">peft</span><span class="kw">:</span></span>
-<span id="cb1-326"><a href="#cb1-326" aria-hidden="true" tabindex="-1"></a><span class="co">  # Configuration options for loftq initialization for LoRA</span></span>
-<span id="cb1-327"><a href="#cb1-327" aria-hidden="true" tabindex="-1"></a><span class="co">  # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization</span></span>
-<span id="cb1-328"><a href="#cb1-328" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="fu">loftq_config</span><span class="kw">:</span></span>
-<span id="cb1-329"><a href="#cb1-329" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">loftq_bits</span><span class="kw">:</span><span class="co">  # typically 4 bits</span></span>
+<span id="cb1-325"><a href="#cb1-325" aria-hidden="true" tabindex="-1"></a><span class="co"># LoRA+ hyperparameters</span></span>
+<span id="cb1-326"><a href="#cb1-326" aria-hidden="true" tabindex="-1"></a><span class="co"># For more details about the following options, see:</span></span>
+<span id="cb1-327"><a href="#cb1-327" aria-hidden="true" tabindex="-1"></a><span class="co"># https://arxiv.org/abs/2402.12354  and `src/axolotl/core/train_builder.py`</span></span>
+<span id="cb1-328"><a href="#cb1-328" aria-hidden="true" tabindex="-1"></a><span class="fu">loraplus_lr_ratio</span><span class="kw">:</span><span class="co"> # loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4.</span></span>
+<span id="cb1-329"><a href="#cb1-329" aria-hidden="true" tabindex="-1"></a><span class="fu">loraplus_lr_embedding</span><span class="kw">:</span><span class="co"> #  loraplus learning rate for lora embedding layers. Default value is 1e-6.</span></span>
 <span id="cb1-330"><a href="#cb1-330" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-331"><a href="#cb1-331" aria-hidden="true" tabindex="-1"></a><span class="co"># ReLoRA configuration</span></span>
-<span id="cb1-332"><a href="#cb1-332" aria-hidden="true" tabindex="-1"></a><span class="co"># Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed</span></span>
-<span id="cb1-333"><a href="#cb1-333" aria-hidden="true" tabindex="-1"></a><span class="fu">relora_steps</span><span class="kw">:</span><span class="co"> # Number of steps per ReLoRA restart</span></span>
-<span id="cb1-334"><a href="#cb1-334" aria-hidden="true" tabindex="-1"></a><span class="fu">relora_warmup_steps</span><span class="kw">:</span><span class="co"> # Number of per-restart warmup steps</span></span>
-<span id="cb1-335"><a href="#cb1-335" aria-hidden="true" tabindex="-1"></a><span class="fu">relora_anneal_steps</span><span class="kw">:</span><span class="co"> # Number of anneal steps for each relora cycle</span></span>
-<span id="cb1-336"><a href="#cb1-336" aria-hidden="true" tabindex="-1"></a><span class="fu">relora_prune_ratio</span><span class="kw">:</span><span class="co"> # threshold for optimizer magnitude when pruning</span></span>
-<span id="cb1-337"><a href="#cb1-337" aria-hidden="true" tabindex="-1"></a><span class="fu">relora_cpu_offload</span><span class="kw">:</span><span class="co"> # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings</span></span>
-<span id="cb1-338"><a href="#cb1-338" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-339"><a href="#cb1-339" aria-hidden="true" tabindex="-1"></a><span class="co"># wandb configuration if you're using it</span></span>
-<span id="cb1-340"><a href="#cb1-340" aria-hidden="true" tabindex="-1"></a><span class="co"># Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.</span></span>
-<span id="cb1-341"><a href="#cb1-341" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_mode</span><span class="kw">:</span><span class="co"> # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb</span></span>
-<span id="cb1-342"><a href="#cb1-342" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_project</span><span class="kw">:</span><span class="co"> # Your wandb project name</span></span>
-<span id="cb1-343"><a href="#cb1-343" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_entity</span><span class="kw">:</span><span class="co"> # A wandb Team name if using a Team</span></span>
-<span id="cb1-344"><a href="#cb1-344" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_watch</span><span class="kw">:</span></span>
-<span id="cb1-345"><a href="#cb1-345" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_name</span><span class="kw">:</span><span class="co"> # Set the name of your wandb run</span></span>
-<span id="cb1-346"><a href="#cb1-346" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_run_id</span><span class="kw">:</span><span class="co"> # Set the ID of your wandb run</span></span>
-<span id="cb1-347"><a href="#cb1-347" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_log_model</span><span class="kw">:</span><span class="co"> # "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only at the end of training</span></span>
-<span id="cb1-348"><a href="#cb1-348" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-349"><a href="#cb1-349" aria-hidden="true" tabindex="-1"></a><span class="co"># mlflow configuration if you're using it</span></span>
-<span id="cb1-350"><a href="#cb1-350" aria-hidden="true" tabindex="-1"></a><span class="fu">mlflow_tracking_uri</span><span class="kw">:</span><span class="co"> # URI to mlflow</span></span>
-<span id="cb1-351"><a href="#cb1-351" aria-hidden="true" tabindex="-1"></a><span class="fu">mlflow_experiment_name</span><span class="kw">:</span><span class="co"> # Your experiment name</span></span>
-<span id="cb1-352"><a href="#cb1-352" aria-hidden="true" tabindex="-1"></a><span class="fu">mlflow_run_name</span><span class="kw">:</span><span class="co"> # Your run name</span></span>
-<span id="cb1-353"><a href="#cb1-353" aria-hidden="true" tabindex="-1"></a><span class="fu">hf_mlflow_log_artifacts</span><span class="kw">:</span><span class="co">  # set to true to copy each saved checkpoint on each save to mlflow artifact registry</span></span>
+<span id="cb1-331"><a href="#cb1-331" aria-hidden="true" tabindex="-1"></a><span class="fu">peft</span><span class="kw">:</span></span>
+<span id="cb1-332"><a href="#cb1-332" aria-hidden="true" tabindex="-1"></a><span class="co">  # Configuration options for loftq initialization for LoRA</span></span>
+<span id="cb1-333"><a href="#cb1-333" aria-hidden="true" tabindex="-1"></a><span class="co">  # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization</span></span>
+<span id="cb1-334"><a href="#cb1-334" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="fu">loftq_config</span><span class="kw">:</span></span>
+<span id="cb1-335"><a href="#cb1-335" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">loftq_bits</span><span class="kw">:</span><span class="co">  # typically 4 bits</span></span>
+<span id="cb1-336"><a href="#cb1-336" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-337"><a href="#cb1-337" aria-hidden="true" tabindex="-1"></a><span class="co"># ReLoRA configuration</span></span>
+<span id="cb1-338"><a href="#cb1-338" aria-hidden="true" tabindex="-1"></a><span class="co"># Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed</span></span>
+<span id="cb1-339"><a href="#cb1-339" aria-hidden="true" tabindex="-1"></a><span class="fu">relora_steps</span><span class="kw">:</span><span class="co"> # Number of steps per ReLoRA restart</span></span>
+<span id="cb1-340"><a href="#cb1-340" aria-hidden="true" tabindex="-1"></a><span class="fu">relora_warmup_steps</span><span class="kw">:</span><span class="co"> # Number of per-restart warmup steps</span></span>
+<span id="cb1-341"><a href="#cb1-341" aria-hidden="true" tabindex="-1"></a><span class="fu">relora_anneal_steps</span><span class="kw">:</span><span class="co"> # Number of anneal steps for each relora cycle</span></span>
+<span id="cb1-342"><a href="#cb1-342" aria-hidden="true" tabindex="-1"></a><span class="fu">relora_prune_ratio</span><span class="kw">:</span><span class="co"> # threshold for optimizer magnitude when pruning</span></span>
+<span id="cb1-343"><a href="#cb1-343" aria-hidden="true" tabindex="-1"></a><span class="fu">relora_cpu_offload</span><span class="kw">:</span><span class="co"> # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings</span></span>
+<span id="cb1-344"><a href="#cb1-344" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-345"><a href="#cb1-345" aria-hidden="true" tabindex="-1"></a><span class="co"># wandb configuration if you're using it</span></span>
+<span id="cb1-346"><a href="#cb1-346" aria-hidden="true" tabindex="-1"></a><span class="co"># Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.</span></span>
+<span id="cb1-347"><a href="#cb1-347" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_mode</span><span class="kw">:</span><span class="co"> # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb</span></span>
+<span id="cb1-348"><a href="#cb1-348" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_project</span><span class="kw">:</span><span class="co"> # Your wandb project name</span></span>
+<span id="cb1-349"><a href="#cb1-349" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_entity</span><span class="kw">:</span><span class="co"> # A wandb Team name if using a Team</span></span>
+<span id="cb1-350"><a href="#cb1-350" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_watch</span><span class="kw">:</span></span>
+<span id="cb1-351"><a href="#cb1-351" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_name</span><span class="kw">:</span><span class="co"> # Set the name of your wandb run</span></span>
+<span id="cb1-352"><a href="#cb1-352" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_run_id</span><span class="kw">:</span><span class="co"> # Set the ID of your wandb run</span></span>
+<span id="cb1-353"><a href="#cb1-353" aria-hidden="true" tabindex="-1"></a><span class="fu">wandb_log_model</span><span class="kw">:</span><span class="co"> # "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only at the end of training</span></span>
 <span id="cb1-354"><a href="#cb1-354" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-355"><a href="#cb1-355" aria-hidden="true" tabindex="-1"></a><span class="co"># Comet configuration if you're using it</span></span>
-<span id="cb1-356"><a href="#cb1-356" aria-hidden="true" tabindex="-1"></a><span class="co"># Make sure your `COMET_API_KEY` environment variable is set (recommended) or you login to Comet with `comet login`.</span></span>
-<span id="cb1-357"><a href="#cb1-357" aria-hidden="true" tabindex="-1"></a><span class="co"># Check out our documentation for more details https://www.comet.com/docs/v2/api-and-sdk/python-sdk/reference/Experiment-Creation/#comet_ml.start</span></span>
-<span id="cb1-358"><a href="#cb1-358" aria-hidden="true" tabindex="-1"></a><span class="fu">use_comet</span><span class="kw">:</span><span class="co"> # Enable or disable Comet integration.</span></span>
-<span id="cb1-359"><a href="#cb1-359" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_api_key</span><span class="kw">:</span><span class="co"> # API key for Comet. Recommended to set via `comet login`.</span></span>
-<span id="cb1-360"><a href="#cb1-360" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_workspace</span><span class="kw">:</span><span class="co"> # Workspace name in Comet. Defaults to the user's default workspace.</span></span>
-<span id="cb1-361"><a href="#cb1-361" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_project_name</span><span class="kw">:</span><span class="co"> # Project name in Comet. Defaults to Uncategorized.</span></span>
-<span id="cb1-362"><a href="#cb1-362" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_experiment_key</span><span class="kw">:</span><span class="co"> # Identifier for the experiment. Used to append data to an existing experiment or control the key of new experiments. Default to a random key.</span></span>
-<span id="cb1-363"><a href="#cb1-363" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_mode</span><span class="kw">:</span><span class="co"> # Create a new experiment ("create") or log to an existing one ("get"). Default ("get_or_create") auto-selects based on configuration.</span></span>
-<span id="cb1-364"><a href="#cb1-364" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_online</span><span class="kw">:</span><span class="co"> # Set to True to log data to Comet server, or False for offline storage. Default is True.</span></span>
-<span id="cb1-365"><a href="#cb1-365" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_experiment_config</span><span class="kw">:</span><span class="co"> # Dictionary for additional configuration settings, see the doc for more details.</span></span>
-<span id="cb1-366"><a href="#cb1-366" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-367"><a href="#cb1-367" aria-hidden="true" tabindex="-1"></a><span class="co"># Tensorboard</span></span>
-<span id="cb1-368"><a href="#cb1-368" aria-hidden="true" tabindex="-1"></a><span class="fu">use_tensorboard</span><span class="kw">:</span><span class="co"> # Optional[bool]</span></span>
-<span id="cb1-369"><a href="#cb1-369" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-370"><a href="#cb1-370" aria-hidden="true" tabindex="-1"></a><span class="co"># Where to save the full-finetuned model to</span></span>
-<span id="cb1-371"><a href="#cb1-371" aria-hidden="true" tabindex="-1"></a><span class="fu">output_dir</span><span class="kw">:</span><span class="at"> ./completed-model</span></span>
+<span id="cb1-355"><a href="#cb1-355" aria-hidden="true" tabindex="-1"></a><span class="co"># mlflow configuration if you're using it</span></span>
+<span id="cb1-356"><a href="#cb1-356" aria-hidden="true" tabindex="-1"></a><span class="fu">mlflow_tracking_uri</span><span class="kw">:</span><span class="co"> # URI to mlflow</span></span>
+<span id="cb1-357"><a href="#cb1-357" aria-hidden="true" tabindex="-1"></a><span class="fu">mlflow_experiment_name</span><span class="kw">:</span><span class="co"> # Your experiment name</span></span>
+<span id="cb1-358"><a href="#cb1-358" aria-hidden="true" tabindex="-1"></a><span class="fu">mlflow_run_name</span><span class="kw">:</span><span class="co"> # Your run name</span></span>
+<span id="cb1-359"><a href="#cb1-359" aria-hidden="true" tabindex="-1"></a><span class="fu">hf_mlflow_log_artifacts</span><span class="kw">:</span><span class="co">  # set to true to copy each saved checkpoint on each save to mlflow artifact registry</span></span>
+<span id="cb1-360"><a href="#cb1-360" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-361"><a href="#cb1-361" aria-hidden="true" tabindex="-1"></a><span class="co"># Comet configuration if you're using it</span></span>
+<span id="cb1-362"><a href="#cb1-362" aria-hidden="true" tabindex="-1"></a><span class="co"># Make sure your `COMET_API_KEY` environment variable is set (recommended) or you login to Comet with `comet login`.</span></span>
+<span id="cb1-363"><a href="#cb1-363" aria-hidden="true" tabindex="-1"></a><span class="co"># Check out our documentation for more details https://www.comet.com/docs/v2/api-and-sdk/python-sdk/reference/Experiment-Creation/#comet_ml.start</span></span>
+<span id="cb1-364"><a href="#cb1-364" aria-hidden="true" tabindex="-1"></a><span class="fu">use_comet</span><span class="kw">:</span><span class="co"> # Enable or disable Comet integration.</span></span>
+<span id="cb1-365"><a href="#cb1-365" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_api_key</span><span class="kw">:</span><span class="co"> # API key for Comet. Recommended to set via `comet login`.</span></span>
+<span id="cb1-366"><a href="#cb1-366" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_workspace</span><span class="kw">:</span><span class="co"> # Workspace name in Comet. Defaults to the user's default workspace.</span></span>
+<span id="cb1-367"><a href="#cb1-367" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_project_name</span><span class="kw">:</span><span class="co"> # Project name in Comet. Defaults to Uncategorized.</span></span>
+<span id="cb1-368"><a href="#cb1-368" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_experiment_key</span><span class="kw">:</span><span class="co"> # Identifier for the experiment. Used to append data to an existing experiment or control the key of new experiments. Default to a random key.</span></span>
+<span id="cb1-369"><a href="#cb1-369" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_mode</span><span class="kw">:</span><span class="co"> # Create a new experiment ("create") or log to an existing one ("get"). Default ("get_or_create") auto-selects based on configuration.</span></span>
+<span id="cb1-370"><a href="#cb1-370" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_online</span><span class="kw">:</span><span class="co"> # Set to True to log data to Comet server, or False for offline storage. Default is True.</span></span>
+<span id="cb1-371"><a href="#cb1-371" aria-hidden="true" tabindex="-1"></a><span class="fu">comet_experiment_config</span><span class="kw">:</span><span class="co"> # Dictionary for additional configuration settings, see the doc for more details.</span></span>
 <span id="cb1-372"><a href="#cb1-372" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-373"><a href="#cb1-373" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use torch.compile and which backend to use</span></span>
-<span id="cb1-374"><a href="#cb1-374" aria-hidden="true" tabindex="-1"></a><span class="co"># setting to `auto` will enable torch compile when torch&gt;=2.5.1</span></span>
-<span id="cb1-375"><a href="#cb1-375" aria-hidden="true" tabindex="-1"></a><span class="fu">torch_compile</span><span class="kw">:</span><span class="co">  # Optional[Union[Literal["auto"], bool]]</span></span>
-<span id="cb1-376"><a href="#cb1-376" aria-hidden="true" tabindex="-1"></a><span class="fu">torch_compile_backend</span><span class="kw">:</span><span class="co">  # Optional[str]</span></span>
-<span id="cb1-377"><a href="#cb1-377" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-378"><a href="#cb1-378" aria-hidden="true" tabindex="-1"></a><span class="co"># Training hyperparameters</span></span>
-<span id="cb1-379"><a href="#cb1-379" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-380"><a href="#cb1-380" aria-hidden="true" tabindex="-1"></a><span class="co"># If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.</span></span>
-<span id="cb1-381"><a href="#cb1-381" aria-hidden="true" tabindex="-1"></a><span class="fu">gradient_accumulation_steps</span><span class="kw">:</span><span class="at"> </span><span class="dv">1</span></span>
-<span id="cb1-382"><a href="#cb1-382" aria-hidden="true" tabindex="-1"></a><span class="co"># The number of samples to include in each batch. This is the number of samples sent to each GPU.</span></span>
-<span id="cb1-383"><a href="#cb1-383" aria-hidden="true" tabindex="-1"></a><span class="co"># Batch size per gpu = micro_batch_size * gradient_accumulation_steps</span></span>
-<span id="cb1-384"><a href="#cb1-384" aria-hidden="true" tabindex="-1"></a><span class="fu">micro_batch_size</span><span class="kw">:</span><span class="at"> </span><span class="dv">2</span></span>
-<span id="cb1-385"><a href="#cb1-385" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_batch_size</span><span class="kw">:</span></span>
-<span id="cb1-386"><a href="#cb1-386" aria-hidden="true" tabindex="-1"></a><span class="fu">num_epochs</span><span class="kw">:</span><span class="at"> </span><span class="dv">4</span></span>
-<span id="cb1-387"><a href="#cb1-387" aria-hidden="true" tabindex="-1"></a><span class="fu">warmup_steps</span><span class="kw">:</span><span class="at"> </span><span class="dv">100</span><span class="co">  # cannot use with warmup_ratio</span></span>
-<span id="cb1-388"><a href="#cb1-388" aria-hidden="true" tabindex="-1"></a><span class="fu">warmup_ratio</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.05</span><span class="co">  # cannot use with warmup_steps</span></span>
-<span id="cb1-389"><a href="#cb1-389" aria-hidden="true" tabindex="-1"></a><span class="fu">learning_rate</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.00003</span></span>
-<span id="cb1-390"><a href="#cb1-390" aria-hidden="true" tabindex="-1"></a><span class="fu">lr_quadratic_warmup</span><span class="kw">:</span></span>
-<span id="cb1-391"><a href="#cb1-391" aria-hidden="true" tabindex="-1"></a><span class="fu">logging_steps</span><span class="kw">:</span></span>
-<span id="cb1-392"><a href="#cb1-392" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_steps</span><span class="kw">:</span><span class="co"> # Leave empty to eval at each epoch, integer for every N steps. float for fraction of total steps</span></span>
-<span id="cb1-393"><a href="#cb1-393" aria-hidden="true" tabindex="-1"></a><span class="fu">evals_per_epoch</span><span class="kw">:</span><span class="co"> # number of times per epoch to run evals, mutually exclusive with eval_steps</span></span>
-<span id="cb1-394"><a href="#cb1-394" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_strategy</span><span class="kw">:</span><span class="co"> # Set to `"no"` to skip evaluation, `"epoch"` at end of each epoch, leave empty to infer from `eval_steps`.</span></span>
-<span id="cb1-395"><a href="#cb1-395" aria-hidden="true" tabindex="-1"></a><span class="fu">save_strategy</span><span class="kw">:</span><span class="co"> # Set to `"no"` to skip checkpoint saves, `"epoch"` at end of each epoch, `"best"` when better result is achieved, leave empty to infer from `save_steps`.</span></span>
-<span id="cb1-396"><a href="#cb1-396" aria-hidden="true" tabindex="-1"></a><span class="fu">save_steps</span><span class="kw">:</span><span class="co"> # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps</span></span>
-<span id="cb1-397"><a href="#cb1-397" aria-hidden="true" tabindex="-1"></a><span class="fu">saves_per_epoch</span><span class="kw">:</span><span class="co"> # number of times per epoch to save a checkpoint, mutually exclusive with save_steps</span></span>
-<span id="cb1-398"><a href="#cb1-398" aria-hidden="true" tabindex="-1"></a><span class="fu">save_total_limit</span><span class="kw">:</span><span class="co"> # Checkpoints saved at a time</span></span>
-<span id="cb1-399"><a href="#cb1-399" aria-hidden="true" tabindex="-1"></a><span class="co"># Maximum number of iterations to train for. It precedes num_epochs which means that</span></span>
-<span id="cb1-400"><a href="#cb1-400" aria-hidden="true" tabindex="-1"></a><span class="co"># if both are set, num_epochs will not be guaranteed.</span></span>
-<span id="cb1-401"><a href="#cb1-401" aria-hidden="true" tabindex="-1"></a><span class="co"># e.g., when 1 epoch is 1000 steps =&gt; `num_epochs: 2` and `max_steps: 100` will train for 100 steps</span></span>
-<span id="cb1-402"><a href="#cb1-402" aria-hidden="true" tabindex="-1"></a><span class="fu">max_steps</span><span class="kw">:</span></span>
-<span id="cb1-403"><a href="#cb1-403" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-404"><a href="#cb1-404" aria-hidden="true" tabindex="-1"></a><span class="co"># bool of whether to include tokens trainer per second in the training metrics. This iterates over the entire dataset once, so it takes some time.</span></span>
-<span id="cb1-405"><a href="#cb1-405" aria-hidden="true" tabindex="-1"></a><span class="fu">include_tokens_per_second</span><span class="kw">:</span><span class="co"> # Optional[bool]</span></span>
-<span id="cb1-406"><a href="#cb1-406" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-407"><a href="#cb1-407" aria-hidden="true" tabindex="-1"></a><span class="co"># whether to find batch size that fits in memory. Passed to underlying transformers Trainer</span></span>
-<span id="cb1-408"><a href="#cb1-408" aria-hidden="true" tabindex="-1"></a><span class="fu">auto_find_batch_size</span><span class="kw">:</span><span class="co"> # Optional[bool]</span></span>
+<span id="cb1-373"><a href="#cb1-373" aria-hidden="true" tabindex="-1"></a><span class="co"># Tensorboard</span></span>
+<span id="cb1-374"><a href="#cb1-374" aria-hidden="true" tabindex="-1"></a><span class="fu">use_tensorboard</span><span class="kw">:</span><span class="co"> # Optional[bool]</span></span>
+<span id="cb1-375"><a href="#cb1-375" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-376"><a href="#cb1-376" aria-hidden="true" tabindex="-1"></a><span class="co"># Where to save the full-finetuned model to</span></span>
+<span id="cb1-377"><a href="#cb1-377" aria-hidden="true" tabindex="-1"></a><span class="fu">output_dir</span><span class="kw">:</span><span class="at"> ./completed-model</span></span>
+<span id="cb1-378"><a href="#cb1-378" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-379"><a href="#cb1-379" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use torch.compile and which backend to use</span></span>
+<span id="cb1-380"><a href="#cb1-380" aria-hidden="true" tabindex="-1"></a><span class="co"># setting to `auto` will enable torch compile when torch&gt;=2.5.1</span></span>
+<span id="cb1-381"><a href="#cb1-381" aria-hidden="true" tabindex="-1"></a><span class="fu">torch_compile</span><span class="kw">:</span><span class="co">  # Optional[Union[Literal["auto"], bool]]</span></span>
+<span id="cb1-382"><a href="#cb1-382" aria-hidden="true" tabindex="-1"></a><span class="fu">torch_compile_backend</span><span class="kw">:</span><span class="co">  # Optional[str]</span></span>
+<span id="cb1-383"><a href="#cb1-383" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-384"><a href="#cb1-384" aria-hidden="true" tabindex="-1"></a><span class="co"># Training hyperparameters</span></span>
+<span id="cb1-385"><a href="#cb1-385" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-386"><a href="#cb1-386" aria-hidden="true" tabindex="-1"></a><span class="co"># If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.</span></span>
+<span id="cb1-387"><a href="#cb1-387" aria-hidden="true" tabindex="-1"></a><span class="fu">gradient_accumulation_steps</span><span class="kw">:</span><span class="at"> </span><span class="dv">1</span></span>
+<span id="cb1-388"><a href="#cb1-388" aria-hidden="true" tabindex="-1"></a><span class="co"># The number of samples to include in each batch. This is the number of samples sent to each GPU.</span></span>
+<span id="cb1-389"><a href="#cb1-389" aria-hidden="true" tabindex="-1"></a><span class="co"># Batch size per gpu = micro_batch_size * gradient_accumulation_steps</span></span>
+<span id="cb1-390"><a href="#cb1-390" aria-hidden="true" tabindex="-1"></a><span class="fu">micro_batch_size</span><span class="kw">:</span><span class="at"> </span><span class="dv">2</span></span>
+<span id="cb1-391"><a href="#cb1-391" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_batch_size</span><span class="kw">:</span></span>
+<span id="cb1-392"><a href="#cb1-392" aria-hidden="true" tabindex="-1"></a><span class="fu">num_epochs</span><span class="kw">:</span><span class="at"> </span><span class="dv">4</span></span>
+<span id="cb1-393"><a href="#cb1-393" aria-hidden="true" tabindex="-1"></a><span class="fu">warmup_steps</span><span class="kw">:</span><span class="at"> </span><span class="dv">100</span><span class="co">  # cannot use with warmup_ratio</span></span>
+<span id="cb1-394"><a href="#cb1-394" aria-hidden="true" tabindex="-1"></a><span class="fu">warmup_ratio</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.05</span><span class="co">  # cannot use with warmup_steps</span></span>
+<span id="cb1-395"><a href="#cb1-395" aria-hidden="true" tabindex="-1"></a><span class="fu">learning_rate</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.00003</span></span>
+<span id="cb1-396"><a href="#cb1-396" aria-hidden="true" tabindex="-1"></a><span class="fu">lr_quadratic_warmup</span><span class="kw">:</span></span>
+<span id="cb1-397"><a href="#cb1-397" aria-hidden="true" tabindex="-1"></a><span class="fu">logging_steps</span><span class="kw">:</span></span>
+<span id="cb1-398"><a href="#cb1-398" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_steps</span><span class="kw">:</span><span class="co"> # Leave empty to eval at each epoch, integer for every N steps. float for fraction of total steps</span></span>
+<span id="cb1-399"><a href="#cb1-399" aria-hidden="true" tabindex="-1"></a><span class="fu">evals_per_epoch</span><span class="kw">:</span><span class="co"> # number of times per epoch to run evals, mutually exclusive with eval_steps</span></span>
+<span id="cb1-400"><a href="#cb1-400" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_strategy</span><span class="kw">:</span><span class="co"> # Set to `"no"` to skip evaluation, `"epoch"` at end of each epoch, leave empty to infer from `eval_steps`.</span></span>
+<span id="cb1-401"><a href="#cb1-401" aria-hidden="true" tabindex="-1"></a><span class="fu">save_strategy</span><span class="kw">:</span><span class="co"> # Set to `"no"` to skip checkpoint saves, `"epoch"` at end of each epoch, `"best"` when better result is achieved, leave empty to infer from `save_steps`.</span></span>
+<span id="cb1-402"><a href="#cb1-402" aria-hidden="true" tabindex="-1"></a><span class="fu">save_steps</span><span class="kw">:</span><span class="co"> # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps</span></span>
+<span id="cb1-403"><a href="#cb1-403" aria-hidden="true" tabindex="-1"></a><span class="fu">saves_per_epoch</span><span class="kw">:</span><span class="co"> # number of times per epoch to save a checkpoint, mutually exclusive with save_steps</span></span>
+<span id="cb1-404"><a href="#cb1-404" aria-hidden="true" tabindex="-1"></a><span class="fu">save_total_limit</span><span class="kw">:</span><span class="co"> # Checkpoints saved at a time</span></span>
+<span id="cb1-405"><a href="#cb1-405" aria-hidden="true" tabindex="-1"></a><span class="co"># Maximum number of iterations to train for. It precedes num_epochs which means that</span></span>
+<span id="cb1-406"><a href="#cb1-406" aria-hidden="true" tabindex="-1"></a><span class="co"># if both are set, num_epochs will not be guaranteed.</span></span>
+<span id="cb1-407"><a href="#cb1-407" aria-hidden="true" tabindex="-1"></a><span class="co"># e.g., when 1 epoch is 1000 steps =&gt; `num_epochs: 2` and `max_steps: 100` will train for 100 steps</span></span>
+<span id="cb1-408"><a href="#cb1-408" aria-hidden="true" tabindex="-1"></a><span class="fu">max_steps</span><span class="kw">:</span></span>
 <span id="cb1-409"><a href="#cb1-409" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-410"><a href="#cb1-410" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_table_size</span><span class="kw">:</span><span class="co"> # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0</span></span>
-<span id="cb1-411"><a href="#cb1-411" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_max_new_tokens</span><span class="kw">:</span><span class="co"> # Total number of tokens generated for predictions sent to wandb. Default is 128</span></span>
-<span id="cb1-412"><a href="#cb1-412" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_causal_lm_metrics</span><span class="kw">:</span><span class="co"> # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", "chrf", "perplexity"]</span></span>
-<span id="cb1-413"><a href="#cb1-413" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-414"><a href="#cb1-414" aria-hidden="true" tabindex="-1"></a><span class="fu">profiler_steps</span><span class="kw">:</span><span class="co"> # enable the pytorch profiler to capture the first N steps of training to the output_dir.</span></span>
-<span id="cb1-415"><a href="#cb1-415" aria-hidden="true" tabindex="-1"></a><span class="co">                # see https://pytorch.org/blog/understanding-gpu-memory-1/ for more information</span></span>
-<span id="cb1-416"><a href="#cb1-416" aria-hidden="true" tabindex="-1"></a><span class="co">                # snapshots can be visualized @ https://pytorch.org/memory_viz</span></span>
-<span id="cb1-417"><a href="#cb1-417" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-418"><a href="#cb1-418" aria-hidden="true" tabindex="-1"></a><span class="fu">loss_watchdog_threshold</span><span class="kw">:</span><span class="co"> # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)</span></span>
-<span id="cb1-419"><a href="#cb1-419" aria-hidden="true" tabindex="-1"></a><span class="fu">loss_watchdog_patience</span><span class="kw">:</span><span class="co"> # Number of high-loss steps in a row before the trainer aborts (default: 3)</span></span>
-<span id="cb1-420"><a href="#cb1-420" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-421"><a href="#cb1-421" aria-hidden="true" tabindex="-1"></a><span class="co"># Save model as safetensors (require safetensors package)</span></span>
-<span id="cb1-422"><a href="#cb1-422" aria-hidden="true" tabindex="-1"></a><span class="fu">save_safetensors</span><span class="kw">:</span></span>
+<span id="cb1-410"><a href="#cb1-410" aria-hidden="true" tabindex="-1"></a><span class="co"># bool of whether to include tokens trainer per second in the training metrics. This iterates over the entire dataset once, so it takes some time.</span></span>
+<span id="cb1-411"><a href="#cb1-411" aria-hidden="true" tabindex="-1"></a><span class="fu">include_tokens_per_second</span><span class="kw">:</span><span class="co"> # Optional[bool]</span></span>
+<span id="cb1-412"><a href="#cb1-412" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-413"><a href="#cb1-413" aria-hidden="true" tabindex="-1"></a><span class="co"># whether to find batch size that fits in memory. Passed to underlying transformers Trainer</span></span>
+<span id="cb1-414"><a href="#cb1-414" aria-hidden="true" tabindex="-1"></a><span class="fu">auto_find_batch_size</span><span class="kw">:</span><span class="co"> # Optional[bool]</span></span>
+<span id="cb1-415"><a href="#cb1-415" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-416"><a href="#cb1-416" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_table_size</span><span class="kw">:</span><span class="co"> # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0</span></span>
+<span id="cb1-417"><a href="#cb1-417" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_max_new_tokens</span><span class="kw">:</span><span class="co"> # Total number of tokens generated for predictions sent to wandb. Default is 128</span></span>
+<span id="cb1-418"><a href="#cb1-418" aria-hidden="true" tabindex="-1"></a><span class="fu">eval_causal_lm_metrics</span><span class="kw">:</span><span class="co"> # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", "chrf", "perplexity"]</span></span>
+<span id="cb1-419"><a href="#cb1-419" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-420"><a href="#cb1-420" aria-hidden="true" tabindex="-1"></a><span class="fu">profiler_steps</span><span class="kw">:</span><span class="co"> # enable the pytorch profiler to capture the first N steps of training to the output_dir.</span></span>
+<span id="cb1-421"><a href="#cb1-421" aria-hidden="true" tabindex="-1"></a><span class="co">                # see https://pytorch.org/blog/understanding-gpu-memory-1/ for more information</span></span>
+<span id="cb1-422"><a href="#cb1-422" aria-hidden="true" tabindex="-1"></a><span class="co">                # snapshots can be visualized @ https://pytorch.org/memory_viz</span></span>
 <span id="cb1-423"><a href="#cb1-423" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-424"><a href="#cb1-424" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to mask out or include the human's prompt from the training labels</span></span>
-<span id="cb1-425"><a href="#cb1-425" aria-hidden="true" tabindex="-1"></a><span class="fu">train_on_inputs</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
-<span id="cb1-426"><a href="#cb1-426" aria-hidden="true" tabindex="-1"></a><span class="co"># Group similarly sized data to minimize padding.</span></span>
-<span id="cb1-427"><a href="#cb1-427" aria-hidden="true" tabindex="-1"></a><span class="co"># May be slower to start, as it must download and sort the entire dataset.</span></span>
-<span id="cb1-428"><a href="#cb1-428" aria-hidden="true" tabindex="-1"></a><span class="co"># Note that training loss may have an oscillating pattern with this enabled.</span></span>
-<span id="cb1-429"><a href="#cb1-429" aria-hidden="true" tabindex="-1"></a><span class="fu">group_by_length</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
-<span id="cb1-430"><a href="#cb1-430" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-431"><a href="#cb1-431" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing</span></span>
-<span id="cb1-432"><a href="#cb1-432" aria-hidden="true" tabindex="-1"></a><span class="fu">gradient_checkpointing</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
-<span id="cb1-433"><a href="#cb1-433" aria-hidden="true" tabindex="-1"></a><span class="co"># additional kwargs to pass to the trainer for gradient checkpointing</span></span>
-<span id="cb1-434"><a href="#cb1-434" aria-hidden="true" tabindex="-1"></a><span class="co"># gradient_checkpointing_kwargs:</span></span>
-<span id="cb1-435"><a href="#cb1-435" aria-hidden="true" tabindex="-1"></a><span class="co">#   use_reentrant: true</span></span>
+<span id="cb1-424"><a href="#cb1-424" aria-hidden="true" tabindex="-1"></a><span class="fu">loss_watchdog_threshold</span><span class="kw">:</span><span class="co"> # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)</span></span>
+<span id="cb1-425"><a href="#cb1-425" aria-hidden="true" tabindex="-1"></a><span class="fu">loss_watchdog_patience</span><span class="kw">:</span><span class="co"> # Number of high-loss steps in a row before the trainer aborts (default: 3)</span></span>
+<span id="cb1-426"><a href="#cb1-426" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-427"><a href="#cb1-427" aria-hidden="true" tabindex="-1"></a><span class="co"># Save model as safetensors (require safetensors package)</span></span>
+<span id="cb1-428"><a href="#cb1-428" aria-hidden="true" tabindex="-1"></a><span class="fu">save_safetensors</span><span class="kw">:</span></span>
+<span id="cb1-429"><a href="#cb1-429" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-430"><a href="#cb1-430" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to mask out or include the human's prompt from the training labels</span></span>
+<span id="cb1-431"><a href="#cb1-431" aria-hidden="true" tabindex="-1"></a><span class="fu">train_on_inputs</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
+<span id="cb1-432"><a href="#cb1-432" aria-hidden="true" tabindex="-1"></a><span class="co"># Group similarly sized data to minimize padding.</span></span>
+<span id="cb1-433"><a href="#cb1-433" aria-hidden="true" tabindex="-1"></a><span class="co"># May be slower to start, as it must download and sort the entire dataset.</span></span>
+<span id="cb1-434"><a href="#cb1-434" aria-hidden="true" tabindex="-1"></a><span class="co"># Note that training loss may have an oscillating pattern with this enabled.</span></span>
+<span id="cb1-435"><a href="#cb1-435" aria-hidden="true" tabindex="-1"></a><span class="fu">group_by_length</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
 <span id="cb1-436"><a href="#cb1-436" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-437"><a href="#cb1-437" aria-hidden="true" tabindex="-1"></a><span class="co"># Stop training after this many evaluation losses have increased in a row</span></span>
-<span id="cb1-438"><a href="#cb1-438" aria-hidden="true" tabindex="-1"></a><span class="co"># https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback</span></span>
-<span id="cb1-439"><a href="#cb1-439" aria-hidden="true" tabindex="-1"></a><span class="fu">early_stopping_patience</span><span class="kw">:</span><span class="at"> </span><span class="dv">3</span></span>
-<span id="cb1-440"><a href="#cb1-440" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-441"><a href="#cb1-441" aria-hidden="true" tabindex="-1"></a><span class="co"># Specify a scheduler and kwargs to use with the optimizer</span></span>
-<span id="cb1-442"><a href="#cb1-442" aria-hidden="true" tabindex="-1"></a><span class="fu">lr_scheduler</span><span class="kw">:</span><span class="co"> # 'one_cycle' | 'log_sweep' | empty for cosine</span></span>
-<span id="cb1-443"><a href="#cb1-443" aria-hidden="true" tabindex="-1"></a><span class="fu">lr_scheduler_kwargs</span><span class="kw">:</span></span>
-<span id="cb1-444"><a href="#cb1-444" aria-hidden="true" tabindex="-1"></a><span class="fu">cosine_min_lr_ratio</span><span class="kw">:</span><span class="co"> # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr</span></span>
-<span id="cb1-445"><a href="#cb1-445" aria-hidden="true" tabindex="-1"></a><span class="fu">cosine_constant_lr_ratio</span><span class="kw">:</span><span class="co"> # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)</span></span>
+<span id="cb1-437"><a href="#cb1-437" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing</span></span>
+<span id="cb1-438"><a href="#cb1-438" aria-hidden="true" tabindex="-1"></a><span class="fu">gradient_checkpointing</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
+<span id="cb1-439"><a href="#cb1-439" aria-hidden="true" tabindex="-1"></a><span class="co"># additional kwargs to pass to the trainer for gradient checkpointing</span></span>
+<span id="cb1-440"><a href="#cb1-440" aria-hidden="true" tabindex="-1"></a><span class="co"># gradient_checkpointing_kwargs:</span></span>
+<span id="cb1-441"><a href="#cb1-441" aria-hidden="true" tabindex="-1"></a><span class="co">#   use_reentrant: true</span></span>
+<span id="cb1-442"><a href="#cb1-442" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-443"><a href="#cb1-443" aria-hidden="true" tabindex="-1"></a><span class="co"># Stop training after this many evaluation losses have increased in a row</span></span>
+<span id="cb1-444"><a href="#cb1-444" aria-hidden="true" tabindex="-1"></a><span class="co"># https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback</span></span>
+<span id="cb1-445"><a href="#cb1-445" aria-hidden="true" tabindex="-1"></a><span class="fu">early_stopping_patience</span><span class="kw">:</span><span class="at"> </span><span class="dv">3</span></span>
 <span id="cb1-446"><a href="#cb1-446" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-447"><a href="#cb1-447" aria-hidden="true" tabindex="-1"></a><span class="co"># For one_cycle optim</span></span>
-<span id="cb1-448"><a href="#cb1-448" aria-hidden="true" tabindex="-1"></a><span class="fu">lr_div_factor</span><span class="kw">:</span><span class="co"> # Learning rate div factor</span></span>
-<span id="cb1-449"><a href="#cb1-449" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-450"><a href="#cb1-450" aria-hidden="true" tabindex="-1"></a><span class="co"># Specify optimizer</span></span>
-<span id="cb1-451"><a href="#cb1-451" aria-hidden="true" tabindex="-1"></a><span class="co"># Valid values are driven by the Transformers OptimizerNames class, see:</span></span>
-<span id="cb1-452"><a href="#cb1-452" aria-hidden="true" tabindex="-1"></a><span class="co"># https://github.com/huggingface/transformers/blob/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134</span></span>
-<span id="cb1-453"><a href="#cb1-453" aria-hidden="true" tabindex="-1"></a><span class="co">#</span></span>
-<span id="cb1-454"><a href="#cb1-454" aria-hidden="true" tabindex="-1"></a><span class="co"># Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of</span></span>
-<span id="cb1-455"><a href="#cb1-455" aria-hidden="true" tabindex="-1"></a><span class="co"># torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used</span></span>
-<span id="cb1-456"><a href="#cb1-456" aria-hidden="true" tabindex="-1"></a><span class="co"># in the examples/ for your model and fine-tuning use case.</span></span>
-<span id="cb1-457"><a href="#cb1-457" aria-hidden="true" tabindex="-1"></a><span class="co">#</span></span>
-<span id="cb1-458"><a href="#cb1-458" aria-hidden="true" tabindex="-1"></a><span class="co"># Valid values for 'optimizer' include:</span></span>
-<span id="cb1-459"><a href="#cb1-459" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_hf</span></span>
-<span id="cb1-460"><a href="#cb1-460" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_torch</span></span>
-<span id="cb1-461"><a href="#cb1-461" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_torch_fused</span></span>
-<span id="cb1-462"><a href="#cb1-462" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_torch_xla</span></span>
-<span id="cb1-463"><a href="#cb1-463" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_apex_fused</span></span>
-<span id="cb1-464"><a href="#cb1-464" aria-hidden="true" tabindex="-1"></a><span class="co"># - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version &gt;= 2.5.1)</span></span>
-<span id="cb1-465"><a href="#cb1-465" aria-hidden="true" tabindex="-1"></a><span class="co"># - adafactor</span></span>
-<span id="cb1-466"><a href="#cb1-466" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_anyprecision</span></span>
-<span id="cb1-467"><a href="#cb1-467" aria-hidden="true" tabindex="-1"></a><span class="co"># - sgd</span></span>
-<span id="cb1-468"><a href="#cb1-468" aria-hidden="true" tabindex="-1"></a><span class="co"># - adagrad</span></span>
-<span id="cb1-469"><a href="#cb1-469" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_bnb_8bit</span></span>
-<span id="cb1-470"><a href="#cb1-470" aria-hidden="true" tabindex="-1"></a><span class="co"># - lion_8bit</span></span>
-<span id="cb1-471"><a href="#cb1-471" aria-hidden="true" tabindex="-1"></a><span class="co"># - lion_32bit</span></span>
-<span id="cb1-472"><a href="#cb1-472" aria-hidden="true" tabindex="-1"></a><span class="co"># - paged_adamw_32bit</span></span>
-<span id="cb1-473"><a href="#cb1-473" aria-hidden="true" tabindex="-1"></a><span class="co"># - paged_adamw_8bit</span></span>
-<span id="cb1-474"><a href="#cb1-474" aria-hidden="true" tabindex="-1"></a><span class="co"># - paged_lion_32bit</span></span>
-<span id="cb1-475"><a href="#cb1-475" aria-hidden="true" tabindex="-1"></a><span class="co"># - paged_lion_8bit</span></span>
-<span id="cb1-476"><a href="#cb1-476" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adamw</span></span>
-<span id="cb1-477"><a href="#cb1-477" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adamw_8bit</span></span>
-<span id="cb1-478"><a href="#cb1-478" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adafactor</span></span>
-<span id="cb1-479"><a href="#cb1-479" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adamw_layerwise</span></span>
-<span id="cb1-480"><a href="#cb1-480" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adamw_8bit_layerwise</span></span>
-<span id="cb1-481"><a href="#cb1-481" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adafactor_layerwise</span></span>
-<span id="cb1-482"><a href="#cb1-482" aria-hidden="true" tabindex="-1"></a><span class="fu">optimizer</span><span class="kw">:</span></span>
-<span id="cb1-483"><a href="#cb1-483" aria-hidden="true" tabindex="-1"></a><span class="co"># Dictionary of arguments to pass to the optimizer</span></span>
-<span id="cb1-484"><a href="#cb1-484" aria-hidden="true" tabindex="-1"></a><span class="fu">optim_args</span><span class="kw">:</span></span>
-<span id="cb1-485"><a href="#cb1-485" aria-hidden="true" tabindex="-1"></a><span class="co"># For Galore Optimizers the following optim_args are available</span></span>
-<span id="cb1-486"><a href="#cb1-486" aria-hidden="true" tabindex="-1"></a><span class="co"># rank:  # type: int</span></span>
-<span id="cb1-487"><a href="#cb1-487" aria-hidden="true" tabindex="-1"></a><span class="co"># update_proj_gap  # type: int</span></span>
-<span id="cb1-488"><a href="#cb1-488" aria-hidden="true" tabindex="-1"></a><span class="co"># scale  # type: float</span></span>
-<span id="cb1-489"><a href="#cb1-489" aria-hidden="true" tabindex="-1"></a><span class="co"># proj_type:  # type: str, default = std</span></span>
-<span id="cb1-490"><a href="#cb1-490" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-491"><a href="#cb1-491" aria-hidden="true" tabindex="-1"></a><span class="co"># The target modules to optimize, i.e. the module names that you would like to train, right now this is used only for GaLore algorithm</span></span>
-<span id="cb1-492"><a href="#cb1-492" aria-hidden="true" tabindex="-1"></a><span class="fu">optim_target_modules</span><span class="kw">:</span></span>
-<span id="cb1-493"><a href="#cb1-493" aria-hidden="true" tabindex="-1"></a><span class="co"># - self_attn  # for llama</span></span>
-<span id="cb1-494"><a href="#cb1-494" aria-hidden="true" tabindex="-1"></a><span class="co"># - mlp</span></span>
-<span id="cb1-495"><a href="#cb1-495" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-496"><a href="#cb1-496" aria-hidden="true" tabindex="-1"></a><span class="co"># Specify weight decay</span></span>
-<span id="cb1-497"><a href="#cb1-497" aria-hidden="true" tabindex="-1"></a><span class="fu">weight_decay</span><span class="kw">:</span></span>
-<span id="cb1-498"><a href="#cb1-498" aria-hidden="true" tabindex="-1"></a><span class="co"># adamw hyperparams</span></span>
-<span id="cb1-499"><a href="#cb1-499" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_beta1</span><span class="kw">:</span></span>
-<span id="cb1-500"><a href="#cb1-500" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_beta2</span><span class="kw">:</span></span>
-<span id="cb1-501"><a href="#cb1-501" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_epsilon</span><span class="kw">:</span></span>
-<span id="cb1-502"><a href="#cb1-502" aria-hidden="true" tabindex="-1"></a><span class="co"># Gradient clipping max norm</span></span>
-<span id="cb1-503"><a href="#cb1-503" aria-hidden="true" tabindex="-1"></a><span class="fu">max_grad_norm</span><span class="kw">:</span></span>
-<span id="cb1-504"><a href="#cb1-504" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-505"><a href="#cb1-505" aria-hidden="true" tabindex="-1"></a><span class="co"># Augmentation techniques</span></span>
-<span id="cb1-506"><a href="#cb1-506" aria-hidden="true" tabindex="-1"></a><span class="co"># NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings</span></span>
-<span id="cb1-507"><a href="#cb1-507" aria-hidden="true" tabindex="-1"></a><span class="co"># currently only supported on Llama and Mistral</span></span>
-<span id="cb1-508"><a href="#cb1-508" aria-hidden="true" tabindex="-1"></a><span class="fu">neftune_noise_alpha</span><span class="kw">:</span></span>
-<span id="cb1-509"><a href="#cb1-509" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-510"><a href="#cb1-510" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to bettertransformers</span></span>
-<span id="cb1-511"><a href="#cb1-511" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_optimum</span><span class="kw">:</span></span>
-<span id="cb1-512"><a href="#cb1-512" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use xformers attention patch https://github.com/facebookresearch/xformers:</span></span>
-<span id="cb1-513"><a href="#cb1-513" aria-hidden="true" tabindex="-1"></a><span class="fu">xformers_attention</span><span class="kw">:</span></span>
-<span id="cb1-514"><a href="#cb1-514" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:</span></span>
-<span id="cb1-515"><a href="#cb1-515" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attention</span><span class="kw">:</span></span>
-<span id="cb1-516"><a href="#cb1-516" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_cross_entropy</span><span class="kw">:</span><span class="co">  # Whether to use flash-attention cross entropy implementation - advanced use only</span></span>
-<span id="cb1-517"><a href="#cb1-517" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_rms_norm</span><span class="kw">:</span><span class="co">  # Whether to use flash-attention rms norm implementation - advanced use only</span></span>
-<span id="cb1-518"><a href="#cb1-518" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_qkv</span><span class="kw">:</span><span class="co"> # Whether to fuse QKV into a single operation</span></span>
-<span id="cb1-519"><a href="#cb1-519" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_mlp</span><span class="kw">:</span><span class="co"> # Whether to fuse part of the MLP into a single operation</span></span>
-<span id="cb1-520"><a href="#cb1-520" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use scaled-dot-product attention</span></span>
-<span id="cb1-521"><a href="#cb1-521" aria-hidden="true" tabindex="-1"></a><span class="co"># https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html</span></span>
-<span id="cb1-522"><a href="#cb1-522" aria-hidden="true" tabindex="-1"></a><span class="fu">sdp_attention</span><span class="kw">:</span></span>
-<span id="cb1-523"><a href="#cb1-523" aria-hidden="true" tabindex="-1"></a><span class="co"># Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf</span></span>
-<span id="cb1-524"><a href="#cb1-524" aria-hidden="true" tabindex="-1"></a><span class="fu">s2_attention</span><span class="kw">:</span></span>
-<span id="cb1-525"><a href="#cb1-525" aria-hidden="true" tabindex="-1"></a><span class="co"># Resume from a specific checkpoint dir</span></span>
-<span id="cb1-526"><a href="#cb1-526" aria-hidden="true" tabindex="-1"></a><span class="fu">resume_from_checkpoint</span><span class="kw">:</span></span>
-<span id="cb1-527"><a href="#cb1-527" aria-hidden="true" tabindex="-1"></a><span class="co"># If resume_from_checkpoint isn't set and you simply want it to start where it left off.</span></span>
-<span id="cb1-528"><a href="#cb1-528" aria-hidden="true" tabindex="-1"></a><span class="co"># Be careful with this being turned on between different models.</span></span>
-<span id="cb1-529"><a href="#cb1-529" aria-hidden="true" tabindex="-1"></a><span class="fu">auto_resume_from_checkpoints</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
-<span id="cb1-530"><a href="#cb1-530" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-531"><a href="#cb1-531" aria-hidden="true" tabindex="-1"></a><span class="co"># Don't mess with this, it's here for accelerate and torchrun</span></span>
-<span id="cb1-532"><a href="#cb1-532" aria-hidden="true" tabindex="-1"></a><span class="fu">local_rank</span><span class="kw">:</span></span>
-<span id="cb1-533"><a href="#cb1-533" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-534"><a href="#cb1-534" aria-hidden="true" tabindex="-1"></a><span class="co"># Add or change special tokens.</span></span>
-<span id="cb1-535"><a href="#cb1-535" aria-hidden="true" tabindex="-1"></a><span class="co"># If you add tokens here, you don't need to add them to the `tokens` list.</span></span>
-<span id="cb1-536"><a href="#cb1-536" aria-hidden="true" tabindex="-1"></a><span class="fu">special_tokens</span><span class="kw">:</span></span>
-<span id="cb1-537"><a href="#cb1-537" aria-hidden="true" tabindex="-1"></a><span class="co">  # bos_token: "&lt;s&gt;"</span></span>
-<span id="cb1-538"><a href="#cb1-538" aria-hidden="true" tabindex="-1"></a><span class="co">  # eos_token: "&lt;/s&gt;"</span></span>
-<span id="cb1-539"><a href="#cb1-539" aria-hidden="true" tabindex="-1"></a><span class="co">  # unk_token: "&lt;unk&gt;"</span></span>
-<span id="cb1-540"><a href="#cb1-540" aria-hidden="true" tabindex="-1"></a><span class="co">  # pad_token: "[PAD]"</span></span>
+<span id="cb1-447"><a href="#cb1-447" aria-hidden="true" tabindex="-1"></a><span class="co"># Specify a scheduler and kwargs to use with the optimizer</span></span>
+<span id="cb1-448"><a href="#cb1-448" aria-hidden="true" tabindex="-1"></a><span class="fu">lr_scheduler</span><span class="kw">:</span><span class="co"> # 'one_cycle' | 'log_sweep' | empty for cosine</span></span>
+<span id="cb1-449"><a href="#cb1-449" aria-hidden="true" tabindex="-1"></a><span class="fu">lr_scheduler_kwargs</span><span class="kw">:</span></span>
+<span id="cb1-450"><a href="#cb1-450" aria-hidden="true" tabindex="-1"></a><span class="fu">cosine_min_lr_ratio</span><span class="kw">:</span><span class="co"> # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr</span></span>
+<span id="cb1-451"><a href="#cb1-451" aria-hidden="true" tabindex="-1"></a><span class="fu">cosine_constant_lr_ratio</span><span class="kw">:</span><span class="co"> # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)</span></span>
+<span id="cb1-452"><a href="#cb1-452" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-453"><a href="#cb1-453" aria-hidden="true" tabindex="-1"></a><span class="co"># For one_cycle optim</span></span>
+<span id="cb1-454"><a href="#cb1-454" aria-hidden="true" tabindex="-1"></a><span class="fu">lr_div_factor</span><span class="kw">:</span><span class="co"> # Learning rate div factor</span></span>
+<span id="cb1-455"><a href="#cb1-455" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-456"><a href="#cb1-456" aria-hidden="true" tabindex="-1"></a><span class="co"># Specify optimizer</span></span>
+<span id="cb1-457"><a href="#cb1-457" aria-hidden="true" tabindex="-1"></a><span class="co"># Valid values are driven by the Transformers OptimizerNames class, see:</span></span>
+<span id="cb1-458"><a href="#cb1-458" aria-hidden="true" tabindex="-1"></a><span class="co"># https://github.com/huggingface/transformers/blob/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134</span></span>
+<span id="cb1-459"><a href="#cb1-459" aria-hidden="true" tabindex="-1"></a><span class="co">#</span></span>
+<span id="cb1-460"><a href="#cb1-460" aria-hidden="true" tabindex="-1"></a><span class="co"># Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of</span></span>
+<span id="cb1-461"><a href="#cb1-461" aria-hidden="true" tabindex="-1"></a><span class="co"># torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used</span></span>
+<span id="cb1-462"><a href="#cb1-462" aria-hidden="true" tabindex="-1"></a><span class="co"># in the examples/ for your model and fine-tuning use case.</span></span>
+<span id="cb1-463"><a href="#cb1-463" aria-hidden="true" tabindex="-1"></a><span class="co">#</span></span>
+<span id="cb1-464"><a href="#cb1-464" aria-hidden="true" tabindex="-1"></a><span class="co"># Valid values for 'optimizer' include:</span></span>
+<span id="cb1-465"><a href="#cb1-465" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_hf</span></span>
+<span id="cb1-466"><a href="#cb1-466" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_torch</span></span>
+<span id="cb1-467"><a href="#cb1-467" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_torch_fused</span></span>
+<span id="cb1-468"><a href="#cb1-468" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_torch_xla</span></span>
+<span id="cb1-469"><a href="#cb1-469" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_apex_fused</span></span>
+<span id="cb1-470"><a href="#cb1-470" aria-hidden="true" tabindex="-1"></a><span class="co"># - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version &gt;= 2.5.1)</span></span>
+<span id="cb1-471"><a href="#cb1-471" aria-hidden="true" tabindex="-1"></a><span class="co"># - adafactor</span></span>
+<span id="cb1-472"><a href="#cb1-472" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_anyprecision</span></span>
+<span id="cb1-473"><a href="#cb1-473" aria-hidden="true" tabindex="-1"></a><span class="co"># - sgd</span></span>
+<span id="cb1-474"><a href="#cb1-474" aria-hidden="true" tabindex="-1"></a><span class="co"># - adagrad</span></span>
+<span id="cb1-475"><a href="#cb1-475" aria-hidden="true" tabindex="-1"></a><span class="co"># - adamw_bnb_8bit</span></span>
+<span id="cb1-476"><a href="#cb1-476" aria-hidden="true" tabindex="-1"></a><span class="co"># - lion_8bit</span></span>
+<span id="cb1-477"><a href="#cb1-477" aria-hidden="true" tabindex="-1"></a><span class="co"># - lion_32bit</span></span>
+<span id="cb1-478"><a href="#cb1-478" aria-hidden="true" tabindex="-1"></a><span class="co"># - paged_adamw_32bit</span></span>
+<span id="cb1-479"><a href="#cb1-479" aria-hidden="true" tabindex="-1"></a><span class="co"># - paged_adamw_8bit</span></span>
+<span id="cb1-480"><a href="#cb1-480" aria-hidden="true" tabindex="-1"></a><span class="co"># - paged_lion_32bit</span></span>
+<span id="cb1-481"><a href="#cb1-481" aria-hidden="true" tabindex="-1"></a><span class="co"># - paged_lion_8bit</span></span>
+<span id="cb1-482"><a href="#cb1-482" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adamw</span></span>
+<span id="cb1-483"><a href="#cb1-483" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adamw_8bit</span></span>
+<span id="cb1-484"><a href="#cb1-484" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adafactor</span></span>
+<span id="cb1-485"><a href="#cb1-485" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adamw_layerwise</span></span>
+<span id="cb1-486"><a href="#cb1-486" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adamw_8bit_layerwise</span></span>
+<span id="cb1-487"><a href="#cb1-487" aria-hidden="true" tabindex="-1"></a><span class="co"># - galore_adafactor_layerwise</span></span>
+<span id="cb1-488"><a href="#cb1-488" aria-hidden="true" tabindex="-1"></a><span class="fu">optimizer</span><span class="kw">:</span></span>
+<span id="cb1-489"><a href="#cb1-489" aria-hidden="true" tabindex="-1"></a><span class="co"># Dictionary of arguments to pass to the optimizer</span></span>
+<span id="cb1-490"><a href="#cb1-490" aria-hidden="true" tabindex="-1"></a><span class="fu">optim_args</span><span class="kw">:</span></span>
+<span id="cb1-491"><a href="#cb1-491" aria-hidden="true" tabindex="-1"></a><span class="co"># For Galore Optimizers the following optim_args are available</span></span>
+<span id="cb1-492"><a href="#cb1-492" aria-hidden="true" tabindex="-1"></a><span class="co"># rank:  # type: int</span></span>
+<span id="cb1-493"><a href="#cb1-493" aria-hidden="true" tabindex="-1"></a><span class="co"># update_proj_gap  # type: int</span></span>
+<span id="cb1-494"><a href="#cb1-494" aria-hidden="true" tabindex="-1"></a><span class="co"># scale  # type: float</span></span>
+<span id="cb1-495"><a href="#cb1-495" aria-hidden="true" tabindex="-1"></a><span class="co"># proj_type:  # type: str, default = std</span></span>
+<span id="cb1-496"><a href="#cb1-496" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-497"><a href="#cb1-497" aria-hidden="true" tabindex="-1"></a><span class="co"># The target modules to optimize, i.e. the module names that you would like to train, right now this is used only for GaLore algorithm</span></span>
+<span id="cb1-498"><a href="#cb1-498" aria-hidden="true" tabindex="-1"></a><span class="fu">optim_target_modules</span><span class="kw">:</span></span>
+<span id="cb1-499"><a href="#cb1-499" aria-hidden="true" tabindex="-1"></a><span class="co"># - self_attn  # for llama</span></span>
+<span id="cb1-500"><a href="#cb1-500" aria-hidden="true" tabindex="-1"></a><span class="co"># - mlp</span></span>
+<span id="cb1-501"><a href="#cb1-501" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-502"><a href="#cb1-502" aria-hidden="true" tabindex="-1"></a><span class="co"># Specify weight decay</span></span>
+<span id="cb1-503"><a href="#cb1-503" aria-hidden="true" tabindex="-1"></a><span class="fu">weight_decay</span><span class="kw">:</span></span>
+<span id="cb1-504"><a href="#cb1-504" aria-hidden="true" tabindex="-1"></a><span class="co"># adamw hyperparams</span></span>
+<span id="cb1-505"><a href="#cb1-505" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_beta1</span><span class="kw">:</span></span>
+<span id="cb1-506"><a href="#cb1-506" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_beta2</span><span class="kw">:</span></span>
+<span id="cb1-507"><a href="#cb1-507" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_epsilon</span><span class="kw">:</span></span>
+<span id="cb1-508"><a href="#cb1-508" aria-hidden="true" tabindex="-1"></a><span class="co"># Gradient clipping max norm</span></span>
+<span id="cb1-509"><a href="#cb1-509" aria-hidden="true" tabindex="-1"></a><span class="fu">max_grad_norm</span><span class="kw">:</span></span>
+<span id="cb1-510"><a href="#cb1-510" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-511"><a href="#cb1-511" aria-hidden="true" tabindex="-1"></a><span class="co"># Augmentation techniques</span></span>
+<span id="cb1-512"><a href="#cb1-512" aria-hidden="true" tabindex="-1"></a><span class="co"># NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings</span></span>
+<span id="cb1-513"><a href="#cb1-513" aria-hidden="true" tabindex="-1"></a><span class="co"># currently only supported on Llama and Mistral</span></span>
+<span id="cb1-514"><a href="#cb1-514" aria-hidden="true" tabindex="-1"></a><span class="fu">neftune_noise_alpha</span><span class="kw">:</span></span>
+<span id="cb1-515"><a href="#cb1-515" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-516"><a href="#cb1-516" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to bettertransformers</span></span>
+<span id="cb1-517"><a href="#cb1-517" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_optimum</span><span class="kw">:</span></span>
+<span id="cb1-518"><a href="#cb1-518" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use xformers attention patch https://github.com/facebookresearch/xformers:</span></span>
+<span id="cb1-519"><a href="#cb1-519" aria-hidden="true" tabindex="-1"></a><span class="fu">xformers_attention</span><span class="kw">:</span></span>
+<span id="cb1-520"><a href="#cb1-520" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:</span></span>
+<span id="cb1-521"><a href="#cb1-521" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attention</span><span class="kw">:</span></span>
+<span id="cb1-522"><a href="#cb1-522" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_cross_entropy</span><span class="kw">:</span><span class="co">  # Whether to use flash-attention cross entropy implementation - advanced use only</span></span>
+<span id="cb1-523"><a href="#cb1-523" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_rms_norm</span><span class="kw">:</span><span class="co">  # Whether to use flash-attention rms norm implementation - advanced use only</span></span>
+<span id="cb1-524"><a href="#cb1-524" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_qkv</span><span class="kw">:</span><span class="co"> # Whether to fuse QKV into a single operation</span></span>
+<span id="cb1-525"><a href="#cb1-525" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_mlp</span><span class="kw">:</span><span class="co"> # Whether to fuse part of the MLP into a single operation</span></span>
+<span id="cb1-526"><a href="#cb1-526" aria-hidden="true" tabindex="-1"></a><span class="co"># Whether to use scaled-dot-product attention</span></span>
+<span id="cb1-527"><a href="#cb1-527" aria-hidden="true" tabindex="-1"></a><span class="co"># https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html</span></span>
+<span id="cb1-528"><a href="#cb1-528" aria-hidden="true" tabindex="-1"></a><span class="fu">sdp_attention</span><span class="kw">:</span></span>
+<span id="cb1-529"><a href="#cb1-529" aria-hidden="true" tabindex="-1"></a><span class="co"># Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf</span></span>
+<span id="cb1-530"><a href="#cb1-530" aria-hidden="true" tabindex="-1"></a><span class="fu">s2_attention</span><span class="kw">:</span></span>
+<span id="cb1-531"><a href="#cb1-531" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use low_cpu_mem_usage</span></span>
+<span id="cb1-532"><a href="#cb1-532" aria-hidden="true" tabindex="-1"></a><span class="fu">low_cpu_mem_usage</span><span class="kw">:</span></span>
+<span id="cb1-533"><a href="#cb1-533" aria-hidden="true" tabindex="-1"></a><span class="co"># Resume from a specific checkpoint dir</span></span>
+<span id="cb1-534"><a href="#cb1-534" aria-hidden="true" tabindex="-1"></a><span class="fu">resume_from_checkpoint</span><span class="kw">:</span></span>
+<span id="cb1-535"><a href="#cb1-535" aria-hidden="true" tabindex="-1"></a><span class="co"># If resume_from_checkpoint isn't set and you simply want it to start where it left off.</span></span>
+<span id="cb1-536"><a href="#cb1-536" aria-hidden="true" tabindex="-1"></a><span class="co"># Be careful with this being turned on between different models.</span></span>
+<span id="cb1-537"><a href="#cb1-537" aria-hidden="true" tabindex="-1"></a><span class="fu">auto_resume_from_checkpoints</span><span class="kw">:</span><span class="at"> </span><span class="ch">false</span></span>
+<span id="cb1-538"><a href="#cb1-538" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-539"><a href="#cb1-539" aria-hidden="true" tabindex="-1"></a><span class="co"># Don't mess with this, it's here for accelerate and torchrun</span></span>
+<span id="cb1-540"><a href="#cb1-540" aria-hidden="true" tabindex="-1"></a><span class="fu">local_rank</span><span class="kw">:</span></span>
 <span id="cb1-541"><a href="#cb1-541" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-542"><a href="#cb1-542" aria-hidden="true" tabindex="-1"></a><span class="co"># Add extra tokens.</span></span>
-<span id="cb1-543"><a href="#cb1-543" aria-hidden="true" tabindex="-1"></a><span class="fu">tokens</span><span class="kw">:</span></span>
-<span id="cb1-544"><a href="#cb1-544" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-545"><a href="#cb1-545" aria-hidden="true" tabindex="-1"></a><span class="co"># FSDP</span></span>
-<span id="cb1-546"><a href="#cb1-546" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp</span><span class="kw">:</span></span>
-<span id="cb1-547"><a href="#cb1-547" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp_config</span><span class="kw">:</span></span>
-<span id="cb1-548"><a href="#cb1-548" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-549"><a href="#cb1-549" aria-hidden="true" tabindex="-1"></a><span class="co"># Deepspeed config path. e.g., deepspeed_configs/zero3.json</span></span>
-<span id="cb1-550"><a href="#cb1-550" aria-hidden="true" tabindex="-1"></a><span class="fu">deepspeed</span><span class="kw">:</span></span>
-<span id="cb1-551"><a href="#cb1-551" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-552"><a href="#cb1-552" aria-hidden="true" tabindex="-1"></a><span class="co"># Advanced DDP Arguments</span></span>
-<span id="cb1-553"><a href="#cb1-553" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_timeout</span><span class="kw">:</span></span>
-<span id="cb1-554"><a href="#cb1-554" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_bucket_cap_mb</span><span class="kw">:</span></span>
-<span id="cb1-555"><a href="#cb1-555" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_broadcast_buffers</span><span class="kw">:</span></span>
+<span id="cb1-542"><a href="#cb1-542" aria-hidden="true" tabindex="-1"></a><span class="co"># Add or change special tokens.</span></span>
+<span id="cb1-543"><a href="#cb1-543" aria-hidden="true" tabindex="-1"></a><span class="co"># If you add tokens here, you don't need to add them to the `tokens` list.</span></span>
+<span id="cb1-544"><a href="#cb1-544" aria-hidden="true" tabindex="-1"></a><span class="fu">special_tokens</span><span class="kw">:</span></span>
+<span id="cb1-545"><a href="#cb1-545" aria-hidden="true" tabindex="-1"></a><span class="co">  # bos_token: "&lt;s&gt;"</span></span>
+<span id="cb1-546"><a href="#cb1-546" aria-hidden="true" tabindex="-1"></a><span class="co">  # eos_token: "&lt;/s&gt;"</span></span>
+<span id="cb1-547"><a href="#cb1-547" aria-hidden="true" tabindex="-1"></a><span class="co">  # unk_token: "&lt;unk&gt;"</span></span>
+<span id="cb1-548"><a href="#cb1-548" aria-hidden="true" tabindex="-1"></a><span class="co">  # pad_token: "[PAD]"</span></span>
+<span id="cb1-549"><a href="#cb1-549" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-550"><a href="#cb1-550" aria-hidden="true" tabindex="-1"></a><span class="co"># Add extra tokens.</span></span>
+<span id="cb1-551"><a href="#cb1-551" aria-hidden="true" tabindex="-1"></a><span class="fu">tokens</span><span class="kw">:</span></span>
+<span id="cb1-552"><a href="#cb1-552" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-553"><a href="#cb1-553" aria-hidden="true" tabindex="-1"></a><span class="co"># FSDP</span></span>
+<span id="cb1-554"><a href="#cb1-554" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp</span><span class="kw">:</span></span>
+<span id="cb1-555"><a href="#cb1-555" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp_config</span><span class="kw">:</span></span>
 <span id="cb1-556"><a href="#cb1-556" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-557"><a href="#cb1-557" aria-hidden="true" tabindex="-1"></a><span class="co"># Path to torch distx for optim 'adamw_anyprecision'</span></span>
-<span id="cb1-558"><a href="#cb1-558" aria-hidden="true" tabindex="-1"></a><span class="fu">torchdistx_path</span><span class="kw">:</span></span>
+<span id="cb1-557"><a href="#cb1-557" aria-hidden="true" tabindex="-1"></a><span class="co"># Deepspeed config path. e.g., deepspeed_configs/zero3.json</span></span>
+<span id="cb1-558"><a href="#cb1-558" aria-hidden="true" tabindex="-1"></a><span class="fu">deepspeed</span><span class="kw">:</span></span>
 <span id="cb1-559"><a href="#cb1-559" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-560"><a href="#cb1-560" aria-hidden="true" tabindex="-1"></a><span class="co"># Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize</span></span>
-<span id="cb1-561"><a href="#cb1-561" aria-hidden="true" tabindex="-1"></a><span class="fu">pretraining_dataset</span><span class="kw">:</span></span>
-<span id="cb1-562"><a href="#cb1-562" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-563"><a href="#cb1-563" aria-hidden="true" tabindex="-1"></a><span class="co"># Debug mode</span></span>
-<span id="cb1-564"><a href="#cb1-564" aria-hidden="true" tabindex="-1"></a><span class="fu">debug</span><span class="kw">:</span></span>
-<span id="cb1-565"><a href="#cb1-565" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-566"><a href="#cb1-566" aria-hidden="true" tabindex="-1"></a><span class="co"># Seed</span></span>
-<span id="cb1-567"><a href="#cb1-567" aria-hidden="true" tabindex="-1"></a><span class="fu">seed</span><span class="kw">:</span></span>
-<span id="cb1-568"><a href="#cb1-568" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-569"><a href="#cb1-569" aria-hidden="true" tabindex="-1"></a><span class="co"># Allow overwrite yml config using from cli</span></span>
-<span id="cb1-570"><a href="#cb1-570" aria-hidden="true" tabindex="-1"></a><span class="fu">strict</span><span class="kw">:</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<span id="cb1-560"><a href="#cb1-560" aria-hidden="true" tabindex="-1"></a><span class="co"># Advanced DDP Arguments</span></span>
+<span id="cb1-561"><a href="#cb1-561" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_timeout</span><span class="kw">:</span></span>
+<span id="cb1-562"><a href="#cb1-562" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_bucket_cap_mb</span><span class="kw">:</span></span>
+<span id="cb1-563"><a href="#cb1-563" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_broadcast_buffers</span><span class="kw">:</span></span>
+<span id="cb1-564"><a href="#cb1-564" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-565"><a href="#cb1-565" aria-hidden="true" tabindex="-1"></a><span class="co"># Path to torch distx for optim 'adamw_anyprecision'</span></span>
+<span id="cb1-566"><a href="#cb1-566" aria-hidden="true" tabindex="-1"></a><span class="fu">torchdistx_path</span><span class="kw">:</span></span>
+<span id="cb1-567"><a href="#cb1-567" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-568"><a href="#cb1-568" aria-hidden="true" tabindex="-1"></a><span class="co"># Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize</span></span>
+<span id="cb1-569"><a href="#cb1-569" aria-hidden="true" tabindex="-1"></a><span class="fu">pretraining_dataset</span><span class="kw">:</span></span>
+<span id="cb1-570"><a href="#cb1-570" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-571"><a href="#cb1-571" aria-hidden="true" tabindex="-1"></a><span class="co"># Debug mode</span></span>
+<span id="cb1-572"><a href="#cb1-572" aria-hidden="true" tabindex="-1"></a><span class="fu">debug</span><span class="kw">:</span></span>
+<span id="cb1-573"><a href="#cb1-573" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-574"><a href="#cb1-574" aria-hidden="true" tabindex="-1"></a><span class="co"># Seed</span></span>
+<span id="cb1-575"><a href="#cb1-575" aria-hidden="true" tabindex="-1"></a><span class="fu">seed</span><span class="kw">:</span></span>
+<span id="cb1-576"><a href="#cb1-576" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-577"><a href="#cb1-577" aria-hidden="true" tabindex="-1"></a><span class="co"># Allow overwrite yml config using from cli</span></span>
+<span id="cb1-578"><a href="#cb1-578" aria-hidden="true" tabindex="-1"></a><span class="fu">strict</span><span class="kw">:</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 
 
 
diff --git a/docs/custom_integrations.html b/docs/custom_integrations.html
index e077159ff..f5579bd25 100644
--- a/docs/custom_integrations.html
+++ b/docs/custom_integrations.html
@@ -240,6 +240,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
@@ -433,6 +439,8 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
   <ul>
   <li><a href="#cut-cross-entropy" id="toc-cut-cross-entropy" class="nav-link active" data-scroll-target="#cut-cross-entropy">Cut Cross Entropy</a>
   <ul class="collapse">
+  <li><a href="#requirements" id="toc-requirements" class="nav-link" data-scroll-target="#requirements">Requirements</a></li>
+  <li><a href="#installation" id="toc-installation" class="nav-link" data-scroll-target="#installation">Installation</a></li>
   <li><a href="#usage" id="toc-usage" class="nav-link" data-scroll-target="#usage">Usage</a></li>
   <li><a href="#citation" id="toc-citation" class="nav-link" data-scroll-target="#citation">Citation</a></li>
   </ul></li>
@@ -492,26 +500,39 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 <h2 class="anchored" data-anchor-id="cut-cross-entropy">Cut Cross Entropy</h2>
 <p>Cut Cross Entropy reduces VRAM usage through optimization on the cross-entropy operation during loss calculation.</p>
 <p>See https://github.com/apple/ml-cross-entropy</p>
+<section id="requirements" class="level3">
+<h3 class="anchored" data-anchor-id="requirements">Requirements</h3>
+<ul>
+<li>PyTorch 2.4.0 or higher</li>
+</ul>
+</section>
+<section id="installation" class="level3">
+<h3 class="anchored" data-anchor-id="installation">Installation</h3>
+<p>Run the following command to install <code>cut_cross_entropy[transformers]</code> if you don’t have it already.</p>
+<div class="sourceCode" id="cb1"><pre class="sourceCode bash code-with-copy"><code class="sourceCode bash"><span id="cb1-1"><a href="#cb1-1" aria-hidden="true" tabindex="-1"></a><span class="ex">python</span> scripts/cutcrossentropy_install.py <span class="kw">|</span> <span class="fu">sh</span></span>
+<span id="cb1-2"><a href="#cb1-2" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb1-3"><a href="#cb1-3" aria-hidden="true" tabindex="-1"></a><span class="ex">pip3</span> uninstall <span class="at">-y</span> cut-cross-entropy <span class="kw">&amp;&amp;</span> <span class="ex">pip3</span> install <span class="st">"cut-cross-entropy @ git+https://github.com/apple/ml-cross-entropy.git@9c297c905f55b73594b5d650722d1e78183b77bd"'</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+</section>
 <section id="usage" class="level3">
 <h3 class="anchored" data-anchor-id="usage">Usage</h3>
-<div class="sourceCode" id="cb1"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb1-1"><a href="#cb1-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
-<span id="cb1-2"><a href="#cb1-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin</span></span>
-<span id="cb1-3"><a href="#cb1-3" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb1-4"><a href="#cb1-4" aria-hidden="true" tabindex="-1"></a><span class="fu">cut_cross_entropy</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="sourceCode" id="cb2"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb2-1"><a href="#cb2-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
+<span id="cb2-2"><a href="#cb2-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin</span></span>
+<span id="cb2-3"><a href="#cb2-3" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb2-4"><a href="#cb2-4" aria-hidden="true" tabindex="-1"></a><span class="fu">cut_cross_entropy</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 </section>
 <section id="citation" class="level3">
 <h3 class="anchored" data-anchor-id="citation">Citation</h3>
-<div class="sourceCode" id="cb2"><pre class="sourceCode bib code-with-copy"><code class="sourceCode bibtex"><span id="cb2-1"><a href="#cb2-1" aria-hidden="true" tabindex="-1"></a><span class="va">@article</span>{<span class="ot">wijmans2024cut</span>,</span>
-<span id="cb2-2"><a href="#cb2-2" aria-hidden="true" tabindex="-1"></a>  <span class="dt">author</span>       = {Erik Wijmans and</span>
-<span id="cb2-3"><a href="#cb2-3" aria-hidden="true" tabindex="-1"></a>                  Brody Huval and</span>
-<span id="cb2-4"><a href="#cb2-4" aria-hidden="true" tabindex="-1"></a>                  Alexander Hertzberg and</span>
-<span id="cb2-5"><a href="#cb2-5" aria-hidden="true" tabindex="-1"></a>                  Vladlen Koltun and</span>
-<span id="cb2-6"><a href="#cb2-6" aria-hidden="true" tabindex="-1"></a>                  Philipp Kr<span class="ch">\"</span>ahenb<span class="ch">\"</span>uhl},</span>
-<span id="cb2-7"><a href="#cb2-7" aria-hidden="true" tabindex="-1"></a>  <span class="dt">title</span>        = {Cut Your Losses in Large-Vocabulary Language Models},</span>
-<span id="cb2-8"><a href="#cb2-8" aria-hidden="true" tabindex="-1"></a>  <span class="dt">journal</span>      = {arXiv},</span>
-<span id="cb2-9"><a href="#cb2-9" aria-hidden="true" tabindex="-1"></a>  <span class="dt">year</span>         = {2024},</span>
-<span id="cb2-10"><a href="#cb2-10" aria-hidden="true" tabindex="-1"></a>  <span class="dt">url</span>          = {https://arxiv.org/abs/2411.09009},</span>
-<span id="cb2-11"><a href="#cb2-11" aria-hidden="true" tabindex="-1"></a>}</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="sourceCode" id="cb3"><pre class="sourceCode bib code-with-copy"><code class="sourceCode bibtex"><span id="cb3-1"><a href="#cb3-1" aria-hidden="true" tabindex="-1"></a><span class="va">@article</span>{<span class="ot">wijmans2024cut</span>,</span>
+<span id="cb3-2"><a href="#cb3-2" aria-hidden="true" tabindex="-1"></a>  <span class="dt">author</span>       = {Erik Wijmans and</span>
+<span id="cb3-3"><a href="#cb3-3" aria-hidden="true" tabindex="-1"></a>                  Brody Huval and</span>
+<span id="cb3-4"><a href="#cb3-4" aria-hidden="true" tabindex="-1"></a>                  Alexander Hertzberg and</span>
+<span id="cb3-5"><a href="#cb3-5" aria-hidden="true" tabindex="-1"></a>                  Vladlen Koltun and</span>
+<span id="cb3-6"><a href="#cb3-6" aria-hidden="true" tabindex="-1"></a>                  Philipp Kr<span class="ch">\"</span>ahenb<span class="ch">\"</span>uhl},</span>
+<span id="cb3-7"><a href="#cb3-7" aria-hidden="true" tabindex="-1"></a>  <span class="dt">title</span>        = {Cut Your Losses in Large-Vocabulary Language Models},</span>
+<span id="cb3-8"><a href="#cb3-8" aria-hidden="true" tabindex="-1"></a>  <span class="dt">journal</span>      = {arXiv},</span>
+<span id="cb3-9"><a href="#cb3-9" aria-hidden="true" tabindex="-1"></a>  <span class="dt">year</span>         = {2024},</span>
+<span id="cb3-10"><a href="#cb3-10" aria-hidden="true" tabindex="-1"></a>  <span class="dt">url</span>          = {https://arxiv.org/abs/2411.09009},</span>
+<span id="cb3-11"><a href="#cb3-11" aria-hidden="true" tabindex="-1"></a>}</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 <p>Please see reference <a href="https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/cut_cross_entropy">here</a></p>
 </section>
 </section>
@@ -520,20 +541,20 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 <p>See https://github.com/ironjr/grokfast</p>
 <section id="usage-1" class="level3">
 <h3 class="anchored" data-anchor-id="usage-1">Usage</h3>
-<div class="sourceCode" id="cb3"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb3-1"><a href="#cb3-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
-<span id="cb3-2"><a href="#cb3-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> axolotl.integrations.grokfast.GrokfastPlugin</span></span>
-<span id="cb3-3"><a href="#cb3-3" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb3-4"><a href="#cb3-4" aria-hidden="true" tabindex="-1"></a><span class="fu">grokfast_alpha</span><span class="kw">:</span><span class="at"> </span><span class="fl">2.0</span></span>
-<span id="cb3-5"><a href="#cb3-5" aria-hidden="true" tabindex="-1"></a><span class="fu">grokfast_lamb</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.98</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="sourceCode" id="cb4"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb4-1"><a href="#cb4-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
+<span id="cb4-2"><a href="#cb4-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> axolotl.integrations.grokfast.GrokfastPlugin</span></span>
+<span id="cb4-3"><a href="#cb4-3" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb4-4"><a href="#cb4-4" aria-hidden="true" tabindex="-1"></a><span class="fu">grokfast_alpha</span><span class="kw">:</span><span class="at"> </span><span class="fl">2.0</span></span>
+<span id="cb4-5"><a href="#cb4-5" aria-hidden="true" tabindex="-1"></a><span class="fu">grokfast_lamb</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.98</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 </section>
 <section id="citation-1" class="level3">
 <h3 class="anchored" data-anchor-id="citation-1">Citation</h3>
-<div class="sourceCode" id="cb4"><pre class="sourceCode bib code-with-copy"><code class="sourceCode bibtex"><span id="cb4-1"><a href="#cb4-1" aria-hidden="true" tabindex="-1"></a><span class="va">@article</span>{<span class="ot">lee2024grokfast</span>,</span>
-<span id="cb4-2"><a href="#cb4-2" aria-hidden="true" tabindex="-1"></a>    <span class="dt">title</span>={{Grokfast}: Accelerated Grokking by Amplifying Slow Gradients},</span>
-<span id="cb4-3"><a href="#cb4-3" aria-hidden="true" tabindex="-1"></a>    <span class="dt">author</span>={Lee, Jaerin and Kang, Bong Gyun and Kim, Kihoon and Lee, Kyoung Mu},</span>
-<span id="cb4-4"><a href="#cb4-4" aria-hidden="true" tabindex="-1"></a>    <span class="dt">journal</span>={arXiv preprint arXiv:2405.20233},</span>
-<span id="cb4-5"><a href="#cb4-5" aria-hidden="true" tabindex="-1"></a>    <span class="dt">year</span>={2024}</span>
-<span id="cb4-6"><a href="#cb4-6" aria-hidden="true" tabindex="-1"></a>}</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="sourceCode" id="cb5"><pre class="sourceCode bib code-with-copy"><code class="sourceCode bibtex"><span id="cb5-1"><a href="#cb5-1" aria-hidden="true" tabindex="-1"></a><span class="va">@article</span>{<span class="ot">lee2024grokfast</span>,</span>
+<span id="cb5-2"><a href="#cb5-2" aria-hidden="true" tabindex="-1"></a>    <span class="dt">title</span>={{Grokfast}: Accelerated Grokking by Amplifying Slow Gradients},</span>
+<span id="cb5-3"><a href="#cb5-3" aria-hidden="true" tabindex="-1"></a>    <span class="dt">author</span>={Lee, Jaerin and Kang, Bong Gyun and Kim, Kihoon and Lee, Kyoung Mu},</span>
+<span id="cb5-4"><a href="#cb5-4" aria-hidden="true" tabindex="-1"></a>    <span class="dt">journal</span>={arXiv preprint arXiv:2405.20233},</span>
+<span id="cb5-5"><a href="#cb5-5" aria-hidden="true" tabindex="-1"></a>    <span class="dt">year</span>={2024}</span>
+<span id="cb5-6"><a href="#cb5-6" aria-hidden="true" tabindex="-1"></a>}</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 <p>Please see reference <a href="https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/grokfast">here</a></p>
 </section>
 </section>
@@ -541,21 +562,21 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 <h2 class="anchored" data-anchor-id="knowledge-distillation-kd">Knowledge Distillation (KD)</h2>
 <section id="usage-2" class="level3">
 <h3 class="anchored" data-anchor-id="usage-2">Usage</h3>
-<div class="sourceCode" id="cb5"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb5-1"><a href="#cb5-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
-<span id="cb5-2"><a href="#cb5-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> </span><span class="st">"axolotl.integrations.kd.KDPlugin"</span></span>
-<span id="cb5-3"><a href="#cb5-3" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb5-4"><a href="#cb5-4" aria-hidden="true" tabindex="-1"></a><span class="fu">kd_trainer</span><span class="kw">:</span><span class="at"> </span><span class="ch">True</span></span>
-<span id="cb5-5"><a href="#cb5-5" aria-hidden="true" tabindex="-1"></a><span class="fu">kd_ce_alpha</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.1</span></span>
-<span id="cb5-6"><a href="#cb5-6" aria-hidden="true" tabindex="-1"></a><span class="fu">kd_alpha</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.9</span></span>
-<span id="cb5-7"><a href="#cb5-7" aria-hidden="true" tabindex="-1"></a><span class="fu">kd_temperature</span><span class="kw">:</span><span class="at"> </span><span class="fl">1.0</span></span>
-<span id="cb5-8"><a href="#cb5-8" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb5-9"><a href="#cb5-9" aria-hidden="true" tabindex="-1"></a><span class="fu">torch_compile</span><span class="kw">:</span><span class="at"> </span><span class="ch">True</span><span class="co">  # torch&gt;=2.5.1, recommended to reduce vram</span></span>
-<span id="cb5-10"><a href="#cb5-10" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb5-11"><a href="#cb5-11" aria-hidden="true" tabindex="-1"></a><span class="fu">datasets</span><span class="kw">:</span></span>
-<span id="cb5-12"><a href="#cb5-12" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> </span><span class="fu">path</span><span class="kw">:</span><span class="at"> ...</span></span>
-<span id="cb5-13"><a href="#cb5-13" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">type</span><span class="kw">:</span><span class="at"> </span><span class="st">"axolotl.integrations.kd.chat_template"</span></span>
-<span id="cb5-14"><a href="#cb5-14" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">field_messages</span><span class="kw">:</span><span class="at"> </span><span class="st">"messages_combined"</span></span>
-<span id="cb5-15"><a href="#cb5-15" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">logprobs_field</span><span class="kw">:</span><span class="at"> </span><span class="st">"llm_text_generation_vllm_logprobs"</span><span class="co">  # for kd only, field of logprobs</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="sourceCode" id="cb6"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb6-1"><a href="#cb6-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
+<span id="cb6-2"><a href="#cb6-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> </span><span class="st">"axolotl.integrations.kd.KDPlugin"</span></span>
+<span id="cb6-3"><a href="#cb6-3" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb6-4"><a href="#cb6-4" aria-hidden="true" tabindex="-1"></a><span class="fu">kd_trainer</span><span class="kw">:</span><span class="at"> </span><span class="ch">True</span></span>
+<span id="cb6-5"><a href="#cb6-5" aria-hidden="true" tabindex="-1"></a><span class="fu">kd_ce_alpha</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.1</span></span>
+<span id="cb6-6"><a href="#cb6-6" aria-hidden="true" tabindex="-1"></a><span class="fu">kd_alpha</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.9</span></span>
+<span id="cb6-7"><a href="#cb6-7" aria-hidden="true" tabindex="-1"></a><span class="fu">kd_temperature</span><span class="kw">:</span><span class="at"> </span><span class="fl">1.0</span></span>
+<span id="cb6-8"><a href="#cb6-8" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb6-9"><a href="#cb6-9" aria-hidden="true" tabindex="-1"></a><span class="fu">torch_compile</span><span class="kw">:</span><span class="at"> </span><span class="ch">True</span><span class="co">  # torch&gt;=2.5.1, recommended to reduce vram</span></span>
+<span id="cb6-10"><a href="#cb6-10" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb6-11"><a href="#cb6-11" aria-hidden="true" tabindex="-1"></a><span class="fu">datasets</span><span class="kw">:</span></span>
+<span id="cb6-12"><a href="#cb6-12" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> </span><span class="fu">path</span><span class="kw">:</span><span class="at"> ...</span></span>
+<span id="cb6-13"><a href="#cb6-13" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">type</span><span class="kw">:</span><span class="at"> </span><span class="st">"axolotl.integrations.kd.chat_template"</span></span>
+<span id="cb6-14"><a href="#cb6-14" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">field_messages</span><span class="kw">:</span><span class="at"> </span><span class="st">"messages_combined"</span></span>
+<span id="cb6-15"><a href="#cb6-15" aria-hidden="true" tabindex="-1"></a><span class="at">    </span><span class="fu">logprobs_field</span><span class="kw">:</span><span class="at"> </span><span class="st">"llm_text_generation_vllm_logprobs"</span><span class="co">  # for kd only, field of logprobs</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 <p>An example dataset can be found at <a href="https://huggingface.co/datasets/axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample"><code>axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample</code></a></p>
 <p>Please see reference <a href="https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/kd">here</a></p>
 </section>
@@ -571,26 +592,26 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 <p>See https://github.com/linkedin/Liger-Kernel</p>
 <section id="usage-3" class="level3">
 <h3 class="anchored" data-anchor-id="usage-3">Usage</h3>
-<div class="sourceCode" id="cb6"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb6-1"><a href="#cb6-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
-<span id="cb6-2"><a href="#cb6-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> axolotl.integrations.liger.LigerPlugin</span></span>
-<span id="cb6-3"><a href="#cb6-3" aria-hidden="true" tabindex="-1"></a><span class="fu">liger_rope</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
-<span id="cb6-4"><a href="#cb6-4" aria-hidden="true" tabindex="-1"></a><span class="fu">liger_rms_norm</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
-<span id="cb6-5"><a href="#cb6-5" aria-hidden="true" tabindex="-1"></a><span class="fu">liger_glu_activation</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
-<span id="cb6-6"><a href="#cb6-6" aria-hidden="true" tabindex="-1"></a><span class="fu">liger_layer_norm</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
-<span id="cb6-7"><a href="#cb6-7" aria-hidden="true" tabindex="-1"></a><span class="fu">liger_fused_linear_cross_entropy</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="sourceCode" id="cb7"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb7-1"><a href="#cb7-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
+<span id="cb7-2"><a href="#cb7-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> axolotl.integrations.liger.LigerPlugin</span></span>
+<span id="cb7-3"><a href="#cb7-3" aria-hidden="true" tabindex="-1"></a><span class="fu">liger_rope</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
+<span id="cb7-4"><a href="#cb7-4" aria-hidden="true" tabindex="-1"></a><span class="fu">liger_rms_norm</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
+<span id="cb7-5"><a href="#cb7-5" aria-hidden="true" tabindex="-1"></a><span class="fu">liger_glu_activation</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
+<span id="cb7-6"><a href="#cb7-6" aria-hidden="true" tabindex="-1"></a><span class="fu">liger_layer_norm</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
+<span id="cb7-7"><a href="#cb7-7" aria-hidden="true" tabindex="-1"></a><span class="fu">liger_fused_linear_cross_entropy</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 </section>
 <section id="citation-2" class="level3">
 <h3 class="anchored" data-anchor-id="citation-2">Citation</h3>
-<div class="sourceCode" id="cb7"><pre class="sourceCode bib code-with-copy"><code class="sourceCode bibtex"><span id="cb7-1"><a href="#cb7-1" aria-hidden="true" tabindex="-1"></a><span class="va">@article</span>{<span class="ot">hsu2024ligerkernelefficienttriton</span>,</span>
-<span id="cb7-2"><a href="#cb7-2" aria-hidden="true" tabindex="-1"></a>      <span class="dt">title</span>={Liger Kernel: Efficient Triton Kernels for LLM Training},</span>
-<span id="cb7-3"><a href="#cb7-3" aria-hidden="true" tabindex="-1"></a>      <span class="dt">author</span>={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},</span>
-<span id="cb7-4"><a href="#cb7-4" aria-hidden="true" tabindex="-1"></a>      <span class="dt">year</span>={2024},</span>
-<span id="cb7-5"><a href="#cb7-5" aria-hidden="true" tabindex="-1"></a>      <span class="dt">eprint</span>={2410.10989},</span>
-<span id="cb7-6"><a href="#cb7-6" aria-hidden="true" tabindex="-1"></a>      <span class="dt">archivePrefix</span>={arXiv},</span>
-<span id="cb7-7"><a href="#cb7-7" aria-hidden="true" tabindex="-1"></a>      <span class="dt">primaryClass</span>={cs.LG},</span>
-<span id="cb7-8"><a href="#cb7-8" aria-hidden="true" tabindex="-1"></a>      <span class="dt">url</span>={https://arxiv.org/abs/2410.10989},</span>
-<span id="cb7-9"><a href="#cb7-9" aria-hidden="true" tabindex="-1"></a>      <span class="dt">journal</span>={arXiv preprint arXiv:2410.10989},</span>
-<span id="cb7-10"><a href="#cb7-10" aria-hidden="true" tabindex="-1"></a>}</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="sourceCode" id="cb8"><pre class="sourceCode bib code-with-copy"><code class="sourceCode bibtex"><span id="cb8-1"><a href="#cb8-1" aria-hidden="true" tabindex="-1"></a><span class="va">@article</span>{<span class="ot">hsu2024ligerkernelefficienttriton</span>,</span>
+<span id="cb8-2"><a href="#cb8-2" aria-hidden="true" tabindex="-1"></a>      <span class="dt">title</span>={Liger Kernel: Efficient Triton Kernels for LLM Training},</span>
+<span id="cb8-3"><a href="#cb8-3" aria-hidden="true" tabindex="-1"></a>      <span class="dt">author</span>={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},</span>
+<span id="cb8-4"><a href="#cb8-4" aria-hidden="true" tabindex="-1"></a>      <span class="dt">year</span>={2024},</span>
+<span id="cb8-5"><a href="#cb8-5" aria-hidden="true" tabindex="-1"></a>      <span class="dt">eprint</span>={2410.10989},</span>
+<span id="cb8-6"><a href="#cb8-6" aria-hidden="true" tabindex="-1"></a>      <span class="dt">archivePrefix</span>={arXiv},</span>
+<span id="cb8-7"><a href="#cb8-7" aria-hidden="true" tabindex="-1"></a>      <span class="dt">primaryClass</span>={cs.LG},</span>
+<span id="cb8-8"><a href="#cb8-8" aria-hidden="true" tabindex="-1"></a>      <span class="dt">url</span>={https://arxiv.org/abs/2410.10989},</span>
+<span id="cb8-9"><a href="#cb8-9" aria-hidden="true" tabindex="-1"></a>      <span class="dt">journal</span>={arXiv preprint arXiv:2410.10989},</span>
+<span id="cb8-10"><a href="#cb8-10" aria-hidden="true" tabindex="-1"></a>}</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 <p>Please see reference <a href="https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/liger">here</a></p>
 </section>
 </section>
@@ -600,29 +621,29 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 <p>See https://github.com/EleutherAI/lm-evaluation-harness</p>
 <section id="usage-4" class="level3">
 <h3 class="anchored" data-anchor-id="usage-4">Usage</h3>
-<div class="sourceCode" id="cb8"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb8-1"><a href="#cb8-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
-<span id="cb8-2"><a href="#cb8-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> axolotl.integrations.lm_eval.LMEvalPlugin</span></span>
-<span id="cb8-3"><a href="#cb8-3" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb8-4"><a href="#cb8-4" aria-hidden="true" tabindex="-1"></a><span class="fu">lm_eval_tasks</span><span class="kw">:</span></span>
-<span id="cb8-5"><a href="#cb8-5" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> gsm8k</span></span>
-<span id="cb8-6"><a href="#cb8-6" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> hellaswag</span></span>
-<span id="cb8-7"><a href="#cb8-7" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> arc_easy</span></span>
-<span id="cb8-8"><a href="#cb8-8" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb8-9"><a href="#cb8-9" aria-hidden="true" tabindex="-1"></a><span class="fu">lm_eval_batch_size</span><span class="kw">:</span><span class="co"> # Batch size for evaluation</span></span>
-<span id="cb8-10"><a href="#cb8-10" aria-hidden="true" tabindex="-1"></a><span class="fu">output_dir</span><span class="kw">:</span><span class="co"> # Directory to save evaluation results</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="sourceCode" id="cb9"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb9-1"><a href="#cb9-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
+<span id="cb9-2"><a href="#cb9-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> axolotl.integrations.lm_eval.LMEvalPlugin</span></span>
+<span id="cb9-3"><a href="#cb9-3" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb9-4"><a href="#cb9-4" aria-hidden="true" tabindex="-1"></a><span class="fu">lm_eval_tasks</span><span class="kw">:</span></span>
+<span id="cb9-5"><a href="#cb9-5" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> gsm8k</span></span>
+<span id="cb9-6"><a href="#cb9-6" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> hellaswag</span></span>
+<span id="cb9-7"><a href="#cb9-7" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> arc_easy</span></span>
+<span id="cb9-8"><a href="#cb9-8" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb9-9"><a href="#cb9-9" aria-hidden="true" tabindex="-1"></a><span class="fu">lm_eval_batch_size</span><span class="kw">:</span><span class="co"> # Batch size for evaluation</span></span>
+<span id="cb9-10"><a href="#cb9-10" aria-hidden="true" tabindex="-1"></a><span class="fu">output_dir</span><span class="kw">:</span><span class="co"> # Directory to save evaluation results</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 </section>
 <section id="citation-3" class="level3">
 <h3 class="anchored" data-anchor-id="citation-3">Citation</h3>
-<div class="sourceCode" id="cb9"><pre class="sourceCode bib code-with-copy"><code class="sourceCode bibtex"><span id="cb9-1"><a href="#cb9-1" aria-hidden="true" tabindex="-1"></a><span class="va">@misc</span>{<span class="ot">eval</span>-<span class="ot">harness</span>,</span>
-<span id="cb9-2"><a href="#cb9-2" aria-hidden="true" tabindex="-1"></a>  <span class="dt">author</span>       = {Gao, Leo and Tow, Jonathan and Abbasi, Baber and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and Le Noac'h, Alain and Li, Haonan and McDonell, Kyle and Muennighoff, Niklas and Ociepa, Chris and Phang, Jason and Reynolds, Laria and Schoelkopf, Hailey and Skowron, Aviya and Sutawika, Lintang and Tang, Eric and Thite, Anish and Wang, Ben and Wang, Kevin and Zou, Andy},</span>
-<span id="cb9-3"><a href="#cb9-3" aria-hidden="true" tabindex="-1"></a>  <span class="dt">title</span>        = {A framework for few-shot language model evaluation},</span>
-<span id="cb9-4"><a href="#cb9-4" aria-hidden="true" tabindex="-1"></a>  <span class="dt">month</span>        = 07,</span>
-<span id="cb9-5"><a href="#cb9-5" aria-hidden="true" tabindex="-1"></a>  <span class="dt">year</span>         = 2024,</span>
-<span id="cb9-6"><a href="#cb9-6" aria-hidden="true" tabindex="-1"></a>  <span class="dt">publisher</span>    = {Zenodo},</span>
-<span id="cb9-7"><a href="#cb9-7" aria-hidden="true" tabindex="-1"></a>  <span class="dt">version</span>      = {v0.4.3},</span>
-<span id="cb9-8"><a href="#cb9-8" aria-hidden="true" tabindex="-1"></a>  <span class="dt">doi</span>          = {10.5281/zenodo.12608602},</span>
-<span id="cb9-9"><a href="#cb9-9" aria-hidden="true" tabindex="-1"></a>  <span class="dt">url</span>          = {https://zenodo.org/records/12608602}</span>
-<span id="cb9-10"><a href="#cb9-10" aria-hidden="true" tabindex="-1"></a>}</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="sourceCode" id="cb10"><pre class="sourceCode bib code-with-copy"><code class="sourceCode bibtex"><span id="cb10-1"><a href="#cb10-1" aria-hidden="true" tabindex="-1"></a><span class="va">@misc</span>{<span class="ot">eval</span>-<span class="ot">harness</span>,</span>
+<span id="cb10-2"><a href="#cb10-2" aria-hidden="true" tabindex="-1"></a>  <span class="dt">author</span>       = {Gao, Leo and Tow, Jonathan and Abbasi, Baber and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and Le Noac'h, Alain and Li, Haonan and McDonell, Kyle and Muennighoff, Niklas and Ociepa, Chris and Phang, Jason and Reynolds, Laria and Schoelkopf, Hailey and Skowron, Aviya and Sutawika, Lintang and Tang, Eric and Thite, Anish and Wang, Ben and Wang, Kevin and Zou, Andy},</span>
+<span id="cb10-3"><a href="#cb10-3" aria-hidden="true" tabindex="-1"></a>  <span class="dt">title</span>        = {A framework for few-shot language model evaluation},</span>
+<span id="cb10-4"><a href="#cb10-4" aria-hidden="true" tabindex="-1"></a>  <span class="dt">month</span>        = 07,</span>
+<span id="cb10-5"><a href="#cb10-5" aria-hidden="true" tabindex="-1"></a>  <span class="dt">year</span>         = 2024,</span>
+<span id="cb10-6"><a href="#cb10-6" aria-hidden="true" tabindex="-1"></a>  <span class="dt">publisher</span>    = {Zenodo},</span>
+<span id="cb10-7"><a href="#cb10-7" aria-hidden="true" tabindex="-1"></a>  <span class="dt">version</span>      = {v0.4.3},</span>
+<span id="cb10-8"><a href="#cb10-8" aria-hidden="true" tabindex="-1"></a>  <span class="dt">doi</span>          = {10.5281/zenodo.12608602},</span>
+<span id="cb10-9"><a href="#cb10-9" aria-hidden="true" tabindex="-1"></a>  <span class="dt">url</span>          = {https://zenodo.org/records/12608602}</span>
+<span id="cb10-10"><a href="#cb10-10" aria-hidden="true" tabindex="-1"></a>}</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 <p>Please see reference <a href="https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/lm_eval">here</a></p>
 </section>
 </section>
@@ -637,23 +658,23 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 </section>
 <section id="usage-5" class="level3">
 <h3 class="anchored" data-anchor-id="usage-5">Usage</h3>
-<div class="sourceCode" id="cb10"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb10-1"><a href="#cb10-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
-<span id="cb10-2"><a href="#cb10-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> axolotl.integrations.spectrum.SpectrumPlugin</span></span>
-<span id="cb10-3"><a href="#cb10-3" aria-hidden="true" tabindex="-1"></a></span>
-<span id="cb10-4"><a href="#cb10-4" aria-hidden="true" tabindex="-1"></a><span class="fu">spectrum_top_fraction</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.5</span></span>
-<span id="cb10-5"><a href="#cb10-5" aria-hidden="true" tabindex="-1"></a><span class="fu">spectrum_model_name</span><span class="kw">:</span><span class="at"> meta-llama/Meta-Llama-3.1-8B</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="sourceCode" id="cb11"><pre class="sourceCode yaml code-with-copy"><code class="sourceCode yaml"><span id="cb11-1"><a href="#cb11-1" aria-hidden="true" tabindex="-1"></a><span class="fu">plugins</span><span class="kw">:</span></span>
+<span id="cb11-2"><a href="#cb11-2" aria-hidden="true" tabindex="-1"></a><span class="at">  </span><span class="kw">-</span><span class="at"> axolotl.integrations.spectrum.SpectrumPlugin</span></span>
+<span id="cb11-3"><a href="#cb11-3" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb11-4"><a href="#cb11-4" aria-hidden="true" tabindex="-1"></a><span class="fu">spectrum_top_fraction</span><span class="kw">:</span><span class="at"> </span><span class="fl">0.5</span></span>
+<span id="cb11-5"><a href="#cb11-5" aria-hidden="true" tabindex="-1"></a><span class="fu">spectrum_model_name</span><span class="kw">:</span><span class="at"> meta-llama/Meta-Llama-3.1-8B</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 </section>
 <section id="citation-4" class="level3">
 <h3 class="anchored" data-anchor-id="citation-4">Citation</h3>
-<div class="sourceCode" id="cb11"><pre class="sourceCode bib code-with-copy"><code class="sourceCode bibtex"><span id="cb11-1"><a href="#cb11-1" aria-hidden="true" tabindex="-1"></a><span class="va">@misc</span>{<span class="ot">hartford2024spectrumtargetedtrainingsignal</span>,</span>
-<span id="cb11-2"><a href="#cb11-2" aria-hidden="true" tabindex="-1"></a>      <span class="dt">title</span>={Spectrum: Targeted Training on Signal to Noise Ratio},</span>
-<span id="cb11-3"><a href="#cb11-3" aria-hidden="true" tabindex="-1"></a>      <span class="dt">author</span>={Eric Hartford and Lucas Atkins and Fernando Fernandes Neto and David Golchinfar},</span>
-<span id="cb11-4"><a href="#cb11-4" aria-hidden="true" tabindex="-1"></a>      <span class="dt">year</span>={2024},</span>
-<span id="cb11-5"><a href="#cb11-5" aria-hidden="true" tabindex="-1"></a>      <span class="dt">eprint</span>={2406.06623},</span>
-<span id="cb11-6"><a href="#cb11-6" aria-hidden="true" tabindex="-1"></a>      <span class="dt">archivePrefix</span>={arXiv},</span>
-<span id="cb11-7"><a href="#cb11-7" aria-hidden="true" tabindex="-1"></a>      <span class="dt">primaryClass</span>={cs.LG},</span>
-<span id="cb11-8"><a href="#cb11-8" aria-hidden="true" tabindex="-1"></a>      <span class="dt">url</span>={https://arxiv.org/abs/2406.06623},</span>
-<span id="cb11-9"><a href="#cb11-9" aria-hidden="true" tabindex="-1"></a>}</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="sourceCode" id="cb12"><pre class="sourceCode bib code-with-copy"><code class="sourceCode bibtex"><span id="cb12-1"><a href="#cb12-1" aria-hidden="true" tabindex="-1"></a><span class="va">@misc</span>{<span class="ot">hartford2024spectrumtargetedtrainingsignal</span>,</span>
+<span id="cb12-2"><a href="#cb12-2" aria-hidden="true" tabindex="-1"></a>      <span class="dt">title</span>={Spectrum: Targeted Training on Signal to Noise Ratio},</span>
+<span id="cb12-3"><a href="#cb12-3" aria-hidden="true" tabindex="-1"></a>      <span class="dt">author</span>={Eric Hartford and Lucas Atkins and Fernando Fernandes Neto and David Golchinfar},</span>
+<span id="cb12-4"><a href="#cb12-4" aria-hidden="true" tabindex="-1"></a>      <span class="dt">year</span>={2024},</span>
+<span id="cb12-5"><a href="#cb12-5" aria-hidden="true" tabindex="-1"></a>      <span class="dt">eprint</span>={2406.06623},</span>
+<span id="cb12-6"><a href="#cb12-6" aria-hidden="true" tabindex="-1"></a>      <span class="dt">archivePrefix</span>={arXiv},</span>
+<span id="cb12-7"><a href="#cb12-7" aria-hidden="true" tabindex="-1"></a>      <span class="dt">primaryClass</span>={cs.LG},</span>
+<span id="cb12-8"><a href="#cb12-8" aria-hidden="true" tabindex="-1"></a>      <span class="dt">url</span>={https://arxiv.org/abs/2406.06623},</span>
+<span id="cb12-9"><a href="#cb12-9" aria-hidden="true" tabindex="-1"></a>}</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 <p>Please see reference <a href="https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/spectrum">here</a></p>
 
 
diff --git a/docs/dataset-formats/conversation.html b/docs/dataset-formats/conversation.html
index ea7b14d3d..f335a0565 100644
--- a/docs/dataset-formats/conversation.html
+++ b/docs/dataset-formats/conversation.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/dataset-formats/index.html b/docs/dataset-formats/index.html
index 899377ae4..d82870a5b 100644
--- a/docs/dataset-formats/index.html
+++ b/docs/dataset-formats/index.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
@@ -602,7 +608,12 @@ Tip
 <section id="pre-tokenized-dataset" class="level3">
 <h3 class="anchored" data-anchor-id="pre-tokenized-dataset">Pre-Tokenized Dataset</h3>
 <p>We suggest this approach when you want to bring your own tokenized dataset.</p>
-<p>Axolotl expects the dataset to have three keys: - <code>input_ids</code>: from tokenizing formatted prompt - <code>attention_mask</code>: for masking padding. If you don’t add padding, it would be equal to <code>len(input_ids) * [1]</code> - <code>labels</code>: this is the same as <code>input_ids</code>, however, if you want to mask certain tokens, you would set those indices to <code>-100</code>.</p>
+<p>Axolotl expects the dataset to have three keys:</p>
+<ul>
+<li><code>input_ids</code>: from tokenizing formatted prompt</li>
+<li><code>attention_mask</code>: for masking padding. If you don’t add padding, it would be equal to <code>len(input_ids) * [1]</code></li>
+<li><code>labels</code>: this is the same as <code>input_ids</code>, however, if you want to mask certain tokens, you would set those indices to <code>-100</code>.</li>
+</ul>
 <div class="callout callout-style-default callout-tip callout-titled">
 <div class="callout-header d-flex align-content-center">
 <div class="callout-icon-container">
diff --git a/docs/dataset-formats/inst_tune.html b/docs/dataset-formats/inst_tune.html
index 341ecc539..b3c75efd5 100644
--- a/docs/dataset-formats/inst_tune.html
+++ b/docs/dataset-formats/inst_tune.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/dataset-formats/pretraining.html b/docs/dataset-formats/pretraining.html
index bfc42278c..f56e001a0 100644
--- a/docs/dataset-formats/pretraining.html
+++ b/docs/dataset-formats/pretraining.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/dataset-formats/stepwise_supervised.html b/docs/dataset-formats/stepwise_supervised.html
index 765753072..0f1e13317 100644
--- a/docs/dataset-formats/stepwise_supervised.html
+++ b/docs/dataset-formats/stepwise_supervised.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/dataset-formats/template_free.html b/docs/dataset-formats/template_free.html
index ede75609e..2676dcb13 100644
--- a/docs/dataset-formats/template_free.html
+++ b/docs/dataset-formats/template_free.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/dataset-formats/tokenized.html b/docs/dataset-formats/tokenized.html
index 9cfa6add6..8e21f9f5d 100644
--- a/docs/dataset-formats/tokenized.html
+++ b/docs/dataset-formats/tokenized.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/dataset_preprocessing.html b/docs/dataset_preprocessing.html
index 93ba941aa..f2910f1db 100644
--- a/docs/dataset_preprocessing.html
+++ b/docs/dataset_preprocessing.html
@@ -207,6 +207,12 @@ ul.task-list li input[type="checkbox"] {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/debugging.html b/docs/debugging.html
index 7a3382a55..b07f03f4c 100644
--- a/docs/debugging.html
+++ b/docs/debugging.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/docker.html b/docs/docker.html
new file mode 100644
index 000000000..ed36a70e9
--- /dev/null
+++ b/docs/docker.html
@@ -0,0 +1,1061 @@
+<!DOCTYPE html>
+<html xmlns="http://www.w3.org/1999/xhtml" lang="en" xml:lang="en"><head>
+
+<meta charset="utf-8">
+<meta name="generator" content="quarto-1.6.42">
+
+<meta name="viewport" content="width=device-width, initial-scale=1.0, user-scalable=yes">
+
+
+<title>Docker – Axolotl</title>
+<style>
+code{white-space: pre-wrap;}
+span.smallcaps{font-variant: small-caps;}
+div.columns{display: flex; gap: min(4vw, 1.5em);}
+div.column{flex: auto; overflow-x: auto;}
+div.hanging-indent{margin-left: 1.5em; text-indent: -1.5em;}
+ul.task-list{list-style: none;}
+ul.task-list li input[type="checkbox"] {
+  width: 0.8em;
+  margin: 0 0.8em 0.2em -1em; /* quarto-specific, see https://github.com/quarto-dev/quarto-cli/issues/4556 */ 
+  vertical-align: middle;
+}
+/* CSS for syntax highlighting */
+pre > code.sourceCode { white-space: pre; position: relative; }
+pre > code.sourceCode > span { line-height: 1.25; }
+pre > code.sourceCode > span:empty { height: 1.2em; }
+.sourceCode { overflow: visible; }
+code.sourceCode > span { color: inherit; text-decoration: inherit; }
+div.sourceCode { margin: 1em 0; }
+pre.sourceCode { margin: 0; }
+@media screen {
+div.sourceCode { overflow: auto; }
+}
+@media print {
+pre > code.sourceCode { white-space: pre-wrap; }
+pre > code.sourceCode > span { display: inline-block; text-indent: -5em; padding-left: 5em; }
+}
+pre.numberSource code
+  { counter-reset: source-line 0; }
+pre.numberSource code > span
+  { position: relative; left: -4em; counter-increment: source-line; }
+pre.numberSource code > span > a:first-child::before
+  { content: counter(source-line);
+    position: relative; left: -1em; text-align: right; vertical-align: baseline;
+    border: none; display: inline-block;
+    -webkit-touch-callout: none; -webkit-user-select: none;
+    -khtml-user-select: none; -moz-user-select: none;
+    -ms-user-select: none; user-select: none;
+    padding: 0 4px; width: 4em;
+  }
+pre.numberSource { margin-left: 3em;  padding-left: 4px; }
+div.sourceCode
+  {   }
+@media screen {
+pre > code.sourceCode > span > a:first-child::before { text-decoration: underline; }
+}
+</style>
+
+
+<script src="../site_libs/quarto-nav/quarto-nav.js"></script>
+<script src="../site_libs/clipboard/clipboard.min.js"></script>
+<script src="../site_libs/quarto-search/autocomplete.umd.js"></script>
+<script src="../site_libs/quarto-search/fuse.min.js"></script>
+<script src="../site_libs/quarto-search/quarto-search.js"></script>
+<meta name="quarto:offset" content="../">
+<link href="../favicon.jpg" rel="icon" type="image/jpeg">
+<script src="../site_libs/quarto-html/quarto.js"></script>
+<script src="../site_libs/quarto-html/popper.min.js"></script>
+<script src="../site_libs/quarto-html/tippy.umd.min.js"></script>
+<script src="../site_libs/quarto-html/anchor.min.js"></script>
+<link href="../site_libs/quarto-html/tippy.css" rel="stylesheet">
+<link href="../site_libs/quarto-html/quarto-syntax-highlighting-dark-b53751a350365c71b6c909e95f209ed1.css" rel="stylesheet" id="quarto-text-highlighting-styles">
+<script src="../site_libs/bootstrap/bootstrap.min.js"></script>
+<link href="../site_libs/bootstrap/bootstrap-icons.css" rel="stylesheet">
+<link href="../site_libs/bootstrap/bootstrap-71f806479865a0e7fd52beea5cc266f1.min.css" rel="stylesheet" append-hash="true" id="quarto-bootstrap" data-mode="dark">
+<script id="quarto-search-options" type="application/json">{
+  "location": "navbar",
+  "copy-button": false,
+  "collapse-after": 3,
+  "panel-placement": "end",
+  "type": "overlay",
+  "limit": 50,
+  "keyboard-shortcut": [
+    "f",
+    "/",
+    "s"
+  ],
+  "show-item-context": false,
+  "language": {
+    "search-no-results-text": "No results",
+    "search-matching-documents-text": "matching documents",
+    "search-copy-link-title": "Copy link to search",
+    "search-hide-matches-text": "Hide additional matches",
+    "search-more-match-text": "more match in this document",
+    "search-more-matches-text": "more matches in this document",
+    "search-clear-button-title": "Clear",
+    "search-text-placeholder": "",
+    "search-detached-cancel-button-title": "Cancel",
+    "search-submit-button-title": "Submit",
+    "search-label": "Search"
+  }
+}</script>
+
+
+<link rel="stylesheet" href="../styles.css">
+</head>
+
+<body class="nav-sidebar docked nav-fixed">
+
+<div id="quarto-search-results"></div>
+  <header id="quarto-header" class="headroom fixed-top">
+    <nav class="navbar navbar-expand " data-bs-theme="dark">
+      <div class="navbar-container container-fluid">
+      <div class="navbar-brand-container mx-auto">
+    <a href="../index.html" class="navbar-brand navbar-brand-logo">
+    <img src="../image/axolotl_logo_digital_white.svg" alt="" class="navbar-logo">
+    </a>
+  </div>
+        <div class="quarto-navbar-tools tools-wide tools-end">
+    <a href="https://twitter.com/axolotl_ai" title="" class="quarto-navigation-tool px-1" aria-label=""><i class="bi bi-twitter"></i></a>
+    <a href="https://github.com/axolotl-ai-cloud/axolotl/" title="" class="quarto-navigation-tool px-1" aria-label=""><i class="bi bi-github"></i></a>
+    <a href="https://discord.gg/7m9sfhzaf3" title="" class="quarto-navigation-tool px-1" aria-label=""><i class="bi bi-discord"></i></a>
+</div>
+          <div id="quarto-search" class="" title="Search"></div>
+      </div> <!-- /container-fluid -->
+    </nav>
+  <nav class="quarto-secondary-nav">
+    <div class="container-fluid d-flex">
+      <button type="button" class="quarto-btn-toggle btn" data-bs-toggle="collapse" role="button" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">
+        <i class="bi bi-layout-text-sidebar-reverse"></i>
+      </button>
+        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/docker.html">Docker</a></li></ol></nav>
+        <a class="flex-grow-1" role="navigation" data-bs-toggle="collapse" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">      
+        </a>
+    </div>
+  </nav>
+</header>
+<!-- content -->
+<div id="quarto-content" class="quarto-container page-columns page-rows-contents page-layout-article page-navbar">
+<!-- sidebar -->
+  <nav id="quarto-sidebar" class="sidebar collapse collapse-horizontal quarto-sidebar-collapse-item sidebar-navigation docked overflow-auto">
+    <div class="sidebar-menu-container"> 
+    <ul class="list-unstyled mt-1">
+        <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../index.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Home</span></a>
+  </div>
+</li>
+        <li class="sidebar-item sidebar-item-section">
+      <div class="sidebar-item-container"> 
+            <a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-1" role="navigation" aria-expanded="true">
+ <span class="menu-text">Getting Started</span></a>
+          <a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-1" role="navigation" aria-expanded="true" aria-label="Toggle section">
+            <i class="bi bi-chevron-right ms-2"></i>
+          </a> 
+      </div>
+      <ul id="quarto-sidebar-section-1" class="collapse list-unstyled sidebar-section depth1 show">  
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/getting-started.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Quickstart</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/installation.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Installation</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/cli.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">CLI Reference</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/inference.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Inference</span></a>
+  </div>
+</li>
+      </ul>
+  </li>
+        <li class="sidebar-item sidebar-item-section">
+      <div class="sidebar-item-container"> 
+            <a href="../docs/dataset-formats/index.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Dataset Formats</span></a>
+          <a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-2" role="navigation" aria-expanded="true" aria-label="Toggle section">
+            <i class="bi bi-chevron-right ms-2"></i>
+          </a> 
+      </div>
+      <ul id="quarto-sidebar-section-2" class="collapse list-unstyled sidebar-section depth1 show">  
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/dataset-formats/pretraining.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Pre-training</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/dataset-formats/inst_tune.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Instruction Tuning</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/dataset-formats/conversation.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Conversation</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/dataset-formats/stepwise_supervised.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Stepwise Supervised Format</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/dataset-formats/template_free.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Template-Free</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/dataset-formats/tokenized.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Custom Pre-Tokenized Dataset</span></a>
+  </div>
+</li>
+      </ul>
+  </li>
+        <li class="sidebar-item sidebar-item-section">
+      <div class="sidebar-item-container"> 
+            <a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-3" role="navigation" aria-expanded="true">
+ <span class="menu-text">Deployments</span></a>
+          <a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-3" role="navigation" aria-expanded="true" aria-label="Toggle section">
+            <i class="bi bi-chevron-right ms-2"></i>
+          </a> 
+      </div>
+      <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link active">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Multi-GPU</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/multi-node.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Multi Node</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/ray-integration.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Ray Train</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/amd_hpc.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">AMD GPUs on HPC Systems</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/mac.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Mac M-series</span></a>
+  </div>
+</li>
+      </ul>
+  </li>
+        <li class="sidebar-item sidebar-item-section">
+      <div class="sidebar-item-container"> 
+            <a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-4" role="navigation" aria-expanded="true">
+ <span class="menu-text">How To Guides</span></a>
+          <a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-4" role="navigation" aria-expanded="true" aria-label="Toggle section">
+            <i class="bi bi-chevron-right ms-2"></i>
+          </a> 
+      </div>
+      <ul id="quarto-sidebar-section-4" class="collapse list-unstyled sidebar-section depth1 show">  
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/multimodal.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">MultiModal / Vision Language Models (BETA)</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/rlhf.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">RLHF (Beta)</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/reward_modelling.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Reward Modelling</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/lr_groups.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Learning Rate Groups</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/lora_optims.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">LoRA Optimizations</span></a>
+  </div>
+</li>
+      </ul>
+  </li>
+        <li class="sidebar-item sidebar-item-section">
+      <div class="sidebar-item-container"> 
+            <a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-5" role="navigation" aria-expanded="true">
+ <span class="menu-text">Core Concepts</span></a>
+          <a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-5" role="navigation" aria-expanded="true" aria-label="Toggle section">
+            <i class="bi bi-chevron-right ms-2"></i>
+          </a> 
+      </div>
+      <ul id="quarto-sidebar-section-5" class="collapse list-unstyled sidebar-section depth1 show">  
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/batch_vs_grad.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Batch size vs Gradient accumulation</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/dataset_preprocessing.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Dataset Preprocessing</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/multipack.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Multipack (Sample Packing)</span></a>
+  </div>
+</li>
+      </ul>
+  </li>
+        <li class="sidebar-item sidebar-item-section">
+      <div class="sidebar-item-container"> 
+            <a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-6" role="navigation" aria-expanded="true">
+ <span class="menu-text">Advanced Features</span></a>
+          <a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-6" role="navigation" aria-expanded="true" aria-label="Toggle section">
+            <i class="bi bi-chevron-right ms-2"></i>
+          </a> 
+      </div>
+      <ul id="quarto-sidebar-section-6" class="collapse list-unstyled sidebar-section depth1 show">  
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/fsdp_qlora.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">FDSP + QLoRA</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/unsloth.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Unsloth</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/torchao.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">PyTorch ao</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/custom_integrations.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Custom Integrations</span></a>
+  </div>
+</li>
+      </ul>
+  </li>
+        <li class="sidebar-item sidebar-item-section">
+      <div class="sidebar-item-container"> 
+            <a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-7" role="navigation" aria-expanded="true">
+ <span class="menu-text">Troubleshooting</span></a>
+          <a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-7" role="navigation" aria-expanded="true" aria-label="Toggle section">
+            <i class="bi bi-chevron-right ms-2"></i>
+          </a> 
+      </div>
+      <ul id="quarto-sidebar-section-7" class="collapse list-unstyled sidebar-section depth1 show">  
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/faq.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">FAQ</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/debugging.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Debugging</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/nccl.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">NCCL</span></a>
+  </div>
+</li>
+      </ul>
+  </li>
+        <li class="sidebar-item sidebar-item-section">
+      <div class="sidebar-item-container"> 
+            <a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-8" role="navigation" aria-expanded="true">
+ <span class="menu-text">Reference</span></a>
+          <a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-8" role="navigation" aria-expanded="true" aria-label="Toggle section">
+            <i class="bi bi-chevron-right ms-2"></i>
+          </a> 
+      </div>
+      <ul id="quarto-sidebar-section-8" class="collapse list-unstyled sidebar-section depth1 show">  
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
+  <a href="../docs/config.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Config options</span></a>
+  </div>
+</li>
+      </ul>
+  </li>
+    </ul>
+    </div>
+</nav>
+<div id="quarto-sidebar-glass" class="quarto-sidebar-collapse-item" data-bs-toggle="collapse" data-bs-target=".quarto-sidebar-collapse-item"></div>
+<!-- margin-sidebar -->
+    <div id="quarto-margin-sidebar" class="sidebar margin-sidebar">
+        <nav id="TOC" role="doc-toc" class="toc-active">
+    <h2 id="toc-title">On this page</h2>
+   
+  <ul>
+  <li><a href="#base" id="toc-base" class="nav-link active" data-scroll-target="#base">Base</a>
+  <ul>
+  <li><a href="#image" id="toc-image" class="nav-link" data-scroll-target="#image">Image</a></li>
+  <li><a href="#tags-format" id="toc-tags-format" class="nav-link" data-scroll-target="#tags-format">Tags format</a></li>
+  </ul></li>
+  <li><a href="#main" id="toc-main" class="nav-link" data-scroll-target="#main">Main</a>
+  <ul>
+  <li><a href="#image-1" id="toc-image-1" class="nav-link" data-scroll-target="#image-1">Image</a></li>
+  <li><a href="#sec-main-tags" id="toc-sec-main-tags" class="nav-link" data-scroll-target="#sec-main-tags">Tags format</a></li>
+  </ul></li>
+  <li><a href="#cloud" id="toc-cloud" class="nav-link" data-scroll-target="#cloud">Cloud</a>
+  <ul>
+  <li><a href="#image-2" id="toc-image-2" class="nav-link" data-scroll-target="#image-2">Image</a></li>
+  <li><a href="#tags-format-1" id="toc-tags-format-1" class="nav-link" data-scroll-target="#tags-format-1">Tags format</a></li>
+  <li><a href="#environment-variables" id="toc-environment-variables" class="nav-link" data-scroll-target="#environment-variables">Environment variables</a></li>
+  <li><a href="#volume-mounts" id="toc-volume-mounts" class="nav-link" data-scroll-target="#volume-mounts">Volume mounts</a></li>
+  </ul></li>
+  <li><a href="#cloud-no-tmux" id="toc-cloud-no-tmux" class="nav-link" data-scroll-target="#cloud-no-tmux">Cloud-no-tmux</a>
+  <ul>
+  <li><a href="#image-3" id="toc-image-3" class="nav-link" data-scroll-target="#image-3">Image</a></li>
+  <li><a href="#tags-format-2" id="toc-tags-format-2" class="nav-link" data-scroll-target="#tags-format-2">Tags format</a></li>
+  </ul></li>
+  </ul>
+</nav>
+    </div>
+<!-- main -->
+<main class="content" id="quarto-document-content">
+
+<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/docker.html">Docker</a></li></ol></nav>
+<div class="quarto-title">
+<h1 class="title">Docker</h1>
+</div>
+
+
+
+<div class="quarto-title-meta">
+
+    
+  
+    
+  </div>
+  
+
+
+</header>
+
+
+<p>This section describes the different Docker images that are released by AxolotlAI at <a href="https://hub.docker.com/u/axolotlai">Docker Hub</a>.</p>
+<section id="base" class="level2">
+<h2 class="anchored" data-anchor-id="base">Base</h2>
+<p>The base image is the most minimal image that can install Axolotl. It is based on the <code>nvidia/cuda</code> image. It includes python, torch, git, git-lfs, awscli, pydantic, and more.</p>
+<section id="image" class="level4">
+<h4 class="anchored" data-anchor-id="image">Image</h4>
+<pre><code>axolotlai/axolotl-base</code></pre>
+<p>Link: <a href="https://hub.docker.com/r/axolotlai/axolotl-base">Docker Hub</a></p>
+</section>
+<section id="tags-format" class="level4">
+<h4 class="anchored" data-anchor-id="tags-format">Tags format</h4>
+<div class="sourceCode" id="cb2"><pre class="sourceCode bash code-with-copy"><code class="sourceCode bash"><span id="cb2-1"><a href="#cb2-1" aria-hidden="true" tabindex="-1"></a><span class="ex">main-base-py{python_version}-cu{cuda_version}-{pytorch_version}</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<p>Tags examples:</p>
+<ul>
+<li><code>main-base-py3.11-cu124-2.6.0</code></li>
+<li><code>main-base-py3.11-cu124-2.5.1</code></li>
+<li><code>main-base-py3.11-cu124-2.4.1</code></li>
+</ul>
+</section>
+</section>
+<section id="main" class="level2">
+<h2 class="anchored" data-anchor-id="main">Main</h2>
+<p>The main image is the image that is used to run Axolotl. It is based on the <code>axolotlai/axolotl-base</code> image and includes the Axolotl codebase, dependencies, and more.</p>
+<section id="image-1" class="level4">
+<h4 class="anchored" data-anchor-id="image-1">Image</h4>
+<pre><code>axolotlai/axolotl</code></pre>
+<p>Link: <a href="https://hub.docker.com/r/axolotlai/axolotl">Docker Hub</a></p>
+</section>
+<section id="sec-main-tags" class="level4">
+<h4 class="anchored" data-anchor-id="sec-main-tags">Tags format</h4>
+<div class="sourceCode" id="cb4"><pre class="sourceCode bash code-with-copy"><code class="sourceCode bash"><span id="cb4-1"><a href="#cb4-1" aria-hidden="true" tabindex="-1"></a><span class="co"># on push to main</span></span>
+<span id="cb4-2"><a href="#cb4-2" aria-hidden="true" tabindex="-1"></a><span class="ex">main-py{python_version}-cu{cuda_version}-{pytorch_version}</span></span>
+<span id="cb4-3"><a href="#cb4-3" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb4-4"><a href="#cb4-4" aria-hidden="true" tabindex="-1"></a><span class="co"># latest main (currently torch 2.5.1, python 3.11, cuda 12.4)</span></span>
+<span id="cb4-5"><a href="#cb4-5" aria-hidden="true" tabindex="-1"></a><span class="ex">main-latest</span></span>
+<span id="cb4-6"><a href="#cb4-6" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb4-7"><a href="#cb4-7" aria-hidden="true" tabindex="-1"></a><span class="co"># nightly build</span></span>
+<span id="cb4-8"><a href="#cb4-8" aria-hidden="true" tabindex="-1"></a><span class="ex">{branch}-{date_in_YYYYMMDD}-py{python_version}-cu{cuda_version}-{pytorch_version}</span></span>
+<span id="cb4-9"><a href="#cb4-9" aria-hidden="true" tabindex="-1"></a></span>
+<span id="cb4-10"><a href="#cb4-10" aria-hidden="true" tabindex="-1"></a><span class="co"># tagged release</span></span>
+<span id="cb4-11"><a href="#cb4-11" aria-hidden="true" tabindex="-1"></a><span class="ex">{version}</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
+<div class="callout callout-style-default callout-tip callout-titled">
+<div class="callout-header d-flex align-content-center">
+<div class="callout-icon-container">
+<i class="callout-icon"></i>
+</div>
+<div class="callout-title-container flex-fill">
+Tip
+</div>
+</div>
+<div class="callout-body-container callout-body">
+<p>There may be some extra tags appended to the image, like <code>-vllm</code> which installs those packages.</p>
+</div>
+</div>
+<p>Tags examples:</p>
+<ul>
+<li><code>main-py3.11-cu124-2.6.0</code></li>
+<li><code>main-py3.11-cu124-2.5.1</code></li>
+<li><code>main-py3.11-cu124-2.4.1</code></li>
+<li><code>main-latest</code></li>
+<li><code>main-20250303-py3.11-cu124-2.6.0</code></li>
+<li><code>main-20250303-py3.11-cu124-2.5.1</code></li>
+<li><code>main-20250303-py3.11-cu124-2.4.1</code></li>
+<li><code>0.7.1</code></li>
+</ul>
+</section>
+</section>
+<section id="cloud" class="level2">
+<h2 class="anchored" data-anchor-id="cloud">Cloud</h2>
+<p>The cloud image is the image that is used to run Axolotl in the cloud. It is based on the <code>axolotlai/axolotl</code> image and sets ENV variables like HuggingFace cache directories for volume mounts, tmux, and more for different cloud providers.</p>
+<div class="callout callout-style-default callout-tip callout-titled">
+<div class="callout-header d-flex align-content-center">
+<div class="callout-icon-container">
+<i class="callout-icon"></i>
+</div>
+<div class="callout-title-container flex-fill">
+Tip
+</div>
+</div>
+<div class="callout-body-container callout-body">
+<p>Jupyter lab is run by default. Set <code>JUPYTER_DISABLE=1</code> in the environment variables to disable it.</p>
+</div>
+</div>
+<section id="image-2" class="level4">
+<h4 class="anchored" data-anchor-id="image-2">Image</h4>
+<pre><code>axolotlai/axolotl-cloud</code></pre>
+<p>Link: <a href="https://hub.docker.com/r/axolotlai/axolotl-cloud">Docker Hub</a></p>
+</section>
+<section id="tags-format-1" class="level4">
+<h4 class="anchored" data-anchor-id="tags-format-1">Tags format</h4>
+<p>This uses the same tags as the <a href="#sec-main-tags"><code>main</code> image</a>.</p>
+</section>
+<section id="environment-variables" class="level4">
+<h4 class="anchored" data-anchor-id="environment-variables">Environment variables</h4>
+<ul>
+<li><code>JUPYTER_DISABLE</code>: Disable Jupyter lab.</li>
+<li><code>JUPYTER_PASSWORD</code>: Set a password for the Jupyter lab.</li>
+<li><code>PUBLIC_KEY</code>: Add a public key for the SSH service.</li>
+<li><code>SSH_KEY</code>: Add a private key for the SSH service.</li>
+</ul>
+</section>
+<section id="volume-mounts" class="level4">
+<h4 class="anchored" data-anchor-id="volume-mounts">Volume mounts</h4>
+<div class="callout callout-style-default callout-tip callout-titled">
+<div class="callout-header d-flex align-content-center">
+<div class="callout-icon-container">
+<i class="callout-icon"></i>
+</div>
+<div class="callout-title-container flex-fill">
+Tip
+</div>
+</div>
+<div class="callout-body-container callout-body">
+<p>We recommend mounting volumes to <code>/workspace/data</code> for data persistence. <code>/workspace/axolotl</code> contains the source code and is ephemeral.</p>
+</div>
+</div>
+<ul>
+<li><code>/workspace/data/axolotl-artifacts</code>: Directory to store Axolotl artifacts.</li>
+<li><code>/workspace/data/huggingface-cache</code>: Directory to store HuggingFace cache.</li>
+</ul>
+</section>
+</section>
+<section id="cloud-no-tmux" class="level2">
+<h2 class="anchored" data-anchor-id="cloud-no-tmux">Cloud-no-tmux</h2>
+<p>This is the same as the <a href="#sec-cloud"><code>cloud</code> image</a> but without tmux.</p>
+<section id="image-3" class="level4">
+<h4 class="anchored" data-anchor-id="image-3">Image</h4>
+<pre><code>axolotlai/axolotl-cloud-term</code></pre>
+<p>Link: <a href="https://hub.docker.com/r/axolotlai/axolotl-cloud-term">Docker Hub</a></p>
+<div class="callout callout-style-default callout-note callout-titled">
+<div class="callout-header d-flex align-content-center">
+<div class="callout-icon-container">
+<i class="callout-icon"></i>
+</div>
+<div class="callout-title-container flex-fill">
+Note
+</div>
+</div>
+<div class="callout-body-container callout-body">
+<p>The naming may be a bit confusing as it has <code>-term</code> appended to the end.</p>
+</div>
+</div>
+</section>
+<section id="tags-format-2" class="level4">
+<h4 class="anchored" data-anchor-id="tags-format-2">Tags format</h4>
+<p>This uses the same tags as the <a href="#sec-cloud-tags"><code>cloud</code> image</a>.</p>
+
+
+</section>
+</section>
+
+</main> <!-- /main -->
+<script id="quarto-html-after-body" type="application/javascript">
+window.document.addEventListener("DOMContentLoaded", function (event) {
+  const toggleBodyColorMode = (bsSheetEl) => {
+    const mode = bsSheetEl.getAttribute("data-mode");
+    const bodyEl = window.document.querySelector("body");
+    if (mode === "dark") {
+      bodyEl.classList.add("quarto-dark");
+      bodyEl.classList.remove("quarto-light");
+    } else {
+      bodyEl.classList.add("quarto-light");
+      bodyEl.classList.remove("quarto-dark");
+    }
+  }
+  const toggleBodyColorPrimary = () => {
+    const bsSheetEl = window.document.querySelector("link#quarto-bootstrap");
+    if (bsSheetEl) {
+      toggleBodyColorMode(bsSheetEl);
+    }
+  }
+  toggleBodyColorPrimary();  
+  const icon = "";
+  const anchorJS = new window.AnchorJS();
+  anchorJS.options = {
+    placement: 'right',
+    icon: icon
+  };
+  anchorJS.add('.anchored');
+  const isCodeAnnotation = (el) => {
+    for (const clz of el.classList) {
+      if (clz.startsWith('code-annotation-')) {                     
+        return true;
+      }
+    }
+    return false;
+  }
+  const onCopySuccess = function(e) {
+    // button target
+    const button = e.trigger;
+    // don't keep focus
+    button.blur();
+    // flash "checked"
+    button.classList.add('code-copy-button-checked');
+    var currentTitle = button.getAttribute("title");
+    button.setAttribute("title", "Copied!");
+    let tooltip;
+    if (window.bootstrap) {
+      button.setAttribute("data-bs-toggle", "tooltip");
+      button.setAttribute("data-bs-placement", "left");
+      button.setAttribute("data-bs-title", "Copied!");
+      tooltip = new bootstrap.Tooltip(button, 
+        { trigger: "manual", 
+          customClass: "code-copy-button-tooltip",
+          offset: [0, -8]});
+      tooltip.show();    
+    }
+    setTimeout(function() {
+      if (tooltip) {
+        tooltip.hide();
+        button.removeAttribute("data-bs-title");
+        button.removeAttribute("data-bs-toggle");
+        button.removeAttribute("data-bs-placement");
+      }
+      button.setAttribute("title", currentTitle);
+      button.classList.remove('code-copy-button-checked');
+    }, 1000);
+    // clear code selection
+    e.clearSelection();
+  }
+  const getTextToCopy = function(trigger) {
+      const codeEl = trigger.previousElementSibling.cloneNode(true);
+      for (const childEl of codeEl.children) {
+        if (isCodeAnnotation(childEl)) {
+          childEl.remove();
+        }
+      }
+      return codeEl.innerText;
+  }
+  const clipboard = new window.ClipboardJS('.code-copy-button:not([data-in-quarto-modal])', {
+    text: getTextToCopy
+  });
+  clipboard.on('success', onCopySuccess);
+  if (window.document.getElementById('quarto-embedded-source-code-modal')) {
+    const clipboardModal = new window.ClipboardJS('.code-copy-button[data-in-quarto-modal]', {
+      text: getTextToCopy,
+      container: window.document.getElementById('quarto-embedded-source-code-modal')
+    });
+    clipboardModal.on('success', onCopySuccess);
+  }
+    var localhostRegex = new RegExp(/^(?:http|https):\/\/localhost\:?[0-9]*\//);
+    var mailtoRegex = new RegExp(/^mailto:/);
+      var filterRegex = new RegExp("https:\/\/axolotl-ai-cloud\.github\.io\/axolotl\/");
+    var isInternal = (href) => {
+        return filterRegex.test(href) || localhostRegex.test(href) || mailtoRegex.test(href);
+    }
+    // Inspect non-navigation links and adorn them if external
+ 	var links = window.document.querySelectorAll('a[href]:not(.nav-link):not(.navbar-brand):not(.toc-action):not(.sidebar-link):not(.sidebar-item-toggle):not(.pagination-link):not(.no-external):not([aria-hidden]):not(.dropdown-item):not(.quarto-navigation-tool):not(.about-link)');
+    for (var i=0; i<links.length; i++) {
+      const link = links[i];
+      if (!isInternal(link.href)) {
+        // undo the damage that might have been done by quarto-nav.js in the case of
+        // links that we want to consider external
+        if (link.dataset.originalHref !== undefined) {
+          link.href = link.dataset.originalHref;
+        }
+      }
+    }
+  function tippyHover(el, contentFn, onTriggerFn, onUntriggerFn) {
+    const config = {
+      allowHTML: true,
+      maxWidth: 500,
+      delay: 100,
+      arrow: false,
+      appendTo: function(el) {
+          return el.parentElement;
+      },
+      interactive: true,
+      interactiveBorder: 10,
+      theme: 'quarto',
+      placement: 'bottom-start',
+    };
+    if (contentFn) {
+      config.content = contentFn;
+    }
+    if (onTriggerFn) {
+      config.onTrigger = onTriggerFn;
+    }
+    if (onUntriggerFn) {
+      config.onUntrigger = onUntriggerFn;
+    }
+    window.tippy(el, config); 
+  }
+  const noterefs = window.document.querySelectorAll('a[role="doc-noteref"]');
+  for (var i=0; i<noterefs.length; i++) {
+    const ref = noterefs[i];
+    tippyHover(ref, function() {
+      // use id or data attribute instead here
+      let href = ref.getAttribute('data-footnote-href') || ref.getAttribute('href');
+      try { href = new URL(href).hash; } catch {}
+      const id = href.replace(/^#\/?/, "");
+      const note = window.document.getElementById(id);
+      if (note) {
+        return note.innerHTML;
+      } else {
+        return "";
+      }
+    });
+  }
+  const xrefs = window.document.querySelectorAll('a.quarto-xref');
+  const processXRef = (id, note) => {
+    // Strip column container classes
+    const stripColumnClz = (el) => {
+      el.classList.remove("page-full", "page-columns");
+      if (el.children) {
+        for (const child of el.children) {
+          stripColumnClz(child);
+        }
+      }
+    }
+    stripColumnClz(note)
+    if (id === null || id.startsWith('sec-')) {
+      // Special case sections, only their first couple elements
+      const container = document.createElement("div");
+      if (note.children && note.children.length > 2) {
+        container.appendChild(note.children[0].cloneNode(true));
+        for (let i = 1; i < note.children.length; i++) {
+          const child = note.children[i];
+          if (child.tagName === "P" && child.innerText === "") {
+            continue;
+          } else {
+            container.appendChild(child.cloneNode(true));
+            break;
+          }
+        }
+        if (window.Quarto?.typesetMath) {
+          window.Quarto.typesetMath(container);
+        }
+        return container.innerHTML
+      } else {
+        if (window.Quarto?.typesetMath) {
+          window.Quarto.typesetMath(note);
+        }
+        return note.innerHTML;
+      }
+    } else {
+      // Remove any anchor links if they are present
+      const anchorLink = note.querySelector('a.anchorjs-link');
+      if (anchorLink) {
+        anchorLink.remove();
+      }
+      if (window.Quarto?.typesetMath) {
+        window.Quarto.typesetMath(note);
+      }
+      if (note.classList.contains("callout")) {
+        return note.outerHTML;
+      } else {
+        return note.innerHTML;
+      }
+    }
+  }
+  for (var i=0; i<xrefs.length; i++) {
+    const xref = xrefs[i];
+    tippyHover(xref, undefined, function(instance) {
+      instance.disable();
+      let url = xref.getAttribute('href');
+      let hash = undefined; 
+      if (url.startsWith('#')) {
+        hash = url;
+      } else {
+        try { hash = new URL(url).hash; } catch {}
+      }
+      if (hash) {
+        const id = hash.replace(/^#\/?/, "");
+        const note = window.document.getElementById(id);
+        if (note !== null) {
+          try {
+            const html = processXRef(id, note.cloneNode(true));
+            instance.setContent(html);
+          } finally {
+            instance.enable();
+            instance.show();
+          }
+        } else {
+          // See if we can fetch this
+          fetch(url.split('#')[0])
+          .then(res => res.text())
+          .then(html => {
+            const parser = new DOMParser();
+            const htmlDoc = parser.parseFromString(html, "text/html");
+            const note = htmlDoc.getElementById(id);
+            if (note !== null) {
+              const html = processXRef(id, note);
+              instance.setContent(html);
+            } 
+          }).finally(() => {
+            instance.enable();
+            instance.show();
+          });
+        }
+      } else {
+        // See if we can fetch a full url (with no hash to target)
+        // This is a special case and we should probably do some content thinning / targeting
+        fetch(url)
+        .then(res => res.text())
+        .then(html => {
+          const parser = new DOMParser();
+          const htmlDoc = parser.parseFromString(html, "text/html");
+          const note = htmlDoc.querySelector('main.content');
+          if (note !== null) {
+            // This should only happen for chapter cross references
+            // (since there is no id in the URL)
+            // remove the first header
+            if (note.children.length > 0 && note.children[0].tagName === "HEADER") {
+              note.children[0].remove();
+            }
+            const html = processXRef(null, note);
+            instance.setContent(html);
+          } 
+        }).finally(() => {
+          instance.enable();
+          instance.show();
+        });
+      }
+    }, function(instance) {
+    });
+  }
+      let selectedAnnoteEl;
+      const selectorForAnnotation = ( cell, annotation) => {
+        let cellAttr = 'data-code-cell="' + cell + '"';
+        let lineAttr = 'data-code-annotation="' +  annotation + '"';
+        const selector = 'span[' + cellAttr + '][' + lineAttr + ']';
+        return selector;
+      }
+      const selectCodeLines = (annoteEl) => {
+        const doc = window.document;
+        const targetCell = annoteEl.getAttribute("data-target-cell");
+        const targetAnnotation = annoteEl.getAttribute("data-target-annotation");
+        const annoteSpan = window.document.querySelector(selectorForAnnotation(targetCell, targetAnnotation));
+        const lines = annoteSpan.getAttribute("data-code-lines").split(",");
+        const lineIds = lines.map((line) => {
+          return targetCell + "-" + line;
+        })
+        let top = null;
+        let height = null;
+        let parent = null;
+        if (lineIds.length > 0) {
+            //compute the position of the single el (top and bottom and make a div)
+            const el = window.document.getElementById(lineIds[0]);
+            top = el.offsetTop;
+            height = el.offsetHeight;
+            parent = el.parentElement.parentElement;
+          if (lineIds.length > 1) {
+            const lastEl = window.document.getElementById(lineIds[lineIds.length - 1]);
+            const bottom = lastEl.offsetTop + lastEl.offsetHeight;
+            height = bottom - top;
+          }
+          if (top !== null && height !== null && parent !== null) {
+            // cook up a div (if necessary) and position it 
+            let div = window.document.getElementById("code-annotation-line-highlight");
+            if (div === null) {
+              div = window.document.createElement("div");
+              div.setAttribute("id", "code-annotation-line-highlight");
+              div.style.position = 'absolute';
+              parent.appendChild(div);
+            }
+            div.style.top = top - 2 + "px";
+            div.style.height = height + 4 + "px";
+            div.style.left = 0;
+            let gutterDiv = window.document.getElementById("code-annotation-line-highlight-gutter");
+            if (gutterDiv === null) {
+              gutterDiv = window.document.createElement("div");
+              gutterDiv.setAttribute("id", "code-annotation-line-highlight-gutter");
+              gutterDiv.style.position = 'absolute';
+              const codeCell = window.document.getElementById(targetCell);
+              const gutter = codeCell.querySelector('.code-annotation-gutter');
+              gutter.appendChild(gutterDiv);
+            }
+            gutterDiv.style.top = top - 2 + "px";
+            gutterDiv.style.height = height + 4 + "px";
+          }
+          selectedAnnoteEl = annoteEl;
+        }
+      };
+      const unselectCodeLines = () => {
+        const elementsIds = ["code-annotation-line-highlight", "code-annotation-line-highlight-gutter"];
+        elementsIds.forEach((elId) => {
+          const div = window.document.getElementById(elId);
+          if (div) {
+            div.remove();
+          }
+        });
+        selectedAnnoteEl = undefined;
+      };
+        // Handle positioning of the toggle
+    window.addEventListener(
+      "resize",
+      throttle(() => {
+        elRect = undefined;
+        if (selectedAnnoteEl) {
+          selectCodeLines(selectedAnnoteEl);
+        }
+      }, 10)
+    );
+    function throttle(fn, ms) {
+    let throttle = false;
+    let timer;
+      return (...args) => {
+        if(!throttle) { // first call gets through
+            fn.apply(this, args);
+            throttle = true;
+        } else { // all the others get throttled
+            if(timer) clearTimeout(timer); // cancel #2
+            timer = setTimeout(() => {
+              fn.apply(this, args);
+              timer = throttle = false;
+            }, ms);
+        }
+      };
+    }
+      // Attach click handler to the DT
+      const annoteDls = window.document.querySelectorAll('dt[data-target-cell]');
+      for (const annoteDlNode of annoteDls) {
+        annoteDlNode.addEventListener('click', (event) => {
+          const clickedEl = event.target;
+          if (clickedEl !== selectedAnnoteEl) {
+            unselectCodeLines();
+            const activeEl = window.document.querySelector('dt[data-target-cell].code-annotation-active');
+            if (activeEl) {
+              activeEl.classList.remove('code-annotation-active');
+            }
+            selectCodeLines(clickedEl);
+            clickedEl.classList.add('code-annotation-active');
+          } else {
+            // Unselect the line
+            unselectCodeLines();
+            clickedEl.classList.remove('code-annotation-active');
+          }
+        });
+      }
+  const findCites = (el) => {
+    const parentEl = el.parentElement;
+    if (parentEl) {
+      const cites = parentEl.dataset.cites;
+      if (cites) {
+        return {
+          el,
+          cites: cites.split(' ')
+        };
+      } else {
+        return findCites(el.parentElement)
+      }
+    } else {
+      return undefined;
+    }
+  };
+  var bibliorefs = window.document.querySelectorAll('a[role="doc-biblioref"]');
+  for (var i=0; i<bibliorefs.length; i++) {
+    const ref = bibliorefs[i];
+    const citeInfo = findCites(ref);
+    if (citeInfo) {
+      tippyHover(citeInfo.el, function() {
+        var popup = window.document.createElement('div');
+        citeInfo.cites.forEach(function(cite) {
+          var citeDiv = window.document.createElement('div');
+          citeDiv.classList.add('hanging-indent');
+          citeDiv.classList.add('csl-entry');
+          var biblioDiv = window.document.getElementById('ref-' + cite);
+          if (biblioDiv) {
+            citeDiv.innerHTML = biblioDiv.innerHTML;
+          }
+          popup.appendChild(citeDiv);
+        });
+        return popup.innerHTML;
+      });
+    }
+  }
+});
+</script>
+</div> <!-- /content -->
+
+
+
+
+</body></html>
\ No newline at end of file
diff --git a/docs/faq.html b/docs/faq.html
index 046fc3f91..89f080c03 100644
--- a/docs/faq.html
+++ b/docs/faq.html
@@ -207,6 +207,12 @@ ul.task-list li input[type="checkbox"] {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
@@ -445,8 +451,9 @@ ul.task-list li input[type="checkbox"] {
 <p>A: Try upgrading deepspeed w: <code>pip install -U deepspeed</code></p>
 </blockquote>
 <p><strong>Q: AttributeError: ‘DummyOptim’ object has no attribute ‘step’</strong></p>
+<p><strong>Q: ModuleNotFoundError: No module named ‘mpi4py’ using single GPU with deepspeed</strong></p>
 <blockquote class="blockquote">
-<p>A: You may be using deepspeed with single gpu. Please don’t set <code>deepspeed:</code> in yaml or cli.</p>
+<p>A: You may be using deepspeed with single gpu. Please remove the <code>deepspeed:</code> section in the yaml file or <code>--deepspeed</code> CLI flag.</p>
 </blockquote>
 <p><strong>Q: The codes is stuck on saving preprocessed datasets.</strong></p>
 <blockquote class="blockquote">
diff --git a/docs/fsdp_qlora.html b/docs/fsdp_qlora.html
index dc7c5c165..f3a31ae82 100644
--- a/docs/fsdp_qlora.html
+++ b/docs/fsdp_qlora.html
@@ -207,6 +207,12 @@ ul.task-list li input[type="checkbox"] {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/getting-started.html b/docs/getting-started.html
index aa611c261..cd04bb868 100644
--- a/docs/getting-started.html
+++ b/docs/getting-started.html
@@ -240,6 +240,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/inference.html b/docs/inference.html
index 9efec1553..08cd06e70 100644
--- a/docs/inference.html
+++ b/docs/inference.html
@@ -240,6 +240,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/input_output.html b/docs/input_output.html
index f8511d912..a5bc91737 100644
--- a/docs/input_output.html
+++ b/docs/input_output.html
@@ -207,6 +207,12 @@ ul.task-list li input[type="checkbox"] {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/installation.html b/docs/installation.html
index 1aca1b0d6..53f4ae4d2 100644
--- a/docs/installation.html
+++ b/docs/installation.html
@@ -240,6 +240,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
@@ -525,6 +531,7 @@ Advanced Docker Configuration
 <span id="cb5-6"><a href="#cb5-6" aria-hidden="true" tabindex="-1"></a>  axolotlai/axolotl:main-latest</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
 </div>
 </div>
+<p>Please refer to the <a href="../docs/docker.html">Docker documentation</a> for more information on the different Docker images that are available.</p>
 </section>
 </section>
 <section id="sec-cloud" class="level2" data-number="3">
diff --git a/docs/lora_optims.html b/docs/lora_optims.html
index 55b8f410d..02da7dbc9 100644
--- a/docs/lora_optims.html
+++ b/docs/lora_optims.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/lr_groups.html b/docs/lr_groups.html
index 65e8c3042..993c1d142 100644
--- a/docs/lr_groups.html
+++ b/docs/lr_groups.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/mac.html b/docs/mac.html
index dad9cf504..9ebf008d7 100644
--- a/docs/mac.html
+++ b/docs/mac.html
@@ -96,7 +96,7 @@ ul.task-list li input[type="checkbox"] {
       <button type="button" class="quarto-btn-toggle btn" data-bs-toggle="collapse" role="button" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">
         <i class="bi bi-layout-text-sidebar-reverse"></i>
       </button>
-        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/mac.html">Mac M-series</a></li></ol></nav>
+        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/mac.html">Mac M-series</a></li></ol></nav>
         <a class="flex-grow-1" role="navigation" data-bs-toggle="collapse" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">      
         </a>
     </div>
@@ -207,6 +207,12 @@ ul.task-list li input[type="checkbox"] {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
@@ -399,7 +405,7 @@ ul.task-list li input[type="checkbox"] {
 <!-- main -->
 <main class="content" id="quarto-document-content">
 
-<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/mac.html">Mac M-series</a></li></ol></nav>
+<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/mac.html">Mac M-series</a></li></ol></nav>
 <div class="quarto-title">
 <h1 class="title">Mac M-series</h1>
 </div>
diff --git a/docs/multi-gpu.html b/docs/multi-gpu.html
index d610eed5a..ed820f31d 100644
--- a/docs/multi-gpu.html
+++ b/docs/multi-gpu.html
@@ -129,7 +129,7 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <button type="button" class="quarto-btn-toggle btn" data-bs-toggle="collapse" role="button" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">
         <i class="bi bi-layout-text-sidebar-reverse"></i>
       </button>
-        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Multi-GPU</a></li></ol></nav>
+        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Multi-GPU</a></li></ol></nav>
         <a class="flex-grow-1" role="navigation" data-bs-toggle="collapse" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">      
         </a>
     </div>
@@ -240,6 +240,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link active">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
@@ -458,7 +464,7 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 <!-- main -->
 <main class="content" id="quarto-document-content">
 
-<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Multi-GPU</a></li></ol></nav>
+<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Multi-GPU</a></li></ol></nav>
 <div class="quarto-title">
 <div class="quarto-title-block"><div><h1 class="title">Multi-GPU</h1><button type="button" class="btn code-tools-button" id="quarto-code-tools-source"><i class="bi"></i> Code</button></div></div>
 </div>
diff --git a/docs/multi-node.html b/docs/multi-node.html
index 6514e7d96..1ca67a97b 100644
--- a/docs/multi-node.html
+++ b/docs/multi-node.html
@@ -130,7 +130,7 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <button type="button" class="quarto-btn-toggle btn" data-bs-toggle="collapse" role="button" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">
         <i class="bi bi-layout-text-sidebar-reverse"></i>
       </button>
-        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/multi-node.html">Multi Node</a></li></ol></nav>
+        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/multi-node.html">Multi Node</a></li></ol></nav>
         <a class="flex-grow-1" role="navigation" data-bs-toggle="collapse" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">      
         </a>
     </div>
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
@@ -441,7 +447,7 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 <!-- main -->
 <main class="content" id="quarto-document-content">
 
-<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/multi-node.html">Multi Node</a></li></ol></nav>
+<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/multi-node.html">Multi Node</a></li></ol></nav>
 <div class="quarto-title">
 <h1 class="title">Multi Node</h1>
 </div>
diff --git a/docs/multimodal.html b/docs/multimodal.html
index afd7aa476..50ce457ac 100644
--- a/docs/multimodal.html
+++ b/docs/multimodal.html
@@ -240,6 +240,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/multipack.html b/docs/multipack.html
index 110c63f2e..de80047c4 100644
--- a/docs/multipack.html
+++ b/docs/multipack.html
@@ -207,6 +207,12 @@ ul.task-list li input[type="checkbox"] {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/nccl.html b/docs/nccl.html
index 07f0f72c0..9bd145091 100644
--- a/docs/nccl.html
+++ b/docs/nccl.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/ray-integration.html b/docs/ray-integration.html
index 632d4bba1..9a862c786 100644
--- a/docs/ray-integration.html
+++ b/docs/ray-integration.html
@@ -130,7 +130,7 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <button type="button" class="quarto-btn-toggle btn" data-bs-toggle="collapse" role="button" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">
         <i class="bi bi-layout-text-sidebar-reverse"></i>
       </button>
-        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/ray-integration.html">Ray Train</a></li></ol></nav>
+        <nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/ray-integration.html">Ray Train</a></li></ol></nav>
         <a class="flex-grow-1" role="navigation" data-bs-toggle="collapse" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">      
         </a>
     </div>
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
@@ -442,7 +448,7 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 <!-- main -->
 <main class="content" id="quarto-document-content">
 
-<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/multi-gpu.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/ray-integration.html">Ray Train</a></li></ol></nav>
+<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item"><a href="../docs/docker.html">Deployments</a></li><li class="breadcrumb-item"><a href="../docs/ray-integration.html">Ray Train</a></li></ol></nav>
 <div class="quarto-title">
 <h1 class="title">Ray Train</h1>
 </div>
diff --git a/docs/reward_modelling.html b/docs/reward_modelling.html
index 6aa385590..f21cc0021 100644
--- a/docs/reward_modelling.html
+++ b/docs/reward_modelling.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/rlhf.html b/docs/rlhf.html
index 53c9011f8..767826029 100644
--- a/docs/rlhf.html
+++ b/docs/rlhf.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/torchao.html b/docs/torchao.html
index c289cbdc6..30211edf7 100644
--- a/docs/torchao.html
+++ b/docs/torchao.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/docs/unsloth.html b/docs/unsloth.html
index d11e2442b..d9b19d833 100644
--- a/docs/unsloth.html
+++ b/docs/unsloth.html
@@ -241,6 +241,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/examples/colab-notebooks/colab-axolotl-example.html b/examples/colab-notebooks/colab-axolotl-example.html
index 4c11d530c..2c5ebd840 100644
--- a/examples/colab-notebooks/colab-axolotl-example.html
+++ b/examples/colab-notebooks/colab-axolotl-example.html
@@ -269,6 +269,12 @@ window.Quarto = {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/index.html b/index.html
index 7ceb057f6..78e58be2f 100644
--- a/index.html
+++ b/index.html
@@ -240,6 +240,12 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="./docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="./docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
@@ -456,7 +462,7 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
 <picture> <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/887513285d98132142bf5db2a74eb5e0928787f1/image/axolotl_logo_digital_white.svg"> <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/887513285d98132142bf5db2a74eb5e0928787f1/image/axolotl_logo_digital_black.svg"> <img alt="Axolotl" src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/887513285d98132142bf5db2a74eb5e0928787f1/image/axolotl_logo_digital_black.svg" width="400" height="104" style="max-width: 100%;"> </picture>
 </p>
 <p align="center">
-<img src="https://img.shields.io/github/license/axolotl-ai-cloud/axolotl.svg?color=blue" alt="GitHub License"> <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests.yml/badge.svg" alt="tests"> <a href="https://github.com/axolotl-ai-cloud/axolotl/releases"><img src="https://img.shields.io/github/release/axolotl-ai-cloud/axolotl.svg" alt="Releases"></a> <br> <a href="https://github.com/axolotl-ai-cloud/axolotl/graphs/contributors"><img src="https://img.shields.io/github/contributors-anon/axolotl-ai-cloud/axolotl?color=yellow&amp;style=flat-square" alt="contributors" style="height: 20px;"></a> <img src="https://img.shields.io/github/stars/axolotl-ai-cloud/axolotl" alt="GitHub Repo stars"> <br> <a href="https://discord.com/invite/HhrNrHJPRb"><img src="https://img.shields.io/badge/discord-7289da.svg?style=flat-square&amp;logo=discord" alt="discord" style="height: 20px;"></a> <a href="https://twitter.com/axolotl_ai"><img src="https://img.shields.io/twitter/follow/axolotl_ai?style=social" alt="twitter" style="height: 20px;"></a> <br> <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests-nightly.yml/badge.svg" alt="tests-nightly"> <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/multi-gpu-e2e.yml/badge.svg" alt="multigpu-semi-weekly tests"> <a href="https://www.phorm.ai/query?projectId=e315ba4a-4e14-421f-ab05-38a1f9076f25"> <img alt="phorm.ai" src="https://img.shields.io/badge/Phorm-Ask_AI-%23F2777A.svg?&amp;logo=data:image/svg+xml;base64,PHN2ZyB3aWR0aD0iNSIgaGVpZ2h0PSI0IiBmaWxsPSJub25lIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPgogIDxwYXRoIGQ9Ik00LjQzIDEuODgyYTEuNDQgMS40NCAwIDAgMS0uMDk4LjQyNmMtLjA1LjEyMy0uMTE1LjIzLS4xOTIuMzIyLS4wNzUuMDktLjE2LjE2NS0uMjU1LjIyNmExLjM1MyAxLjM1MyAwIDAgMS0uNTk1LjIxMmMtLjA5OS4wMTItLjE5Mi4wMTQtLjI3OS4wMDZsLTEuNTkzLS4xNHYtLjQwNmgxLjY1OGMuMDkuMDAxLjE3LS4xNjkuMjQ2LS4xOTFhLjYwMy42MDMgMCAwIDAgLjItLjEwNi41MjkuNTI5IDAgMCAwIC4xMzgtLjE3LjY1NC42NTQgMCAwIDAgLjA2NS0uMjRsLjAyOC0uMzJhLjkzLjkzIDAgMCAwLS4wMzYtLjI0OS41NjcuNTY3IDAgMCAwLS4xMDMtLjIuNTAyLjUwMiAwIDAgMC0uMTY4LS4xMzguNjA4LjYwOCAwIDAgMC0uMjQtLjA2N0wyLjQzNy43MjkgMS42MjUuNjcxYS4zMjIuMzIyIDAgMCAwLS4yMzIuMDU4LjM3NS4zNzUgMCAwIDAtLjExNi4yMzJsLS4xMTYgMS40NS0uMDU4LjY5Ny0uMDU4Ljc1NEwuNzA1IDRsLS4zNTctLjA3OUwuNjAyLjkwNkMuNjE3LjcyNi42NjMuNTc0LjczOS40NTRhLjk1OC45NTggMCAwIDEgLjI3NC0uMjg1Ljk3MS45NzEgMCAwIDEgLjMzNy0uMTRjLjExOS0uMDI2LjIyNy0uMDM0LjMyNS0uMDI2TDMuMjMyLjE2Yy4xNTkuMDE0LjMzNi4wMy40NTkuMDgyYTEuMTczIDEuMTczIDAgMCAxIC41NDUuNDQ3Yy4wNi4wOTQuMTA5LjE5Mi4xNDQuMjkzYTEuMzkyIDEuMzkyIDAgMCAxIC4wNzguNThsLS4wMjkuMzJaIiBmaWxsPSIjRjI3NzdBIi8+CiAgPHBhdGggZD0iTTQuMDgyIDIuMDA3YTEuNDU1IDEuNDU1IDAgMCAxLS4wOTguNDI3Yy0uMDUuMTI0LS4xMTQuMjMyLS4xOTIuMzI0YTEuMTMgMS4xMyAwIDAgMS0uMjU0LjIyNyAxLjM1MyAxLjM1MyAwIDAgMS0uNTk1LjIxNGMtLjEuMDEyLS4xOTMuMDE0LS4yOC4wMDZsLTEuNTYtLjEwOC4wMzQtLjQwNi4wMy0uMzQ4IDEuNTU5LjE1NGMuMDkgMCAuMTczLS4wMS4yNDgtLjAzM2EuNjAzLjYwMyAwIDAgMCAuMi0uMTA2LjUzMi41MzIgMCAwIDAgLjEzOS0uMTcyLjY2LjY2IDAgMCAwIC4wNjQtLjI0MWwuMDI5LS4zMjFhLjk0Ljk0IDAgMCAwLS4wMzYtLjI1LjU3LjU3IDAgMCAwLS4xMDMtLjIwMi41MDIuNTAyIDAgMCAwLS4xNjgtLjEzOC42MDUuNjA1IDAgMCAwLS4yNC0uMDY3TDEuMjczLjgyN2MtLjA5NC0uMDA4LS4xNjguMDEtLjIyMS4wNTUtLjA1My4wNDUtLjA4NC4xMTQtLjA5Mi4yMDZMLjcwNSA0IDAgMy45MzhsLjI1NS0yLjkxMUExLjAxIDEuMDEgMCAwIDEgLjM5My41NzIuOTYyLjk2MiAwIDAgMSAuNjY2LjI4NmEuOTcuOTcgMCAwIDEgLjMzOC0uMTRDMS4xMjIuMTIgMS4yMy4xMSAxLjMyOC4xMTlsMS41OTMuMTRjLjE2LjAxNC4zLjA0Ny40MjMuMWExLjE3IDEuMTcgMCAwIDEgLjU0NS40NDhjLjA2MS4wOTUuMTA5LjE5My4xNDQuMjk1YTEuNDA2IDEuNDA2IDAgMCAxIC4wNzcuNTgzbC0uMDI4LjMyMloiIGZpbGw9IndoaXRlIi8+CiAgPHBhdGggZD0iTTQuMDgyIDIuMDA3YTEuNDU1IDEuNDU1IDAgMCAxLS4wOTguNDI3Yy0uMDUuMTI0LS4xMTQuMjMyLS4xOTIuMzI0YTEuMTMgMS4xMyAwIDAgMS0uMjU0LjIyNyAxLjM1MyAxLjM1MyAwIDAgMS0uNTk1LjIxNGMtLjEuMDEyLS4xOTMuMDE0LS4yOC4wMDZsLTEuNTYtLjEwOC4wMzQtLjQwNi4wMy0uMzQ4IDEuNTU5LjE1NGMuMDkgMCAuMTczLS4wMS4yNDgtLjAzM2EuNjAzLjYwMyAwIDAgMCAuMi0uMTA2LjUzMi41MzIgMCAwIDAgLjEzOS0uMTcyLjY2LjY2IDAgMCAwIC4wNjQtLjI0MWwuMDI5LS4zMjFhLjk0Ljk0IDAgMCAwLS4wMzYtLjI1LjU3LjU3IDAgMCAwLS4xMDMtLjIwMi41MDIuNTAyIDAgMCAwLS4xNjgtLjEzOC42MDUuNjA1IDAgMCAwLS4yNC0uMDY3TDEuMjczLjgyN2MtLjA5NC0uMDA4LS4xNjguMDEtLjIyMS4wNTUtLjA1My4wNDUtLjA4NC4xMTQtLjA5Mi4yMDZMLjcwNSA0IDAgMy45MzhsLjI1NS0yLjkxMUExLjAxIDEuMDEgMCAwIDEgLjM5My41NzIuOTYyLjk2MiAwIDAgMSAuNjY2LjI4NmEuOTcuOTcgMCAwIDEgLjMzOC0uMTRDMS4xMjIuMTIgMS4yMy4xMSAxLjMyOC4xMTlsMS41OTMuMTRjLjE2LjAxNC4zLjA0Ny40MjMuMWExLjE3IDEuMTcgMCAwIDEgLjU0NS40NDhjLjA2MS4wOTUuMTA5LjE5My4xNDQuMjk1YTEuNDA2IDEuNDA2IDAgMCAxIC4wNzcuNTgzbC0uMDI4LjMyMloiIGZpbGw9IndoaXRlIi8+Cjwvc3ZnPgo="> </a>
+<img src="https://img.shields.io/github/license/axolotl-ai-cloud/axolotl.svg?color=blue" alt="GitHub License"> <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests.yml/badge.svg" alt="tests"> <a href="https://github.com/axolotl-ai-cloud/axolotl/releases"><img src="https://img.shields.io/github/release/axolotl-ai-cloud/axolotl.svg" alt="Releases"></a> <br> <a href="https://github.com/axolotl-ai-cloud/axolotl/graphs/contributors"><img src="https://img.shields.io/github/contributors-anon/axolotl-ai-cloud/axolotl?color=yellow&amp;style=flat-square" alt="contributors" style="height: 20px;"></a> <img src="https://img.shields.io/github/stars/axolotl-ai-cloud/axolotl" alt="GitHub Repo stars"> <br> <a href="https://discord.com/invite/HhrNrHJPRb"><img src="https://img.shields.io/badge/discord-7289da.svg?style=flat-square&amp;logo=discord" alt="discord" style="height: 20px;"></a> <a href="https://twitter.com/axolotl_ai"><img src="https://img.shields.io/twitter/follow/axolotl_ai?style=social" alt="twitter" style="height: 20px;"></a> <br> <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests-nightly.yml/badge.svg" alt="tests-nightly"> <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/multi-gpu-e2e.yml/badge.svg" alt="multigpu-semi-weekly tests">
 </p>
 <p>Axolotl is a tool designed to streamline post-training for various AI models. Post-training refers to any modifications or additional training performed on pre-trained models - including full model fine-tuning, parameter-efficient tuning (like LoRA and QLoRA), supervised fine-tuning (SFT), instruction tuning, and alignment techniques. With support for multiple model architectures and training configurations, Axolotl makes it easy to get started with these techniques.</p>
 <p>Axolotl is designed to work with YAML config files that contain everything you need to preprocess a dataset, train or fine-tune a model, run model inference or evaluation, and much more.</p>
diff --git a/search.json b/search.json
index 25d7e7799..c0e419eb3 100644
--- a/search.json
+++ b/search.json
@@ -148,414 +148,483 @@
     ]
   },
   {
-    "objectID": "docs/mac.html",
-    "href": "docs/mac.html",
-    "title": "Mac M-series",
+    "objectID": "docs/config.html",
+    "href": "docs/config.html",
+    "title": "Config options",
     "section": "",
-    "text": "Currently Axolotl on Mac is partially usable, many of the dependencies of Axolotl including Pytorch do not support MPS or have incomplete support.\nCurrent support:\n\nSupport for all models\nFull training of models\nLoRA training\nSample packing\nFP16 and BF16 (awaiting AMP support for MPS in Pytorch)\nTri-dao’s flash-attn (until it is supported use spd_attention as an alternative)\nxformers\nbitsandbytes (meaning no 4/8 bits loading and bnb optimizers)\nqlora\nDeepSpeed\n\nUntested:\n\nFSDP",
+    "text": "# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files\n# This can also be a relative path to a model on disk\nbase_model: ./llama-7b-hf\n# You can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)\nbase_model_ignore_patterns:\n# If the base_model repo on hf hub doesn't include configuration .json files,\n# You can set that here, or leave this empty to default to base_model\nbase_model_config: ./llama-7b-hf\n# You can specify to choose a specific model revision from huggingface hub\nrevision_of_model:\n# Optional tokenizer configuration path in case you want to use a different tokenizer\n# than the one defined in the base model\ntokenizer_config:\n# If you want to specify the type of model to load, AutoModelForCausalLM is a good choice too\nmodel_type: AutoModelForCausalLM\n# Corresponding tokenizer for the model AutoTokenizer is a good choice\ntokenizer_type: AutoTokenizer\n# Trust remote code for untrusted source\ntrust_remote_code:\n# use_fast option for tokenizer loading from_pretrained, default to True\ntokenizer_use_fast:\n# Whether to use the legacy tokenizer setting, defaults to True\ntokenizer_legacy:\n# Resize the model embeddings when new tokens are added to multiples of 32\n# This is reported to improve training speed on some models\nresize_token_embeddings_to_32x:\n\n# (Internal use only)\n# Used to identify which the model is based on\nis_falcon_derived_model:\nis_llama_derived_model:\nis_qwen_derived_model:\n# Please note that if you set this to true, `padding_side` will be set to \"left\" by default\nis_mistral_derived_model:\n\n# optional overrides to the base model configuration\noverrides_of_model_config:\n  # RoPE Scaling https://github.com/huggingface/transformers/pull/24653\n  rope_scaling:\n    type: # linear | dynamic\n    factor: # float\n\n# optional overrides the base model loading from_pretrained\noverrides_of_model_kwargs:\n  # use_cache: False\n\n# optional overrides to the bnb 4bit quantization configuration\n# https://huggingface.co/docs/transformers/main/main_classes/quantization#transformers.BitsAndBytesConfig\nbnb_config_kwargs:\n  # These are default values\n  llm_int8_has_fp16_weight: false\n  bnb_4bit_quant_type: nf4\n  bnb_4bit_use_double_quant: true\n\n\n# Whether you are training a 4-bit GPTQ quantized model\ngptq: true\n\n# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer\nload_in_8bit: true\n# Use bitsandbytes 4 bit\nload_in_4bit:\n\n# Use CUDA bf16\nbf16: true # bool or 'full' for `bf16_full_eval`. require &gt;=ampere\n# Use CUDA fp16\nfp16: true\n# Use CUDA tf32\ntf32: true # require &gt;=ampere\n\n# No AMP (automatic mixed precision)\nbfloat16: true # require &gt;=ampere\nfloat16: true\n\n# Limit the memory for all available GPUs to this amount (if an integer, expressed in gigabytes); default: unset\ngpu_memory_limit: 20GiB\n# Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge\nlora_on_cpu: true\n\n# A list of one or more datasets to finetune the model with\ndatasets:\n  # HuggingFace dataset repo | s3://,gs:// path | \"json\" for local dataset, make sure to fill data_files\n  - path: vicgalle/alpaca-gpt4\n    # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]\n    type: alpaca # format | format:&lt;prompt_style&gt; (chat/instruct) | &lt;prompt_strategies&gt;.load_&lt;load_fn&gt;\n    ds_type: # Optional[str] (json|arrow|parquet|text|csv) defines the datatype when path is a file\n    data_files: # Optional[str] path to source data files\n\n    shards: # Optional[int] split dataset into N pieces (use with shards_idx)\n    shards_idx: # Optional[int] = 0 the index of sharded dataset to use\n\n    preprocess_shards: # Optional[int] process dataset in N sequential chunks for memory efficiency (exclusive with `shards`)\n\n    name: # Optional[str] name of dataset configuration to load\n    train_on_split: train # Optional[str] name of dataset split to load from\n    revision: # Optional[str] The specific revision of the dataset to use when loading from the Hugging Face Hub. This can be a commit hash, tag, or branch name. If not specified, the latest version will be used. This parameter is ignored for local datasets.\n    trust_remote_code: # Optional[bool] Trust remote code for untrusted source\n\n  # Custom user instruction prompt\n  - path: repo\n    type:\n      # The below are defaults. only set what's needed if you use a different column name.\n      system_prompt: \"\"\n      system_format: \"{system}\"\n      field_system: system\n      field_instruction: instruction\n      field_input: input\n      field_output: output\n\n      # Customizable to be single line or multi-line\n      # Use {instruction}/{input} as key to be replaced\n      # 'format' can include {input}\n      format: |-\n        User: {instruction} {input}\n        Assistant:\n      # 'no_input_format' cannot include {input}\n      no_input_format: \"{instruction} \"\n\n      # For `completion` datsets only, uses the provided field instead of `text` column\n      field:\n\n  # Using chat template\n  - path: ...\n    # Set type to `chat_template` to use this strategy\n    type: chat_template\n    # Specify the name of the chat template to use\n    # The name of the chat template to use for training, following values are supported:\n    # - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default.\n    # - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py\n    # - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to if the tokenizer does not have a chat template else default to tokenizer. E.g. tokenizer_default_fallback_chatml.\n    # - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.\n    chat_template: tokenizer_default\n\n    # Custom jinja chat template. Used only if `chat_template: jinja` or empty.\n    chat_template_jinja:\n\n    # Key containing the messages (default: \"messages\")\n    field_messages: messages\n\n    # Mapping of properties from the input dataset to the chat template.\n    # (default: message_property_mappings={'role':'role', 'content':'content'})\n    # If a property exists in the template but not in this mapping, the system will attempt\n    # to load it directly from the message using the property name as the key.\n    # Example: In the mapping below, 'from' is loaded from input dataset and used as 'role',\n    # while 'value' is loaded and used as 'content' in the chat template.\n    message_property_mappings:\n      role: from\n      content: value\n      # ...\n\n    message_property_mappings:\n\n    # Optional[Dict[str, List]]. Roles mapping in the messages. The default is:\n    roles:\n      user: [\"human\", \"user\"]\n      assistant: [\"gpt\", \"assistant\"]\n      system: [\"system\"]\n      tool: [\"tool\"]\n\n    # Optional[bool]. Whether to drop the system turn from the dataset. Only works with chat_template.\n    # This does not drop the default system message from chat_template if it exists. If you wish to,\n    # we recommend using a custom jinja template with the default system message removed or\n    # adding a system turn with empty content.\n    drop_system_message:\n\n    # IMPORTANT: The following fields determine which parts of the conversation to train on.\n    # Priority order: message_field_training &gt; message_field_training_detail &gt; train_on_inputs or role in roles_to_train\n    # See examples at `docs/dataset-formats/conversation.qmd`\n    # Note: If the below 4 fields are set to empty, defaults to training only on the last message.\n\n    # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.\n    roles_to_train: [\"assistant\"]  # default\n    # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:\n    # - all: train on all EOS tokens\n    # - turn (default): train on the EOS token at the end of each trainable turn\n    # - last: train on the last EOS token in the conversation\n    # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.\n    train_on_eos: last\n    # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.\n    message_field_training: training\n    # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.\n    # The value of the key is a List[Dict] containing `begin_offset` (start character index in content), `end_offset` (end character index in content), and `train` (boolean whether to train).\n    message_field_training_detail: train_detail\n\n\n# If false, the datasets will not be shuffled and will keep their original order in `datasets`.\n# The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.\nshuffle_merged_datasets: true\n\nDeduplicates datasets and test_datasets with identical entries.\ndataset_exact_deduplication: true\n\n# A list of one or more datasets to eval the model with.\n# You can use either test_datasets, or val_set_size, but not both.\ntest_datasets:\n  - path: /workspace/data/eval.jsonl\n    ds_type: json\n    # You need to specify a split. For \"json\" datasets the default split is called \"train\".\n    split: train\n    type: completion\n    data_files:\n      - /workspace/data/eval.jsonl\n\n# use RL training: 'dpo', 'ipo', 'kto'\nrl:\n# whether to perform weighting if doing DPO training. Boolean.\ndpo_use_weighting:\n\n# reward modelling: `True` or `False`\nreward_model:\n\n# process reward modelling: `True` or `False`\nprocess_reward_model:\n\n# The name of the chat template to use for training, following values are supported:\n# - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.\n# - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py\n# - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to. E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not available in the tokenizer.\n# - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.\n# The selected chat template will be saved to the tokenizer_config.json for easier inferencing\n# Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template.\nchat_template: tokenizer_default\n# custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null.\nchat_template_jinja: null\n# Changes the default system message. Currently only supports chatml.\ndefault_system_message: You are a helpful assistant. Please give a long and detailed answer.\n# Axolotl attempts to save the dataset as an arrow after packing the data together so\n# subsequent training attempts load faster, relative path\ndataset_prepared_path: data/last_run_prepared\n# Push prepared dataset to hub\npush_dataset_to_hub: # repo path\n# The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`\n# if not set.\ndataset_processes: # defaults to os.cpu_count() if not set\n# Keep dataset in memory while preprocessing\n# Only needed if cached dataset is taking too much storage\ndataset_keep_in_memory:\n# push checkpoints to hub\nhub_model_id: # private repo path to push finetuned model\n# how to push checkpoints to hub\n# https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy\nhub_strategy:\n# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets\n# Required to be true when used in combination with `push_dataset_to_hub`\nhf_use_auth_token: # boolean\n# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.\nval_set_size: 0.04\n# Num shards for whole dataset\ndataset_shard_num:\n# Index of shard to use for whole dataset\ndataset_shard_idx:\n\n# The maximum length of an input to train with, this should typically be less than 2048\n# as most models have a token/context limit of 2048\nsequence_len: 2048\n# Pad inputs so each step uses constant sized buffers\n# This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently\npad_to_sequence_len:\n# Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'\nsample_packing:\n# Set to 'false' if getting errors during eval with sample_packing on.\neval_sample_packing:\n# You can set these packing optimizations AFTER starting a training at least once.\n# The trainer will provide recommended values for these values.\nsample_packing_eff_est:\ntotal_num_tokens:\n# Increasing the following values helps with packing, but usually only slightly (&lt;%1.)\n# The number of samples packed at a time.\nsample_packing_group_size: 100000\n# The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.\nsample_packing_bin_size: 200\n# whether to concatenate samples during pretraining\npretraining_sample_concatenation:\n\n# Use batch flattening for speedups when not using sample_packing\nbatch_flattening:\n\n# Passed through to transformers when loading the model when launched without accelerate\n# Use `sequential` when training w/ model parallelism to limit memory\ndevice_map:\n# Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model.\nmax_memory:\n\n# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model\nadapter: lora\n# If you already have a lora model trained that you want to load, put that here.\n# This means after training, if you want to test the model, you should set this to the value of `output_dir`.\n# Note that if you merge an adapter to the base model, a new subdirectory `merged` will be created under the `output_dir`.\nlora_model_dir:\n\n# LoRA hyperparameters\n# For more details about the following options, see:\n# https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2\nlora_r: 8\nlora_alpha: 16\nlora_dropout: 0.05\nlora_target_modules:\n  - q_proj\n  - v_proj\n#  - k_proj\n#  - o_proj\n#  - gate_proj\n#  - down_proj\n#  - up_proj\nlora_target_linear: # If true, will target all linear modules\npeft_layers_to_transform: # The layer indices to transform, otherwise, apply to all layers\n\n# If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.\n# For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.\n# `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.\n# https://github.com/huggingface/peft/issues/334#issuecomment-1561727994\nlora_modules_to_save:\n#  - embed_tokens\n#  - lm_head\n\nlora_fan_in_fan_out: false\n\n# Apply custom LoRA autograd functions and activation function Triton kernels for\n# speed and memory savings\n# See: https://axolotl-ai-cloud.github.io/axolotl/docs/lora_optims.html\nlora_mlp_kernel: true\nlora_qkv_kernel: true\nlora_o_kernel: true\n\n# LoRA+ hyperparameters\n# For more details about the following options, see:\n# https://arxiv.org/abs/2402.12354  and `src/axolotl/core/train_builder.py`\nloraplus_lr_ratio: # loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4.\nloraplus_lr_embedding: #  loraplus learning rate for lora embedding layers. Default value is 1e-6.\n\npeft:\n  # Configuration options for loftq initialization for LoRA\n  # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization\n  loftq_config:\n    loftq_bits:  # typically 4 bits\n\n# ReLoRA configuration\n# Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed\nrelora_steps: # Number of steps per ReLoRA restart\nrelora_warmup_steps: # Number of per-restart warmup steps\nrelora_anneal_steps: # Number of anneal steps for each relora cycle\nrelora_prune_ratio: # threshold for optimizer magnitude when pruning\nrelora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings\n\n# wandb configuration if you're using it\n# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.\nwandb_mode: # \"offline\" to save run metadata locally and not sync to the server, \"disabled\" to turn off wandb\nwandb_project: # Your wandb project name\nwandb_entity: # A wandb Team name if using a Team\nwandb_watch:\nwandb_name: # Set the name of your wandb run\nwandb_run_id: # Set the ID of your wandb run\nwandb_log_model: # \"checkpoint\" to log model to wandb Artifacts every `save_steps` or \"end\" to log only at the end of training\n\n# mlflow configuration if you're using it\nmlflow_tracking_uri: # URI to mlflow\nmlflow_experiment_name: # Your experiment name\nmlflow_run_name: # Your run name\nhf_mlflow_log_artifacts:  # set to true to copy each saved checkpoint on each save to mlflow artifact registry\n\n# Comet configuration if you're using it\n# Make sure your `COMET_API_KEY` environment variable is set (recommended) or you login to Comet with `comet login`.\n# Check out our documentation for more details https://www.comet.com/docs/v2/api-and-sdk/python-sdk/reference/Experiment-Creation/#comet_ml.start\nuse_comet: # Enable or disable Comet integration.\ncomet_api_key: # API key for Comet. Recommended to set via `comet login`.\ncomet_workspace: # Workspace name in Comet. Defaults to the user's default workspace.\ncomet_project_name: # Project name in Comet. Defaults to Uncategorized.\ncomet_experiment_key: # Identifier for the experiment. Used to append data to an existing experiment or control the key of new experiments. Default to a random key.\ncomet_mode: # Create a new experiment (\"create\") or log to an existing one (\"get\"). Default (\"get_or_create\") auto-selects based on configuration.\ncomet_online: # Set to True to log data to Comet server, or False for offline storage. Default is True.\ncomet_experiment_config: # Dictionary for additional configuration settings, see the doc for more details.\n\n# Tensorboard\nuse_tensorboard: # Optional[bool]\n\n# Where to save the full-finetuned model to\noutput_dir: ./completed-model\n\n# Whether to use torch.compile and which backend to use\n# setting to `auto` will enable torch compile when torch&gt;=2.5.1\ntorch_compile:  # Optional[Union[Literal[\"auto\"], bool]]\ntorch_compile_backend:  # Optional[str]\n\n# Training hyperparameters\n\n# If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.\ngradient_accumulation_steps: 1\n# The number of samples to include in each batch. This is the number of samples sent to each GPU.\n# Batch size per gpu = micro_batch_size * gradient_accumulation_steps\nmicro_batch_size: 2\neval_batch_size:\nnum_epochs: 4\nwarmup_steps: 100  # cannot use with warmup_ratio\nwarmup_ratio: 0.05  # cannot use with warmup_steps\nlearning_rate: 0.00003\nlr_quadratic_warmup:\nlogging_steps:\neval_steps: # Leave empty to eval at each epoch, integer for every N steps. float for fraction of total steps\nevals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps\neval_strategy: # Set to `\"no\"` to skip evaluation, `\"epoch\"` at end of each epoch, leave empty to infer from `eval_steps`.\nsave_strategy: # Set to `\"no\"` to skip checkpoint saves, `\"epoch\"` at end of each epoch, `\"best\"` when better result is achieved, leave empty to infer from `save_steps`.\nsave_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps\nsaves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps\nsave_total_limit: # Checkpoints saved at a time\n# Maximum number of iterations to train for. It precedes num_epochs which means that\n# if both are set, num_epochs will not be guaranteed.\n# e.g., when 1 epoch is 1000 steps =&gt; `num_epochs: 2` and `max_steps: 100` will train for 100 steps\nmax_steps:\n\n# bool of whether to include tokens trainer per second in the training metrics. This iterates over the entire dataset once, so it takes some time.\ninclude_tokens_per_second: # Optional[bool]\n\n# whether to find batch size that fits in memory. Passed to underlying transformers Trainer\nauto_find_batch_size: # Optional[bool]\n\neval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0\neval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128\neval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is [\"sacrebleu\", \"comet\", \"ter\", \"chrf\", \"perplexity\"]\n\nprofiler_steps: # enable the pytorch profiler to capture the first N steps of training to the output_dir.\n                # see https://pytorch.org/blog/understanding-gpu-memory-1/ for more information\n                # snapshots can be visualized @ https://pytorch.org/memory_viz\n\nloss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)\nloss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)\n\n# Save model as safetensors (require safetensors package)\nsave_safetensors:\n\n# Whether to mask out or include the human's prompt from the training labels\ntrain_on_inputs: false\n# Group similarly sized data to minimize padding.\n# May be slower to start, as it must download and sort the entire dataset.\n# Note that training loss may have an oscillating pattern with this enabled.\ngroup_by_length: false\n\n# Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing\ngradient_checkpointing: false\n# additional kwargs to pass to the trainer for gradient checkpointing\n# gradient_checkpointing_kwargs:\n#   use_reentrant: true\n\n# Stop training after this many evaluation losses have increased in a row\n# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback\nearly_stopping_patience: 3\n\n# Specify a scheduler and kwargs to use with the optimizer\nlr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine\nlr_scheduler_kwargs:\ncosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr\ncosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)\n\n# For one_cycle optim\nlr_div_factor: # Learning rate div factor\n\n# Specify optimizer\n# Valid values are driven by the Transformers OptimizerNames class, see:\n# https://github.com/huggingface/transformers/blob/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134\n#\n# Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of\n# torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used\n# in the examples/ for your model and fine-tuning use case.\n#\n# Valid values for 'optimizer' include:\n# - adamw_hf\n# - adamw_torch\n# - adamw_torch_fused\n# - adamw_torch_xla\n# - adamw_apex_fused\n# - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version &gt;= 2.5.1)\n# - adafactor\n# - adamw_anyprecision\n# - sgd\n# - adagrad\n# - adamw_bnb_8bit\n# - lion_8bit\n# - lion_32bit\n# - paged_adamw_32bit\n# - paged_adamw_8bit\n# - paged_lion_32bit\n# - paged_lion_8bit\n# - galore_adamw\n# - galore_adamw_8bit\n# - galore_adafactor\n# - galore_adamw_layerwise\n# - galore_adamw_8bit_layerwise\n# - galore_adafactor_layerwise\noptimizer:\n# Dictionary of arguments to pass to the optimizer\noptim_args:\n# For Galore Optimizers the following optim_args are available\n# rank:  # type: int\n# update_proj_gap  # type: int\n# scale  # type: float\n# proj_type:  # type: str, default = std\n\n# The target modules to optimize, i.e. the module names that you would like to train, right now this is used only for GaLore algorithm\noptim_target_modules:\n# - self_attn  # for llama\n# - mlp\n\n# Specify weight decay\nweight_decay:\n# adamw hyperparams\nadam_beta1:\nadam_beta2:\nadam_epsilon:\n# Gradient clipping max norm\nmax_grad_norm:\n\n# Augmentation techniques\n# NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings\n# currently only supported on Llama and Mistral\nneftune_noise_alpha:\n\n# Whether to bettertransformers\nflash_optimum:\n# Whether to use xformers attention patch https://github.com/facebookresearch/xformers:\nxformers_attention:\n# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:\nflash_attention:\nflash_attn_cross_entropy:  # Whether to use flash-attention cross entropy implementation - advanced use only\nflash_attn_rms_norm:  # Whether to use flash-attention rms norm implementation - advanced use only\nflash_attn_fuse_qkv: # Whether to fuse QKV into a single operation\nflash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation\n# Whether to use scaled-dot-product attention\n# https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html\nsdp_attention:\n# Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf\ns2_attention:\n# Optional[bool]. Whether to use low_cpu_mem_usage\nlow_cpu_mem_usage:\n# Resume from a specific checkpoint dir\nresume_from_checkpoint:\n# If resume_from_checkpoint isn't set and you simply want it to start where it left off.\n# Be careful with this being turned on between different models.\nauto_resume_from_checkpoints: false\n\n# Don't mess with this, it's here for accelerate and torchrun\nlocal_rank:\n\n# Add or change special tokens.\n# If you add tokens here, you don't need to add them to the `tokens` list.\nspecial_tokens:\n  # bos_token: \"&lt;s&gt;\"\n  # eos_token: \"&lt;/s&gt;\"\n  # unk_token: \"&lt;unk&gt;\"\n  # pad_token: \"[PAD]\"\n\n# Add extra tokens.\ntokens:\n\n# FSDP\nfsdp:\nfsdp_config:\n\n# Deepspeed config path. e.g., deepspeed_configs/zero3.json\ndeepspeed:\n\n# Advanced DDP Arguments\nddp_timeout:\nddp_bucket_cap_mb:\nddp_broadcast_buffers:\n\n# Path to torch distx for optim 'adamw_anyprecision'\ntorchdistx_path:\n\n# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize\npretraining_dataset:\n\n# Debug mode\ndebug:\n\n# Seed\nseed:\n\n# Allow overwrite yml config using from cli\nstrict:",
+    "crumbs": [
+      "Reference",
+      "Config options"
+    ]
+  },
+  {
+    "objectID": "docs/multi-gpu.html",
+    "href": "docs/multi-gpu.html",
+    "title": "Multi-GPU",
+    "section": "",
+    "text": "This guide covers advanced training configurations for multi-GPU setups using Axolotl.",
     "crumbs": [
       "Deployments",
-      "Mac M-series"
+      "Multi-GPU"
     ]
   },
   {
-    "objectID": "docs/custom_integrations.html",
-    "href": "docs/custom_integrations.html",
-    "title": "Custom Integrations",
+    "objectID": "docs/multi-gpu.html#sec-overview",
+    "href": "docs/multi-gpu.html#sec-overview",
+    "title": "Multi-GPU",
+    "section": "1 Overview",
+    "text": "1 Overview\nAxolotl supports several methods for multi-GPU training:\n\nDeepSpeed (recommended)\nFSDP (Fully Sharded Data Parallel)\nFSDP + QLoRA",
+    "crumbs": [
+      "Deployments",
+      "Multi-GPU"
+    ]
+  },
+  {
+    "objectID": "docs/multi-gpu.html#sec-deepspeed",
+    "href": "docs/multi-gpu.html#sec-deepspeed",
+    "title": "Multi-GPU",
+    "section": "2 DeepSpeed",
+    "text": "2 DeepSpeed\nDeepSpeed is the recommended approach for multi-GPU training due to its stability and performance. It provides various optimization levels through ZeRO stages.\n\n2.1 Configuration\nAdd to your YAML config:\ndeepspeed: deepspeed_configs/zero1.json\n\n\n2.2 Usage\n# Passing arg via config\naxolotl train config.yml\n\n# Passing arg via cli\naxolotl train config.yml --deepspeed deepspeed_configs/zero1.json\n\n\n2.3 ZeRO Stages\nWe provide default configurations for:\n\nZeRO Stage 1 (zero1.json)\nZeRO Stage 2 (zero2.json)\nZeRO Stage 3 (zero3.json)\n\nChoose based on your memory requirements and performance needs.",
+    "crumbs": [
+      "Deployments",
+      "Multi-GPU"
+    ]
+  },
+  {
+    "objectID": "docs/multi-gpu.html#sec-fsdp",
+    "href": "docs/multi-gpu.html#sec-fsdp",
+    "title": "Multi-GPU",
+    "section": "3 FSDP",
+    "text": "3 FSDP\n\n3.1 Basic FSDP Configuration\nfsdp:\n  - full_shard\n  - auto_wrap\nfsdp_config:\n  fsdp_offload_params: true\n  fsdp_state_dict_type: FULL_STATE_DICT\n  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer\n\n\n3.2 FSDP + QLoRA\nFor combining FSDP with QLoRA, see our dedicated guide.",
+    "crumbs": [
+      "Deployments",
+      "Multi-GPU"
+    ]
+  },
+  {
+    "objectID": "docs/multi-gpu.html#sec-performance",
+    "href": "docs/multi-gpu.html#sec-performance",
+    "title": "Multi-GPU",
+    "section": "4 Performance Optimization",
+    "text": "4 Performance Optimization\n\n4.1 Liger Kernel Integration\nPlease see docs for more info.",
+    "crumbs": [
+      "Deployments",
+      "Multi-GPU"
+    ]
+  },
+  {
+    "objectID": "docs/multi-gpu.html#sec-troubleshooting",
+    "href": "docs/multi-gpu.html#sec-troubleshooting",
+    "title": "Multi-GPU",
+    "section": "5 Troubleshooting",
+    "text": "5 Troubleshooting\n\n5.1 NCCL Issues\nFor NCCL-related problems, see our NCCL troubleshooting guide.\n\n\n5.2 Common Problems\n\nMemory IssuesTraining Instability\n\n\n\nReduce micro_batch_size\nReduce eval_batch_size\nAdjust gradient_accumulation_steps\nConsider using a higher ZeRO stage\n\n\n\n\nStart with DeepSpeed ZeRO-2\nMonitor loss values\nCheck learning rates\n\n\n\n\nFor more detailed troubleshooting, see our debugging guide.",
+    "crumbs": [
+      "Deployments",
+      "Multi-GPU"
+    ]
+  },
+  {
+    "objectID": "docs/installation.html",
+    "href": "docs/installation.html",
+    "title": "Installation",
     "section": "",
-    "text": "Axolotl adds custom features through integrations. They are located within the src/axolotl/integrations directory.\nTo enable them, please check the respective documentations.",
+    "text": "This guide covers all the ways you can install and set up Axolotl for your environment.",
     "crumbs": [
-      "Advanced Features",
-      "Custom Integrations"
+      "Getting Started",
+      "Installation"
     ]
   },
   {
-    "objectID": "docs/custom_integrations.html#cut-cross-entropy",
-    "href": "docs/custom_integrations.html#cut-cross-entropy",
-    "title": "Custom Integrations",
-    "section": "Cut Cross Entropy",
-    "text": "Cut Cross Entropy\nCut Cross Entropy reduces VRAM usage through optimization on the cross-entropy operation during loss calculation.\nSee https://github.com/apple/ml-cross-entropy\n\nUsage\nplugins:\n  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin\n\ncut_cross_entropy: true\n\n\nCitation\n@article{wijmans2024cut,\n  author       = {Erik Wijmans and\n                  Brody Huval and\n                  Alexander Hertzberg and\n                  Vladlen Koltun and\n                  Philipp Kr\\\"ahenb\\\"uhl},\n  title        = {Cut Your Losses in Large-Vocabulary Language Models},\n  journal      = {arXiv},\n  year         = {2024},\n  url          = {https://arxiv.org/abs/2411.09009},\n}\nPlease see reference here",
+    "objectID": "docs/installation.html#sec-requirements",
+    "href": "docs/installation.html#sec-requirements",
+    "title": "Installation",
+    "section": "1 Requirements",
+    "text": "1 Requirements\n\nNVIDIA GPU (Ampere architecture or newer for bf16 and Flash Attention) or AMD GPU\nPython ≥3.10\nPyTorch ≥2.4.1",
     "crumbs": [
-      "Advanced Features",
-      "Custom Integrations"
+      "Getting Started",
+      "Installation"
     ]
   },
   {
-    "objectID": "docs/custom_integrations.html#grokfast",
-    "href": "docs/custom_integrations.html#grokfast",
-    "title": "Custom Integrations",
-    "section": "Grokfast",
-    "text": "Grokfast\nSee https://github.com/ironjr/grokfast\n\nUsage\nplugins:\n  - axolotl.integrations.grokfast.GrokfastPlugin\n\ngrokfast_alpha: 2.0\ngrokfast_lamb: 0.98\n\n\nCitation\n@article{lee2024grokfast,\n    title={{Grokfast}: Accelerated Grokking by Amplifying Slow Gradients},\n    author={Lee, Jaerin and Kang, Bong Gyun and Kim, Kihoon and Lee, Kyoung Mu},\n    journal={arXiv preprint arXiv:2405.20233},\n    year={2024}\n}\nPlease see reference here",
+    "objectID": "docs/installation.html#sec-installation-methods",
+    "href": "docs/installation.html#sec-installation-methods",
+    "title": "Installation",
+    "section": "2 Installation Methods",
+    "text": "2 Installation Methods\n\n2.1 PyPI Installation (Recommended)\npip3 install --no-build-isolation axolotl[flash-attn,deepspeed]\nWe use --no-build-isolation in order to detect the installed PyTorch version (if installed) in order not to clobber it, and so that we set the correct version of dependencies that are specific to the PyTorch version or other installed co-dependencies.\n\n\n2.2 Edge/Development Build\nFor the latest features between releases:\ngit clone https://github.com/axolotl-ai-cloud/axolotl.git\ncd axolotl\npip3 install packaging ninja\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n\n\n2.3 Docker\ndocker run --gpus '\"all\"' --rm -it axolotlai/axolotl:main-latest\nFor development with Docker:\ndocker compose up -d\n\n\n\n\n\n\nAdvanced Docker Configuration\n\n\n\ndocker run --privileged --gpus '\"all\"' --shm-size 10g --rm -it \\\n  --name axolotl --ipc=host \\\n  --ulimit memlock=-1 --ulimit stack=67108864 \\\n  --mount type=bind,src=\"${PWD}\",target=/workspace/axolotl \\\n  -v ${HOME}/.cache/huggingface:/root/.cache/huggingface \\\n  axolotlai/axolotl:main-latest\n\n\nPlease refer to the Docker documentation for more information on the different Docker images that are available.",
     "crumbs": [
-      "Advanced Features",
-      "Custom Integrations"
+      "Getting Started",
+      "Installation"
     ]
   },
   {
-    "objectID": "docs/custom_integrations.html#knowledge-distillation-kd",
-    "href": "docs/custom_integrations.html#knowledge-distillation-kd",
-    "title": "Custom Integrations",
-    "section": "Knowledge Distillation (KD)",
-    "text": "Knowledge Distillation (KD)\n\nUsage\nplugins:\n  - \"axolotl.integrations.kd.KDPlugin\"\n\nkd_trainer: True\nkd_ce_alpha: 0.1\nkd_alpha: 0.9\nkd_temperature: 1.0\n\ntorch_compile: True  # torch&gt;=2.5.1, recommended to reduce vram\n\ndatasets:\n  - path: ...\n    type: \"axolotl.integrations.kd.chat_template\"\n    field_messages: \"messages_combined\"\n    logprobs_field: \"llm_text_generation_vllm_logprobs\"  # for kd only, field of logprobs\nAn example dataset can be found at axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample\nPlease see reference here",
+    "objectID": "docs/installation.html#sec-cloud",
+    "href": "docs/installation.html#sec-cloud",
+    "title": "Installation",
+    "section": "3 Cloud Environments",
+    "text": "3 Cloud Environments\n\n3.1 Cloud GPU Providers\nFor providers supporting Docker:\n\nUse axolotlai/axolotl-cloud:main-latest\nAvailable on:\n\nLatitude.sh\nJarvisLabs.ai\nRunPod\n\n\n\n\n3.2 Google Colab\nUse our example notebook.",
     "crumbs": [
-      "Advanced Features",
-      "Custom Integrations"
+      "Getting Started",
+      "Installation"
     ]
   },
   {
-    "objectID": "docs/custom_integrations.html#liger-kernels",
-    "href": "docs/custom_integrations.html#liger-kernels",
-    "title": "Custom Integrations",
-    "section": "Liger Kernels",
-    "text": "Liger Kernels\nLiger Kernel provides efficient Triton kernels for LLM training, offering:\n\n20% increase in multi-GPU training throughput\n60% reduction in memory usage\nCompatibility with both FSDP and DeepSpeed\n\nSee https://github.com/linkedin/Liger-Kernel\n\nUsage\nplugins:\n  - axolotl.integrations.liger.LigerPlugin\nliger_rope: true\nliger_rms_norm: true\nliger_glu_activation: true\nliger_layer_norm: true\nliger_fused_linear_cross_entropy: true\n\n\nCitation\n@article{hsu2024ligerkernelefficienttriton,\n      title={Liger Kernel: Efficient Triton Kernels for LLM Training},\n      author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},\n      year={2024},\n      eprint={2410.10989},\n      archivePrefix={arXiv},\n      primaryClass={cs.LG},\n      url={https://arxiv.org/abs/2410.10989},\n      journal={arXiv preprint arXiv:2410.10989},\n}\nPlease see reference here",
+    "objectID": "docs/installation.html#sec-platform-specific",
+    "href": "docs/installation.html#sec-platform-specific",
+    "title": "Installation",
+    "section": "4 Platform-Specific Instructions",
+    "text": "4 Platform-Specific Instructions\n\n4.1 macOS\npip3 install --no-build-isolation -e '.'\nSee Section 6 for Mac-specific issues.\n\n\n4.2 Windows\n\n\n\n\n\n\nImportant\n\n\n\nWe recommend using WSL2 (Windows Subsystem for Linux) or Docker.",
     "crumbs": [
-      "Advanced Features",
-      "Custom Integrations"
+      "Getting Started",
+      "Installation"
     ]
   },
   {
-    "objectID": "docs/custom_integrations.html#language-model-evaluation-harness-lm-eval",
-    "href": "docs/custom_integrations.html#language-model-evaluation-harness-lm-eval",
-    "title": "Custom Integrations",
-    "section": "Language Model Evaluation Harness (LM Eval)",
-    "text": "Language Model Evaluation Harness (LM Eval)\nRun evaluation on model using the popular lm-evaluation-harness library.\nSee https://github.com/EleutherAI/lm-evaluation-harness\n\nUsage\nplugins:\n  - axolotl.integrations.lm_eval.LMEvalPlugin\n\nlm_eval_tasks:\n  - gsm8k\n  - hellaswag\n  - arc_easy\n\nlm_eval_batch_size: # Batch size for evaluation\noutput_dir: # Directory to save evaluation results\n\n\nCitation\n@misc{eval-harness,\n  author       = {Gao, Leo and Tow, Jonathan and Abbasi, Baber and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and Le Noac'h, Alain and Li, Haonan and McDonell, Kyle and Muennighoff, Niklas and Ociepa, Chris and Phang, Jason and Reynolds, Laria and Schoelkopf, Hailey and Skowron, Aviya and Sutawika, Lintang and Tang, Eric and Thite, Anish and Wang, Ben and Wang, Kevin and Zou, Andy},\n  title        = {A framework for few-shot language model evaluation},\n  month        = 07,\n  year         = 2024,\n  publisher    = {Zenodo},\n  version      = {v0.4.3},\n  doi          = {10.5281/zenodo.12608602},\n  url          = {https://zenodo.org/records/12608602}\n}\nPlease see reference here",
+    "objectID": "docs/installation.html#sec-env-managers",
+    "href": "docs/installation.html#sec-env-managers",
+    "title": "Installation",
+    "section": "5 Environment Managers",
+    "text": "5 Environment Managers\n\n5.1 Conda/Pip venv\n\nInstall Python ≥3.10\nInstall PyTorch: https://pytorch.org/get-started/locally/\nInstall Axolotl:\npip3 install packaging\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n(Optional) Login to Hugging Face:\nhuggingface-cli login",
     "crumbs": [
-      "Advanced Features",
-      "Custom Integrations"
+      "Getting Started",
+      "Installation"
     ]
   },
   {
-    "objectID": "docs/custom_integrations.html#spectrum",
-    "href": "docs/custom_integrations.html#spectrum",
-    "title": "Custom Integrations",
-    "section": "Spectrum",
-    "text": "Spectrum\nby Eric Hartford, Lucas Atkins, Fernando Fernandes, David Golchinfar\nThis plugin contains code to freeze the bottom fraction of modules in a model, based on the Signal-to-Noise Ratio (SNR).\nSee https://github.com/cognitivecomputations/spectrum\n\nOverview\nSpectrum is a tool for scanning and evaluating the Signal-to-Noise Ratio (SNR) of layers in large language models. By identifying the top n% of layers with the highest SNR, you can optimize training efficiency.\n\n\nUsage\nplugins:\n  - axolotl.integrations.spectrum.SpectrumPlugin\n\nspectrum_top_fraction: 0.5\nspectrum_model_name: meta-llama/Meta-Llama-3.1-8B\n\n\nCitation\n@misc{hartford2024spectrumtargetedtrainingsignal,\n      title={Spectrum: Targeted Training on Signal to Noise Ratio},\n      author={Eric Hartford and Lucas Atkins and Fernando Fernandes Neto and David Golchinfar},\n      year={2024},\n      eprint={2406.06623},\n      archivePrefix={arXiv},\n      primaryClass={cs.LG},\n      url={https://arxiv.org/abs/2406.06623},\n}\nPlease see reference here",
+    "objectID": "docs/installation.html#sec-troubleshooting",
+    "href": "docs/installation.html#sec-troubleshooting",
+    "title": "Installation",
+    "section": "6 Troubleshooting",
+    "text": "6 Troubleshooting\nIf you encounter installation issues, see our FAQ and Debugging Guide.",
     "crumbs": [
-      "Advanced Features",
-      "Custom Integrations"
+      "Getting Started",
+      "Installation"
     ]
   },
   {
-    "objectID": "docs/dataset_preprocessing.html",
-    "href": "docs/dataset_preprocessing.html",
-    "title": "Dataset Preprocessing",
+    "objectID": "docs/torchao.html",
+    "href": "docs/torchao.html",
+    "title": "PyTorch ao",
     "section": "",
-    "text": "Dataset pre-processing is the step where Axolotl takes each dataset you’ve configured alongside the dataset format and prompt strategies to:\n\nparse the dataset based on the dataset format\ntransform the dataset to how you would interact with the model based on the prompt strategy\ntokenize the dataset based on the configured model & tokenizer\nshuffle and merge multiple datasets together if using more than one\n\nThe processing of the datasets can happen one of two ways:\n\nBefore kicking off training by calling axolotl preprocess config.yaml --debug\nWhen training is started\n\n\n\nWhen training interactively or for sweeps (e.g. you are restarting the trainer often), processing the datasets can oftentimes be frustratingly slow. Pre-processing will cache the tokenized/formatted datasets according to a hash of dependent training parameters so that it will intelligently pull from its cache when possible.\nThe path of the cache is controlled by dataset_prepared_path: and is often left blank in example YAMLs as this leads to a more robust solution that prevents unexpectedly reusing cached data.\nIf dataset_prepared_path: is left empty, when training, the processed dataset will be cached in a default path of ./last_run_prepared/, but will ignore anything already cached there. By explicitly setting dataset_prepared_path: ./last_run_prepared, the trainer will use whatever pre-processed data is in the cache.\n\n\n\nLet’s say you are writing a custom prompt strategy or using a user-defined prompt template. Because the trainer cannot readily detect these changes, we cannot change the calculated hash value for the pre-processed dataset.\nIf you have dataset_prepared_path: ... set and change your prompt templating logic, it may not pick up the changes you made and you will be training over the old prompt.",
+    "text": "To use experimental optimizers (AdamWFp8, AdamW4bit, AdamW8bit) from Pytorch Ao, please install the package as shown below.\n\n\n\n\n\n\nTip\n\n\n\nSome experimental optimizers are already present in regular Pytorch, so please re-check if you actually need this package!\n\n\n\nInstallation\nStable Release from the PyTorch index\npip install torchao --extra-index-url https://download.pytorch.org/whl/cu121 # full options are cpu/cu118/cu121/cu124\nNightly release\npip install --pre torchao-nightly --index-url https://download.pytorch.org/whl/nightly/cu121 # full options are cpu/cu118/cu121/cu124",
+    "crumbs": [
+      "Advanced Features",
+      "PyTorch ao"
+    ]
+  },
+  {
+    "objectID": "docs/reward_modelling.html",
+    "href": "docs/reward_modelling.html",
+    "title": "Reward Modelling",
+    "section": "",
+    "text": "Overview\nReward modelling is a technique used to train models to predict the reward or value of a given input. This is particularly useful in reinforcement learning scenarios where the model needs to evaluate the quality of its actions or predictions. We support the reward modelling techniques supported by trl.\n\n\n(Outcome) Reward Models\nOutcome reward models are trained using data which contains preference annotations for an entire interaction between the user and model (e.g. rather than per-turn or per-step).\nbase_model: google/gemma-2-2b\nmodel_type: AutoModelForSequenceClassification\nnum_labels: 1\ntokenizer_type: AutoTokenizer\n\nreward_model: true\nchat_template: gemma\ndatasets:\n  - path: argilla/distilabel-intel-orca-dpo-pairs\n    type: bradley_terry.chat_template\n\nval_set_size: 0.1\neval_steps: 100\n\n\nProcess Reward Models (PRM)\nProcess reward models are trained using data which contains preference annotations for each step in a series of interactions. Typically, PRMs are trained to provide reward signals over each step of a reasoning trace and are used for downstream reinforcement learning.\nbase_model: Qwen/Qwen2.5-3B\nmodel_type: AutoModelForTokenClassification\nnum_labels: 2\n\nprocess_reward_model: true\ndatasets:\n  - path: trl-lib/math_shepherd\n    type: stepwise_supervised\n    split: train\n\nval_set_size: 0.1\neval_steps: 100",
+    "crumbs": [
+      "How To Guides",
+      "Reward Modelling"
+    ]
+  },
+  {
+    "objectID": "docs/input_output.html",
+    "href": "docs/input_output.html",
+    "title": "Template-free prompt construction",
+    "section": "",
+    "text": "The documentation moved to here."
+  },
+  {
+    "objectID": "docs/multimodal.html",
+    "href": "docs/multimodal.html",
+    "title": "MultiModal / Vision Language Models (BETA)",
+    "section": "",
+    "text": "MultiModal / Vision Language Models (BETA)\n\nSupported Models\n\nMllama, i.e. llama with vision models\n\n\n\nUsage\nCurrently multimodal support is limited and doesn’t have full feature parity. To finetune a multimodal Llama w/ LoRA, you’ll need to use the following in YAML in combination with the rest of the required hyperparams.\nbase_model: alpindale/Llama-3.2-11B-Vision-Instruct\nprocessor_type: AutoProcessor\nskip_prepare_dataset: true\n\nchat_template: llama3_2_vision\ndatasets:\n  - path: HuggingFaceH4/llava-instruct-mix-vsft\n    type: chat_template\n    split: train[:1%]\n    field_messages: messages\nremove_unused_columns: false\nsample_packing: false\n\n# only finetune the Language model, leave the vision model and vision tower frozen\nlora_target_modules: 'language_model.model.layers.[\\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'",
+    "crumbs": [
+      "How To Guides",
+      "MultiModal / Vision Language Models (BETA)"
+    ]
+  },
+  {
+    "objectID": "docs/getting-started.html",
+    "href": "docs/getting-started.html",
+    "title": "Quickstart",
+    "section": "",
+    "text": "This guide will walk you through your first model fine-tuning project with Axolotl.",
+    "crumbs": [
+      "Getting Started",
+      "Quickstart"
+    ]
+  },
+  {
+    "objectID": "docs/getting-started.html#sec-quick-example",
+    "href": "docs/getting-started.html#sec-quick-example",
+    "title": "Quickstart",
+    "section": "1 Quick Example",
+    "text": "1 Quick Example\nLet’s start by fine-tuning a small language model using LoRA. This example uses a 1B parameter model to ensure it runs on most GPUs. Assuming axolotl is installed (if not, see our Installation Guide)\n\nDownload example configs:\n\naxolotl fetch examples\n\nRun the training:\n\naxolotl train examples/llama-3/lora-1b.yml\nThat’s it! Let’s understand what just happened.",
+    "crumbs": [
+      "Getting Started",
+      "Quickstart"
+    ]
+  },
+  {
+    "objectID": "docs/getting-started.html#sec-understanding",
+    "href": "docs/getting-started.html#sec-understanding",
+    "title": "Quickstart",
+    "section": "2 Understanding the Process",
+    "text": "2 Understanding the Process\n\n2.1 The Configuration File\nThe YAML configuration file controls everything about your training. Here’s what (part of) our example config looks like:\nbase_model: NousResearch/Llama-3.2-1B\n# hub_model_id: username/custom_model_name\n\ndatasets:\n  - path: teknium/GPT4-LLM-Cleaned\n    type: alpaca\ndataset_prepared_path: last_run_prepared\nval_set_size: 0.1\noutput_dir: ./outputs/lora-out\n\nadapter: lora\nlora_model_dir:\nSee our Config options for more details.\n\n\n2.2 Training\nWhen you run axolotl train, Axolotl:\n\nDownloads the base model\n(If specified) applies LoRA adapter layers\nLoads and processes the dataset\nRuns the training loop\nSaves the trained model and / or LoRA weights",
+    "crumbs": [
+      "Getting Started",
+      "Quickstart"
+    ]
+  },
+  {
+    "objectID": "docs/getting-started.html#sec-custom",
+    "href": "docs/getting-started.html#sec-custom",
+    "title": "Quickstart",
+    "section": "3 Your First Custom Training",
+    "text": "3 Your First Custom Training\nLet’s modify the example for your own data:\n\nCreate a new config file my_training.yml:\n\nbase_model: NousResearch/Nous-Hermes-llama-1b-v1\nadapter: lora\n\n# Training settings\nmicro_batch_size: 2\nnum_epochs: 3\nlearning_rate: 0.0003\n\n# Your dataset\ndatasets:\n  - path: my_data.jsonl        # Your local data file\n    type: alpaca               # Or other format\nThis specific config is for LoRA fine-tuning a model with instruction tuning data using the alpaca dataset format, which has the following format:\n{\n    \"instruction\": \"Write a description of alpacas.\",\n    \"input\": \"\",\n    \"output\": \"Alpacas are domesticated South American camelids...\"\n}\nPlease see our Dataset Formats for more dataset formats and how to format them.\n\nPrepare your JSONL data in the specified format (in this case, the expected `alpaca format):\n\n{\"instruction\": \"Classify this text\", \"input\": \"I love this!\", \"output\": \"positive\"}\n{\"instruction\": \"Classify this text\", \"input\": \"Not good at all\", \"output\": \"negative\"}\nPlease consult the supported Dataset Formats for more details.\n\nRun the training:\n\naxolotl train my_training.yml",
+    "crumbs": [
+      "Getting Started",
+      "Quickstart"
+    ]
+  },
+  {
+    "objectID": "docs/getting-started.html#sec-common-tasks",
+    "href": "docs/getting-started.html#sec-common-tasks",
+    "title": "Quickstart",
+    "section": "4 Common Tasks",
+    "text": "4 Common Tasks\n\n4.1 Testing Your Model\nAfter training, test your model:\naxolotl inference my_training.yml --lora-model-dir=\"./outputs/lora-out\"\n\n\n4.2 Preprocessing Data\nFor large datasets, preprocess first:\naxolotl preprocess my_training.yml\n\n\n4.3 Using a UI\nLaunch a Gradio interface:\naxolotl inference my_training.yml --lora-model-dir=\"./outputs/lora-out\" --gradio",
+    "crumbs": [
+      "Getting Started",
+      "Quickstart"
+    ]
+  },
+  {
+    "objectID": "docs/getting-started.html#sec-next-steps",
+    "href": "docs/getting-started.html#sec-next-steps",
+    "title": "Quickstart",
+    "section": "5 Next Steps",
+    "text": "5 Next Steps\nNow that you have the basics, you might want to:\n\nTry different model architectures\nExperiment with hyperparameters\nUse more advanced training methods\nScale up to larger models\n\nCheck our other guides for details on these topics:\n\nConfiguration Guide - Full configuration options\nDataset Formats - Working with different data formats\nMulti-GPU Training\nMulti-Node Training",
+    "crumbs": [
+      "Getting Started",
+      "Quickstart"
+    ]
+  },
+  {
+    "objectID": "docs/inference.html",
+    "href": "docs/inference.html",
+    "title": "Inference",
+    "section": "",
+    "text": "This guide covers how to use your trained models for inference, including model loading, interactive testing, and common troubleshooting steps.",
+    "crumbs": [
+      "Getting Started",
+      "Inference"
+    ]
+  },
+  {
+    "objectID": "docs/inference.html#sec-quickstart",
+    "href": "docs/inference.html#sec-quickstart",
+    "title": "Inference",
+    "section": "1 Quick Start",
+    "text": "1 Quick Start\n\n1.1 Basic Inference\n\nLoRA ModelsFull Fine-tuned Models\n\n\naxolotl inference your_config.yml --lora-model-dir=\"./lora-output-dir\"\n\n\naxolotl inference your_config.yml --base-model=\"./completed-model\"",
+    "crumbs": [
+      "Getting Started",
+      "Inference"
+    ]
+  },
+  {
+    "objectID": "docs/inference.html#sec-advanced",
+    "href": "docs/inference.html#sec-advanced",
+    "title": "Inference",
+    "section": "2 Advanced Usage",
+    "text": "2 Advanced Usage\n\n2.1 Gradio Interface\nLaunch an interactive web interface:\naxolotl inference your_config.yml --gradio\n\n\n2.2 File-based Prompts\nProcess prompts from a text file:\ncat /tmp/prompt.txt | axolotl inference your_config.yml \\\n  --base-model=\"./completed-model\" --prompter=None\n\n\n2.3 Memory Optimization\nFor large models or limited memory:\naxolotl inference your_config.yml --load-in-8bit=True",
+    "crumbs": [
+      "Getting Started",
+      "Inference"
+    ]
+  },
+  {
+    "objectID": "docs/inference.html#sec-merging",
+    "href": "docs/inference.html#sec-merging",
+    "title": "Inference",
+    "section": "3 Merging LoRA Weights",
+    "text": "3 Merging LoRA Weights\nMerge LoRA adapters with the base model:\naxolotl merge-lora your_config.yml --lora-model-dir=\"./completed-model\"\n\n3.1 Memory Management for Merging\n\nConfiguration OptionsForce CPU Merging\n\n\ngpu_memory_limit: 20GiB  # Adjust based on your GPU\nlora_on_cpu: true        # Process on CPU if needed\n\n\nCUDA_VISIBLE_DEVICES=\"\" axolotl merge-lora ...",
+    "crumbs": [
+      "Getting Started",
+      "Inference"
+    ]
+  },
+  {
+    "objectID": "docs/inference.html#sec-tokenization",
+    "href": "docs/inference.html#sec-tokenization",
+    "title": "Inference",
+    "section": "4 Tokenization",
+    "text": "4 Tokenization\n\n4.1 Common Issues\n\n\n\n\n\n\nWarning\n\n\n\nTokenization mismatches between training and inference are a common source of problems.\n\n\nTo debug:\n\nCheck training tokenization:\n\naxolotl preprocess your_config.yml --debug\n\nVerify inference tokenization by decoding tokens before model input\nCompare token IDs between training and inference\n\n\n\n4.2 Special Tokens\nConfigure special tokens in your YAML:\nspecial_tokens:\n  bos_token: \"&lt;s&gt;\"\n  eos_token: \"&lt;/s&gt;\"\n  unk_token: \"&lt;unk&gt;\"\ntokens:\n  - \"&lt;|im_start|&gt;\"\n  - \"&lt;|im_end|&gt;\"",
+    "crumbs": [
+      "Getting Started",
+      "Inference"
+    ]
+  },
+  {
+    "objectID": "docs/inference.html#sec-troubleshooting",
+    "href": "docs/inference.html#sec-troubleshooting",
+    "title": "Inference",
+    "section": "5 Troubleshooting",
+    "text": "5 Troubleshooting\n\n5.1 Common Problems\n\nMemory IssuesToken IssuesPerformance Issues\n\n\n\nUse 8-bit loading\nReduce batch sizes\nTry CPU offloading\n\n\n\n\nVerify special tokens\nCheck tokenizer settings\nCompare training and inference preprocessing\n\n\n\n\nVerify model loading\nCheck prompt formatting\nEnsure temperature/sampling settings\n\n\n\n\nFor more details, see our debugging guide.",
+    "crumbs": [
+      "Getting Started",
+      "Inference"
+    ]
+  },
+  {
+    "objectID": "docs/multipack.html",
+    "href": "docs/multipack.html",
+    "title": "Multipack (Sample Packing)",
+    "section": "",
+    "text": "Because Flash Attention simply drops the attention mask, we do not need to construct a 4d attention mask. We only need to concatenate the sequences into a single batch and let flash attention know where each new sequence begins.\n4k context, bsz =4, each character represents 256 tokens X represents a padding token\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B ]\n   C C C C C C C ]\n   D D D D ]]\n\n[[ E E E E E E E E ]\n [ F F F F ]\n [ G G G ]\n [ H H H H ]]\n\n[[ I I I ]\n [ J J J ]\n [ K K K K K]\n [ L L L ]]\nafter padding to longest input in each step\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B X X X X X X ]\n   C C C C C C C X X X X ]\n   D D D D X X X X X X X ]]\n\n[[ E E E E E E E E ]\n [ F F F F X X X X ]\n [ G G G X X X X X ]\n [ H H H H X X X X ]]\n\n[[ I I I X X ]\n [ J J J X X ]\n [ K K K K K ]\n [ L L L X X ]]\nw packing ( note it’s the same effective number of tokens per step, but a true bsz of 1)\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A B B B B B\n   B C C C C C C C D D D D E E E E\n   E E E E F F F F F G G G H H H H\n   I I I J J J J K K K K K L L L X ]]\ncu_seqlens: [[ 0, 11, 17, 24, 28, 36, 41 44, 48, 51, 55, 60, 64]]",
     "crumbs": [
       "Core Concepts",
-      "Dataset Preprocessing"
+      "Multipack (Sample Packing)"
     ]
   },
   {
-    "objectID": "docs/dataset_preprocessing.html#overview",
-    "href": "docs/dataset_preprocessing.html#overview",
-    "title": "Dataset Preprocessing",
+    "objectID": "docs/multipack.html#visualization-of-multipack-with-flash-attention",
+    "href": "docs/multipack.html#visualization-of-multipack-with-flash-attention",
+    "title": "Multipack (Sample Packing)",
     "section": "",
-    "text": "Dataset pre-processing is the step where Axolotl takes each dataset you’ve configured alongside the dataset format and prompt strategies to:\n\nparse the dataset based on the dataset format\ntransform the dataset to how you would interact with the model based on the prompt strategy\ntokenize the dataset based on the configured model & tokenizer\nshuffle and merge multiple datasets together if using more than one\n\nThe processing of the datasets can happen one of two ways:\n\nBefore kicking off training by calling axolotl preprocess config.yaml --debug\nWhen training is started\n\n\n\nWhen training interactively or for sweeps (e.g. you are restarting the trainer often), processing the datasets can oftentimes be frustratingly slow. Pre-processing will cache the tokenized/formatted datasets according to a hash of dependent training parameters so that it will intelligently pull from its cache when possible.\nThe path of the cache is controlled by dataset_prepared_path: and is often left blank in example YAMLs as this leads to a more robust solution that prevents unexpectedly reusing cached data.\nIf dataset_prepared_path: is left empty, when training, the processed dataset will be cached in a default path of ./last_run_prepared/, but will ignore anything already cached there. By explicitly setting dataset_prepared_path: ./last_run_prepared, the trainer will use whatever pre-processed data is in the cache.\n\n\n\nLet’s say you are writing a custom prompt strategy or using a user-defined prompt template. Because the trainer cannot readily detect these changes, we cannot change the calculated hash value for the pre-processed dataset.\nIf you have dataset_prepared_path: ... set and change your prompt templating logic, it may not pick up the changes you made and you will be training over the old prompt.",
+    "text": "Because Flash Attention simply drops the attention mask, we do not need to construct a 4d attention mask. We only need to concatenate the sequences into a single batch and let flash attention know where each new sequence begins.\n4k context, bsz =4, each character represents 256 tokens X represents a padding token\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B ]\n   C C C C C C C ]\n   D D D D ]]\n\n[[ E E E E E E E E ]\n [ F F F F ]\n [ G G G ]\n [ H H H H ]]\n\n[[ I I I ]\n [ J J J ]\n [ K K K K K]\n [ L L L ]]\nafter padding to longest input in each step\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B X X X X X X ]\n   C C C C C C C X X X X ]\n   D D D D X X X X X X X ]]\n\n[[ E E E E E E E E ]\n [ F F F F X X X X ]\n [ G G G X X X X X ]\n [ H H H H X X X X ]]\n\n[[ I I I X X ]\n [ J J J X X ]\n [ K K K K K ]\n [ L L L X X ]]\nw packing ( note it’s the same effective number of tokens per step, but a true bsz of 1)\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A B B B B B\n   B C C C C C C C D D D D E E E E\n   E E E E F F F F F G G G H H H H\n   I I I J J J J K K K K K L L L X ]]\ncu_seqlens: [[ 0, 11, 17, 24, 28, 36, 41 44, 48, 51, 55, 60, 64]]",
     "crumbs": [
       "Core Concepts",
-      "Dataset Preprocessing"
+      "Multipack (Sample Packing)"
     ]
   },
   {
-    "objectID": "docs/fsdp_qlora.html",
-    "href": "docs/fsdp_qlora.html",
-    "title": "FDSP + QLoRA",
-    "section": "",
-    "text": "Using FSDP with QLoRA is essential for fine-tuning larger (70b+ parameter) LLMs on consumer GPUs. For example, you can use FSDP + QLoRA to train a 70b model on two 24GB GPUs1.\nBelow, we describe how to use this feature in Axolotl.",
-    "crumbs": [
-      "Advanced Features",
-      "FDSP + QLoRA"
-    ]
-  },
-  {
-    "objectID": "docs/fsdp_qlora.html#background",
-    "href": "docs/fsdp_qlora.html#background",
-    "title": "FDSP + QLoRA",
-    "section": "",
-    "text": "Using FSDP with QLoRA is essential for fine-tuning larger (70b+ parameter) LLMs on consumer GPUs. For example, you can use FSDP + QLoRA to train a 70b model on two 24GB GPUs1.\nBelow, we describe how to use this feature in Axolotl.",
-    "crumbs": [
-      "Advanced Features",
-      "FDSP + QLoRA"
-    ]
-  },
-  {
-    "objectID": "docs/fsdp_qlora.html#usage",
-    "href": "docs/fsdp_qlora.html#usage",
-    "title": "FDSP + QLoRA",
-    "section": "Usage",
-    "text": "Usage\nTo enable QLoRA with FSDP, you need to perform the following steps:\n\n![Tip] See the example config file in addition to reading these instructions.\n\n\nSet adapter: qlora in your axolotl config file.\nEnable FSDP in your axolotl config, as described here.\nUse one of the supported model types: llama, mistral or mixtral.",
-    "crumbs": [
-      "Advanced Features",
-      "FDSP + QLoRA"
-    ]
-  },
-  {
-    "objectID": "docs/fsdp_qlora.html#example-config",
-    "href": "docs/fsdp_qlora.html#example-config",
-    "title": "FDSP + QLoRA",
-    "section": "Example Config",
-    "text": "Example Config\nexamples/llama-2/qlora-fsdp.yml contains an example of how to enable QLoRA + FSDP in axolotl.",
-    "crumbs": [
-      "Advanced Features",
-      "FDSP + QLoRA"
-    ]
-  },
-  {
-    "objectID": "docs/fsdp_qlora.html#references",
-    "href": "docs/fsdp_qlora.html#references",
-    "title": "FDSP + QLoRA",
-    "section": "References",
-    "text": "References\n\nPR #1378 enabling QLoRA in FSDP in Axolotl.\nBlog Post from the Answer.AI team describing the work that enabled QLoRA in FSDP.\nRelated HuggingFace PRs Enabling FDSP + QLoRA:\n\nAccelerate PR#2544\nTransformers PR#29587\nTRL PR#1416\nPEFT PR#1550",
-    "crumbs": [
-      "Advanced Features",
-      "FDSP + QLoRA"
-    ]
-  },
-  {
-    "objectID": "docs/fsdp_qlora.html#footnotes",
-    "href": "docs/fsdp_qlora.html#footnotes",
-    "title": "FDSP + QLoRA",
-    "section": "Footnotes",
-    "text": "Footnotes\n\n\nThis was enabled by this work from the Answer.AI team.↩︎",
-    "crumbs": [
-      "Advanced Features",
-      "FDSP + QLoRA"
-    ]
-  },
-  {
-    "objectID": "docs/unsloth.html",
-    "href": "docs/unsloth.html",
-    "title": "Unsloth",
-    "section": "",
-    "text": "Overview\nUnsloth provides hand-written optimized kernels for LLM finetuning that slightly improve speed and VRAM over standard industry baselines.\n\n\n\n\n\n\nImportant\n\n\n\nDue to breaking changes in transformers v4.48.0, users will need to downgrade to &lt;=v4.47.1 to use this patch.\nThis will later be deprecated in favor of LoRA Optimizations.\n\n\n\n\nInstallation\nThe following will install the correct unsloth and extras from source.\npython scripts/unsloth_install.py | sh\n\n\nUsage\nAxolotl exposes a few configuration options to try out unsloth and get most of the performance gains.\nOur unsloth integration is currently limited to the following model architectures: - llama\nThese options are specific to LoRA finetuning and cannot be used for multi-GPU finetuning\nunsloth_lora_mlp: true\nunsloth_lora_qkv: true\nunsloth_lora_o: true\nThese options are composable and can be used with multi-gpu finetuning\nunsloth_cross_entropy_loss: true\nunsloth_rms_norm: true\nunsloth_rope: true\n\n\nLimitations\n\nSingle GPU only; e.g. no multi-gpu support\nNo deepspeed or FSDP support (requires multi-gpu)\nLoRA + QLoRA support only. No full fine tunes or fp8 support.\nLimited model architecture support. Llama, Phi, Gemma, Mistral only\nNo MoE support.",
-    "crumbs": [
-      "Advanced Features",
-      "Unsloth"
-    ]
-  },
-  {
-    "objectID": "docs/cli.html",
-    "href": "docs/cli.html",
-    "title": "CLI Reference",
-    "section": "",
-    "text": "The Axolotl CLI provides a streamlined interface for training and fine-tuning large language models. This guide covers the CLI commands, their usage, and common examples.",
-    "crumbs": [
-      "Getting Started",
-      "CLI Reference"
-    ]
-  },
-  {
-    "objectID": "docs/cli.html#basic-commands",
-    "href": "docs/cli.html#basic-commands",
-    "title": "CLI Reference",
-    "section": "Basic Commands",
-    "text": "Basic Commands\nAll Axolotl commands follow this general structure:\naxolotl &lt;command&gt; [config.yml] [options]\nThe config file can be local or a URL to a raw YAML file.",
-    "crumbs": [
-      "Getting Started",
-      "CLI Reference"
-    ]
-  },
-  {
-    "objectID": "docs/cli.html#command-reference",
-    "href": "docs/cli.html#command-reference",
-    "title": "CLI Reference",
-    "section": "Command Reference",
-    "text": "Command Reference\n\nfetch\nDownloads example configurations and deepspeed configs to your local machine.\n# Get example YAML files\naxolotl fetch examples\n\n# Get deepspeed config files\naxolotl fetch deepspeed_configs\n\n# Specify custom destination\naxolotl fetch examples --dest path/to/folder\n\n\npreprocess\nPreprocesses and tokenizes your dataset before training. This is recommended for large datasets.\n# Basic preprocessing\naxolotl preprocess config.yml\n\n# Preprocessing with one GPU\nCUDA_VISIBLE_DEVICES=\"0\" axolotl preprocess config.yml\n\n# Debug mode to see processed examples\naxolotl preprocess config.yml --debug\n\n# Debug with limited examples\naxolotl preprocess config.yml --debug --debug-num-examples 5\nConfiguration options:\ndataset_prepared_path: Local folder for saving preprocessed data\npush_dataset_to_hub: HuggingFace repo to push preprocessed data (optional)\n\n\ntrain\nTrains or fine-tunes a model using the configuration specified in your YAML file.\n# Basic training\naxolotl train config.yml\n\n# Train and set/override specific options\naxolotl train config.yml \\\n    --learning-rate 1e-4 \\\n    --micro-batch-size 2 \\\n    --num-epochs 3\n\n# Training without accelerate\naxolotl train config.yml --no-accelerate\n\n# Resume training from checkpoint\naxolotl train config.yml --resume-from-checkpoint path/to/checkpoint\nIt is possible to run sweeps over multiple hyperparameters by passing in a sweeps config.\n# Basic training with sweeps\naxolotl train config.yml --sweep path/to/sweep.yaml\nExample sweep config:\n_:\n  # This section is for dependent variables we need to fix\n  - load_in_8bit: false\n    load_in_4bit: false\n    adapter: lora\n  - load_in_8bit: true\n    load_in_4bit: false\n    adapter: lora\n\n# These are independent variables\nlearning_rate: [0.0003, 0.0006]\nlora_r:\n  - 16\n  - 32\nlora_alpha:\n  - 16\n  - 32\n  - 64\n\n\ninference\nRuns inference using your trained model in either CLI or Gradio interface mode.\n# CLI inference with LoRA\naxolotl inference config.yml --lora-model-dir=\"./outputs/lora-out\"\n\n# CLI inference with full model\naxolotl inference config.yml --base-model=\"./completed-model\"\n\n# Gradio web interface\naxolotl inference config.yml --gradio \\\n    --lora-model-dir=\"./outputs/lora-out\"\n\n# Inference with input from file\ncat prompt.txt | axolotl inference config.yml \\\n    --base-model=\"./completed-model\"\n\n\nmerge-lora\nMerges trained LoRA adapters into the base model.\n# Basic merge\naxolotl merge-lora config.yml\n\n# Specify LoRA directory (usually used with checkpoints)\naxolotl merge-lora config.yml --lora-model-dir=\"./lora-output/checkpoint-100\"\n\n# Merge using CPU (if out of GPU memory)\nCUDA_VISIBLE_DEVICES=\"\" axolotl merge-lora config.yml\nConfiguration options:\ngpu_memory_limit: Limit GPU memory usage\nlora_on_cpu: Load LoRA weights on CPU\n\n\nmerge-sharded-fsdp-weights\nMerges sharded FSDP model checkpoints into a single combined checkpoint.\n# Basic merge\naxolotl merge-sharded-fsdp-weights config.yml\n\n\nevaluate\nEvaluates a model’s performance using metrics specified in the config.\n# Basic evaluation\naxolotl evaluate config.yml\n\n\nlm-eval\nRuns LM Evaluation Harness on your model.\n# Basic evaluation\naxolotl lm-eval config.yml\nConfiguration options:\n# List of tasks to evaluate\nlm_eval_tasks:\n  - arc_challenge\n  - hellaswag\nlm_eval_batch_size: # Batch size for evaluation\noutput_dir: # Directory to save evaluation results",
-    "crumbs": [
-      "Getting Started",
-      "CLI Reference"
-    ]
-  },
-  {
-    "objectID": "docs/cli.html#legacy-cli-usage",
-    "href": "docs/cli.html#legacy-cli-usage",
-    "title": "CLI Reference",
-    "section": "Legacy CLI Usage",
-    "text": "Legacy CLI Usage\nWhile the new Click-based CLI is preferred, Axolotl still supports the legacy module-based CLI:\n# Preprocess\npython -m axolotl.cli.preprocess config.yml\n\n# Train\naccelerate launch -m axolotl.cli.train config.yml\n\n# Inference\naccelerate launch -m axolotl.cli.inference config.yml \\\n    --lora_model_dir=\"./outputs/lora-out\"\n\n# Gradio interface\naccelerate launch -m axolotl.cli.inference config.yml \\\n    --lora_model_dir=\"./outputs/lora-out\" --gradio\n\n\n\n\n\n\nImportant\n\n\n\nWhen overriding CLI parameters in the legacy CLI, use same notation as in yaml file (e.g., --lora_model_dir).\nNote: This differs from the new Click-based CLI, which uses dash notation (e.g., --lora-model-dir). Keep this in mind if you’re referencing newer documentation or switching between CLI versions.",
-    "crumbs": [
-      "Getting Started",
-      "CLI Reference"
-    ]
-  },
-  {
-    "objectID": "docs/cli.html#remote-compute-with-modal-cloud",
-    "href": "docs/cli.html#remote-compute-with-modal-cloud",
-    "title": "CLI Reference",
-    "section": "Remote Compute with Modal Cloud",
-    "text": "Remote Compute with Modal Cloud\nAxolotl supports running training and inference workloads on Modal cloud infrastructure. This is configured using a cloud YAML file alongside your regular Axolotl config.\n\nCloud Configuration\nCreate a cloud config YAML with your Modal settings:\n# cloud_config.yml\nprovider: modal\ngpu: a100  # Supported: l40s, a100-40gb, a100-80gb, a10g, h100, t4, l4\ngpu_count: 1    # Number of GPUs to use\ntimeout: 86400  # Maximum runtime in seconds (24 hours)\nbranch: main    # Git branch to use (optional)\n\nvolumes:        # Persistent storage volumes\n  - name: axolotl-cache\n    mount: /workspace/cache\n  - name: axolotl-data\n    mount: /workspace/data\n  - name: axolotl-artifacts\n    mount: /workspace/artifacts\n\nenv:            # Environment variables\n  - WANDB_API_KEY\n  - HF_TOKEN\n\n\nRunning on Modal Cloud\nCommands that support the –cloud flag:\n# Preprocess on cloud\naxolotl preprocess config.yml --cloud cloud_config.yml\n\n# Train on cloud\naxolotl train config.yml --cloud cloud_config.yml\n\n# Train without accelerate on cloud\naxolotl train config.yml --cloud cloud_config.yml --no-accelerate\n\n# Run lm-eval on cloud\naxolotl lm-eval config.yml --cloud cloud_config.yml\n\n\nCloud Configuration Options\nprovider: # compute provider, currently only `modal` is supported\ngpu: # GPU type to use\ngpu_count: # Number of GPUs (default: 1)\nmemory: # RAM in GB (default: 128)\ntimeout: # Maximum runtime in seconds\ntimeout_preprocess: # Preprocessing timeout\nbranch: # Git branch to use\ndocker_tag: # Custom Docker image tag\nvolumes: # List of persistent storage volumes\nenv: # Environment variables to pass\nsecrets: # Secrets to inject",
-    "crumbs": [
-      "Getting Started",
-      "CLI Reference"
-    ]
-  },
-  {
-    "objectID": "docs/rlhf.html",
-    "href": "docs/rlhf.html",
-    "title": "RLHF (Beta)",
-    "section": "",
-    "text": "Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human feedback. Various methods include, but not limited to:\n\nDirect Preference Optimization (DPO)\nIdentity Preference Optimization (IPO)\nKahneman-Tversky Optimization (KTO)\nOdds Ratio Preference Optimization (ORPO)\nProximal Policy Optimization (PPO) (not yet supported in axolotl)",
-    "crumbs": [
-      "How To Guides",
-      "RLHF (Beta)"
-    ]
-  },
-  {
-    "objectID": "docs/rlhf.html#overview",
-    "href": "docs/rlhf.html#overview",
-    "title": "RLHF (Beta)",
-    "section": "",
-    "text": "Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human feedback. Various methods include, but not limited to:\n\nDirect Preference Optimization (DPO)\nIdentity Preference Optimization (IPO)\nKahneman-Tversky Optimization (KTO)\nOdds Ratio Preference Optimization (ORPO)\nProximal Policy Optimization (PPO) (not yet supported in axolotl)",
-    "crumbs": [
-      "How To Guides",
-      "RLHF (Beta)"
-    ]
-  },
-  {
-    "objectID": "docs/rlhf.html#rlhf-using-axolotl",
-    "href": "docs/rlhf.html#rlhf-using-axolotl",
-    "title": "RLHF (Beta)",
-    "section": "RLHF using Axolotl",
-    "text": "RLHF using Axolotl\n\n\n\n\n\n\nImportant\n\n\n\nThis is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.\n\n\nWe rely on the TRL library for implementations of various RL training methods, which we wrap around to expose in axolotl. Each method has their own supported ways of loading datasets and prompt formats.\n\n\n\n\n\n\nTip\n\n\n\nYou can find what each method supports by going into src/axolotl/prompt_strategies/{method} where {method} is one of our supported methods. The type: can be retrieved from {method}.{function_name}.\n\n\n\nDPO\nExample config:\nrl: dpo\ndatasets:\n  - path: Intel/orca_dpo_pairs\n    split: train\n    type: chatml.intel\n  - path: argilla/ultrafeedback-binarized-preferences\n    split: train\n    type: chatml\nDPO supports the following types with the following dataset format:\n\nchatml.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"chosen_response\": \"...\",\n    \"rejected_response\": \"...\"\n}\n\n\nchatml.argilla_chat\n{\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nchatml.icr\n{\n    \"system\": \"...\", // optional\n    \"input\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nchatml.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nchatml.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nchatml.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nllama3.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"chosen_response\": \"...\",\n    \"rejected_response\": \"...\"\n}\n\n\nllama3.argilla_chat\n{\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nllama3.icr\n{\n    \"system\": \"...\", // optional\n    \"input\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nllama3.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nllama3.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nllama3.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nzephyr.nectar\n{\n    \"prompt\": \"...\",\n    \"answers\": [\n        {\n            \"answer\": \"...\",\n            \"rank\": 1\n        },\n        {\n            \"answer\": \"...\",\n            \"rank\": 2\n        }\n        // ... more answers with ranks\n    ]\n}\n\n\nchat_template.default\nrl: dpo\ndatasets:\n  - path: ...\n    split: train\n    type: chat_template.default\n    field_messages: \"messages\"\n    field_chosen: \"chosen\"\n    field_rejected: \"rejected\"\n    message_property_mappings:\n      role: role\n      content: content\n    roles:\n      user: [\"user\"]\n      assistant: [\"assistant\"]\n      system: [\"system\"]\nSample input format:\n{\n    \"messages\": [\n        {\n            \"role\": \"system\",\n            \"content\": \"...\"\n        },\n        {\n            \"role\": \"user\",\n            \"content\": \"...\"\n        },\n        // ... more messages\n    ],\n    \"chosen\": {\n        \"role\": \"assistant\",\n        \"content\": \"...\"\n    },\n    \"rejected\": {\n        \"role\": \"assistant\",\n        \"content\": \"...\"\n    }\n}\n\n\nuser_defined.default\nFor custom behaviors,\nrl: dpo\ndatasets:\n  - path: ...\n    split: train\n    type: user_defined.default\n\n    field_prompt: \"prompt\"\n    field_system: \"system\"\n    field_chosen: \"chosen\"\n    field_rejected: \"rejected\"\n    prompt_format: \"{prompt}\"\n    chosen_format: \"{chosen}\"\n    rejected_format: \"{rejected}\"\nThe input format is a simple JSON input with customizable fields based on the above config.\n{\n    \"system\": \"...\",  // optional\n    \"prompt\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\n\nIPO\nAs IPO is just DPO with a different loss function, all supported options for DPO works here.\nrl: ipo\n\n\nORPO\nPaper: https://arxiv.org/abs/2403.07691\nrl: orpo\norpo_alpha: 0.1\nremove_unused_columns: false\n\nchat_template: chatml\ndatasets:\n  - path: argilla/ultrafeedback-binarized-preferences-cleaned\n    type: chat_template.argilla\nORPO supports the following types with the following dataset format:\n\nchat_template.argilla\n{\n    \"system\": \"...\",  // optional\n    \"prompt\": \"...\",  // if available, will be taken as user message for single-turn instead of from list below\n\n    // chosen/rejected should be same till last content and only even-number of alternating user/assistant turns\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\n\nKTO\nrl: kto\nrl_beta: 0.5\nkto_desirable_weight: 0.2\n\nremove_unused_columns: false\n\ndatasets:\n  - path: argilla/ultrafeedback-binarized-preferences-cleaned-kto\n    type: llama3.ultra\n    split: train\n\ngradient_checkpointing: true\ngradient_checkpointing_kwargs:\n  use_reentrant: true\nKTO supports the following types with the following dataset format:\n\nchatml.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nchatml.argilla_chat\n{\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"}\n    ],\n    \"completion\": [\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nchatml.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nchatml.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nchatml.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.argilla_chat\n{\n    \"completion\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nllama3.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nuser_defined.default\nFor custom behaviors,\nrl: kto\ndatasets:\n  - path: ...\n    split: train\n    type: user_defined.default\n\n    field_prompt: \"prompt\"\n    field_system: \"system\"\n    field_completion: \"completion\"\n    field_label: \"label\"\n    prompt_format: \"{prompt}\"\n    completion_format: \"{completion}\"\nThe input format is a simple JSON input with customizable fields based on the above config.\n{\n    \"system\": \"...\",  // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\",\n    \"label\": \"...\"\n}\n\n\n\nGRPO\nGRPO uses custom reward functions and transformations. Please have them ready locally.\nFor ex, to load OpenAI’s GSM8K and use a random reward for completions:\n# rewards.py\nimport random\n\ndef rand_reward_func(completions, **kwargs) -&gt; list[float]:\n    return [random.uniform(0, 1) for _ in completions]\n\ndef oai_gsm8k_transform(cfg, *args, **kwargs):\n    def transform_fn(example, tokenizer=None):\n        label = example[\"answer\"].split(\"####\")[-1].strip().replace(\",\", \"\")\n        return {\n            \"prompt\": [{\"role\": \"user\", \"content\": example[\"question\"]},],\n            \"answer\": label,\n        }\n    return transform_fn, {\"remove_columns\": [\"question\"]}\nrl: grpo\n\ntrl:\n    beta: 0.001\n    max_completion_length: 256\n    use_vllm: True\n    vllm_device: auto\n    vllm_gpu_memory_utilization: 0.15\n    num_generations: 4\n    reward_funcs: [\"rewards.rand_reward_func\"]    # format: '{file_name}.{fn_name}'\ndatasets:\n  - path: openai/gsm8k\n    name: main\n    type: rewards.oai_gsm8k_transform  # format: '{file_name}.{fn_name}'\nTo see other examples of custom reward functions, please see TRL GRPO Docs.\n\n\nUsing local dataset files\ndatasets:\n  - ds_type: json\n    data_files:\n      - orca_rlhf.jsonl\n    split: train\n    type: chatml.intel\n\n\nTRL auto-unwrapping for PEFT\nTRL supports auto-unwrapping PEFT models for RL training paradigms which rely on a reference model. This significantly reduces memory pressure as an additional refreference model does not need to be loaded, and reference model log-probabilities can be obtained by disabling PEFT adapters. This is enabled by default. To turn it off, pass the following config:\n# load ref model when adapter training.\nrl_adapter_ref_model: true",
-    "crumbs": [
-      "How To Guides",
-      "RLHF (Beta)"
-    ]
-  },
-  {
-    "objectID": "docs/lora_optims.html",
-    "href": "docs/lora_optims.html",
-    "title": "LoRA Optimizations",
-    "section": "",
-    "text": "Inspired by Unsloth, we’ve implemented two optimizations for LoRA and QLoRA fine-tuning, supporting both single GPU and multi-GPU (in the DDP and DeepSpeed settings) training. These include (1) SwiGLU and GEGLU activation function Triton kernels, and (2) LoRA MLP and attention custom autograd functions. Our goal was to leverage operator fusion and tensor re-use in order to improve speed and reduce memory usage during the forward and backward passes of these calculations.\nWe currently support several common model architectures, including (but not limited to):",
-    "crumbs": [
-      "How To Guides",
-      "LoRA Optimizations"
-    ]
-  },
-  {
-    "objectID": "docs/lora_optims.html#usage",
-    "href": "docs/lora_optims.html#usage",
-    "title": "LoRA Optimizations",
-    "section": "Usage",
-    "text": "Usage\nThese optimizations can be enabled in your Axolotl config YAML file. The lora_mlp_kernel option enables the optimized MLP path, while lora_qkv_kernel and lora_o_kernel enable the fused query-key-value projection and optimized output projection, respectively.\nlora_mlp_kernel: true\nlora_qkv_kernel: true\nlora_o_kernel: true",
-    "crumbs": [
-      "How To Guides",
-      "LoRA Optimizations"
-    ]
-  },
-  {
-    "objectID": "docs/lora_optims.html#requirements",
-    "href": "docs/lora_optims.html#requirements",
-    "title": "LoRA Optimizations",
-    "section": "Requirements",
-    "text": "Requirements\n\nOne or more NVIDIA or AMD GPUs (in order to use the Triton kernels)\n\nNote: Set TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 to enable memory-efficient attention on AMD GPUs\n\nTargeted LoRA adapters cannot use Dropout\n\nThis may limit model expressivity / cause overfitting\n\nTargeted LoRA adapters cannot have bias terms\n\nThis may limit model expressivity\n\n\nModels with pre-existing LoRA adapters that use Dropout or have bias terms may need to be re-finetuned without these features in order to be useful.",
-    "crumbs": [
-      "How To Guides",
-      "LoRA Optimizations"
-    ]
-  },
-  {
-    "objectID": "docs/lora_optims.html#implementation-details",
-    "href": "docs/lora_optims.html#implementation-details",
-    "title": "LoRA Optimizations",
-    "section": "Implementation details",
-    "text": "Implementation details\n\nCustom autograd functions\nThe LoRA MLP autograd function optimizes the entire MLP computation path. It fuses the LoRA and base weight computations together and provides a single, efficient backward pass for the entire MLP block.\nFor attention components, similar optimizations are provided through a function that handles the query, key, and value projections, and a function that handles the output projection. They are designed to work with the existing transformers attention implementation via some monkey-patching logic.\n\n\nTriton kernels\nTwo activation functions (SwiGLU and GeGLU) are implemented with Triton kernels for improved speed and memory performance. These kernels handle both the forward and backward passes.\n\n\nIntegration\nThe custom autograd functions and Triton kernels are designed to work together. The autograd function manages the high-level computation flow and gradient tracking, while calling the Triton kernels for the activation function computation. During the backward pass, the kernel computes both the activation output and the required gradients, which the autograd function then uses to compute the final gradients for the entire computation path.",
-    "crumbs": [
-      "How To Guides",
-      "LoRA Optimizations"
-    ]
-  },
-  {
-    "objectID": "docs/lora_optims.html#future-work",
-    "href": "docs/lora_optims.html#future-work",
-    "title": "LoRA Optimizations",
-    "section": "Future Work",
-    "text": "Future Work\n\nSupport for additional model architectures\nSupport for the FSDP setting\nSupport for dropout and bias\nAdditional operator fusions",
-    "crumbs": [
-      "How To Guides",
-      "LoRA Optimizations"
-    ]
-  },
-  {
-    "objectID": "docs/batch_vs_grad.html",
-    "href": "docs/batch_vs_grad.html",
-    "title": "Batch size vs Gradient accumulation",
-    "section": "",
-    "text": "Gradient accumulation means accumulating gradients over several mini-batches and updating the model weights afterward. When the samples in each batch are diverse, this technique doesn’t significantly impact learning.\nThis method allows for effective training with larger effective batch sizes without needing proportionally larger memory. Here’s why:\n\nMemory Consumption with Batch Size: The primary reason increasing the batch size impacts memory is due to the storage requirements for intermediate activations. When you forward propagate a batch through a network, you have to store the activations at each layer for each sample in the batch, because these activations are used during backpropagation to compute gradients. Therefore, larger batches mean more activations, leading to greater GPU memory consumption.\nGradient Accumulation: With gradient accumulation, you’re effectively simulating a larger batch size by accumulating gradients over several smaller batches (or micro-batches). However, at any given time, you’re only forward and backward propagating a micro-batch. This means you only store activations for the micro-batch, not the full accumulated batch. As a result, you can simulate the effect of a larger batch size without the memory cost of storing activations for a large batch.\n\nExample 1: Micro batch size: 3 Gradient accumulation steps: 2 Number of GPUs: 3 Total batch size = 3 * 2 * 3 = 18\n| GPU 1          | GPU 2          | GPU 3          |\n|----------------|----------------|----------------|\n| S1, S2, S3     | S4, S5, S6     | S7, S8, S9     |\n| e1, e2, e3     | e4, e5, e6     | e7, e8, e9     |\n|----------------|----------------|----------------|\n| → (accumulate) | → (accumulate) | → (accumulate) |\n|----------------|----------------|----------------|\n| S10, S11, S12  | S13, S14, S15  | S16, S17, S18  |\n| e10, e11, e12  | e13, e14, e15  | e16, e17, e18  |\n|----------------|----------------|----------------|\n| → (apply)      | → (apply)      | → (apply)      |\n\nAccumulated gradient for the weight w1 after the second iteration (considering all GPUs):\nTotal gradient for w1 = e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8 + e9 + e10 + e11 + e12 + e13 + e14 + e15 + e16 + e17 + e18\n\nWeight update for w1:\nw1_new = w1_old - learning rate x (Total gradient for w1 / 18)\nExample 2: Micro batch size: 2 Gradient accumulation steps: 1 Number of GPUs: 3 Total batch size = 2 * 1 * 3 = 6\n| GPU 1     | GPU 2     | GPU 3     |\n|-----------|-----------|-----------|\n| S1, S2    | S3, S4    | S5, S6    |\n| e1, e2    | e3, e4    | e5, e6    |\n|-----------|-----------|-----------|\n| → (apply) | → (apply) | → (apply) |\n\nAccumulated gradient for the weight w1 (considering all GPUs):\nTotal gradient for w1 = e1 + e2 + e3 + e4 + e5 + e6\n\nWeight update for w1:\nw1_new = w1_old - learning rate × (Total gradient for w1 / 6)",
+    "objectID": "docs/multipack.html#multipack-without-flash-attention",
+    "href": "docs/multipack.html#multipack-without-flash-attention",
+    "title": "Multipack (Sample Packing)",
+    "section": "Multipack without Flash Attention",
+    "text": "Multipack without Flash Attention\nMultipack can still be achieved without Flash attention, but with lower packing efficiency as we are not able to join multiple batches into a single batch due to context length limits without flash attention. We can use either Pytorch’s Scaled Dot Product Attention implementation or native Pytorch attention implementation along with 4d attention masks to pack sequences together and avoid cross attention.",
     "crumbs": [
       "Core Concepts",
-      "Batch size vs Gradient accumulation"
+      "Multipack (Sample Packing)"
     ]
   },
   {
-    "objectID": "docs/faq.html",
-    "href": "docs/faq.html",
-    "title": "FAQ",
+    "objectID": "docs/debugging.html",
+    "href": "docs/debugging.html",
+    "title": "Debugging",
     "section": "",
-    "text": "General\nQ: The trainer stopped and hasn’t progressed in several minutes.\n\nA: Usually an issue with the GPUs communicating with each other. See the NCCL doc\n\nQ: Exitcode -9\n\nA: This usually happens when you run out of system RAM.\n\nQ: Exitcode -7 while using deepspeed\n\nA: Try upgrading deepspeed w: pip install -U deepspeed\n\nQ: AttributeError: ‘DummyOptim’ object has no attribute ‘step’\n\nA: You may be using deepspeed with single gpu. Please don’t set deepspeed: in yaml or cli.\n\nQ: The codes is stuck on saving preprocessed datasets.\n\nA: This is usually an issue with the GPU. This can be resolved through setting the os environment variable CUDA_VISIBLE_DEVICES=0. If you are on runpod, this is usually a pod issue. Starting a new pod should take care of it.\n\n\n\nChat templates\nQ: jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____\n\nA: This means that the property mapping for the stated attribute does not exist when building chat_template prompt. For example, if no attribute 'content', please check you have added the correct mapping for content under message_property_mappings.\n\nQ: Empty template generated for turn ___\n\nA: The content is empty for that turn.\n\nQ: Could not find content start/end boundary for turn __\n\nA: The specific turn’s start/end could not be detected. Please ensure you have set the eos_token following your chat_template. Otherwise, this could be a chat_template which doesn’t use proper boundaries for each turn (like system). On the rare occurrence, make sure your content is not [[dummy_message]]. Please let us know about this.\n\nQ: Content end boundary is before start boundary for turn ___\n\nA: This is an edge case which should not occur. Please create an Issue if this happens.\n\nQ: Content end boundary is the same as start boundary for turn ___. This is likely an empty turn.\n\nA: This is likely an empty turn.\n\nQ: The EOS/EOT token is incorrectly being masked or not being masked.\n\nA: This is because of the mismatch between tokenizer.eos_token and EOS/EOT token in template. Please make sure to set eos_token under special_tokens to the same EOS/EOT token as in template.",
+    "text": "This document provides some tips and tricks for debugging Axolotl. It also provides an example configuration for debugging with VSCode. A good debugging setup is essential to understanding how Axolotl code works behind the scenes.",
     "crumbs": [
       "Troubleshooting",
-      "FAQ"
+      "Debugging"
     ]
   },
   {
-    "objectID": "docs/multi-node.html",
-    "href": "docs/multi-node.html",
-    "title": "Multi Node",
+    "objectID": "docs/debugging.html#table-of-contents",
+    "href": "docs/debugging.html#table-of-contents",
+    "title": "Debugging",
+    "section": "Table of Contents",
+    "text": "Table of Contents\n\nGeneral Tips\nDebugging with VSCode\n\nBackground\nConfiguration\nCustomizing your debugger\nVideo Tutorial\n\nDebugging With Docker\n\nSetup\nAttach To Container\nVideo - Attaching To Docker On Remote Host",
+    "crumbs": [
+      "Troubleshooting",
+      "Debugging"
+    ]
+  },
+  {
+    "objectID": "docs/debugging.html#general-tips",
+    "href": "docs/debugging.html#general-tips",
+    "title": "Debugging",
+    "section": "General Tips",
+    "text": "General Tips\nWhile debugging it’s helpful to simplify your test scenario as much as possible. Here are some tips for doing so:\n\n[!Important] All of these tips are incorporated into the example configuration for debugging with VSCode below.\n\n\nMake sure you are using the latest version of axolotl: This project changes often and bugs get fixed fast. Check your git branch and make sure you have pulled the latest changes from main.\nEliminate concurrency: Restrict the number of processes to 1 for both training and data preprocessing:\n\nSet CUDA_VISIBLE_DEVICES to a single GPU, ex: export CUDA_VISIBLE_DEVICES=0.\nSet dataset_processes: 1 in your axolotl config or run the training command with --dataset_processes=1.\n\nUse a small dataset: Construct or use a small dataset from HF Hub. When using a small dataset, you will often have to make sure sample_packing: False and eval_sample_packing: False to avoid errors. If you are in a pinch and don’t have time to construct a small dataset but want to use from the HF Hub, you can shard the data (this will still tokenize the entire dataset, but will only use a fraction of the data for training. For example, to shard the dataset into 20 pieces, add the following to your axolotl config):\ndatasets:\n    ...\n    shards: 20\nUse a small model: A good example of a small model is TinyLlama/TinyLlama-1.1B-Chat-v1.0.\nMinimize iteration time: Make sure the training loop finishes as fast as possible, with these settings.\n\nmicro_batch_size: 1\nmax_steps: 1\nval_set_size: 0\n\nClear Caches: Axolotl caches certain steps and so does the underlying HuggingFace trainer. You may want to clear some of these caches when debugging.\n\nData preprocessing: When debugging data preprocessing, which includes prompt template formation, you may want to delete the directory set in dataset_prepared_path: in your axolotl config. If you didn’t set this value, the default is last_run_prepared.\nHF Hub: If you are debugging data preprocessing, you should clear the relevant HF cache HuggingFace cache, by deleting the appropriate ~/.cache/huggingface/datasets/... folder(s).\nThe recommended approach is to redirect all outputs and caches to a temporary folder and delete selected subfolders before each run. This is demonstrated in the example configuration below.",
+    "crumbs": [
+      "Troubleshooting",
+      "Debugging"
+    ]
+  },
+  {
+    "objectID": "docs/debugging.html#debugging-with-vscode",
+    "href": "docs/debugging.html#debugging-with-vscode",
+    "title": "Debugging",
+    "section": "Debugging with VSCode",
+    "text": "Debugging with VSCode\n\nBackground\nThe below example shows how to configure VSCode to debug data preprocessing of the chat_template format. This is the format used when you have the following in your axolotl config:\ndatasets:\n  - path: &lt;path to your chat_template formatted dataset&gt; # example on HF Hub: fozziethebeat/alpaca_messages_2k_test\n    type: chat_template\n\n[!Important] If you are already familiar with advanced VSCode debugging, you can skip the below explanation and look at the files .vscode/launch.json and .vscode/tasks.json for an example configuration.\n\n\n[!Tip] If you prefer to watch a video, rather than read, you can skip to the video tutorial below (but doing both is recommended).\n\n\n\nSetup\nMake sure you have an editable install of Axolotl, which ensures that changes you make to the code are reflected at runtime. Run the following commands from the root of this project:\npip3 install packaging\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n\nRemote Hosts\nIf you developing on a remote host, you can easily use VSCode to debug remotely. To do so, you will need to follow this remote - SSH guide. You can also see the video below on Docker and Remote SSH debugging.\n\n\n\nConfiguration\nThe easiest way to get started is to modify the .vscode/launch.json file in this project. This is just an example configuration, so you may need to modify or copy it to suit your needs.\nFor example, to mimic the command cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_chat_template.yml, you would use the below configuration1. Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to devtools and set the env variable HF_HOME to a temporary folder that is later partially deleted. This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.\n// .vscode/launch.json\n{\n    \"version\": \"0.2.0\",\n    \"configurations\": [\n        {\n            \"name\": \"Debug axolotl prompt - chat_template\",\n            \"type\": \"python\",\n            \"module\": \"accelerate.commands.launch\",\n            \"request\": \"launch\",\n            \"args\": [\n                \"-m\", \"axolotl.cli.train\", \"dev_chat_template.yml\",\n                // The flags below simplify debugging by overriding the axolotl config\n                // with the debugging tips above.  Modify as needed.\n                \"--dataset_processes=1\",      // limits data preprocessing to one process\n                \"--max_steps=1\",              // limits training to just one step\n                \"--batch_size=1\",             // minimizes batch size\n                \"--micro_batch_size=1\",       // minimizes batch size\n                \"--val_set_size=0\",           // disables validation\n                \"--sample_packing=False\",     // disables sample packing which is necessary for small datasets\n                \"--eval_sample_packing=False\",// disables sample packing on eval set\n                \"--dataset_prepared_path=temp_debug/axolotl_outputs/data\", // send data outputs to a temp folder\n                \"--output_dir=temp_debug/axolotl_outputs/model\" // send model outputs to a temp folder\n                ],\n            \"console\": \"integratedTerminal\",      // show output in the integrated terminal\n            \"cwd\": \"${workspaceFolder}/devtools\", // set working directory to devtools from the root of the project\n            \"justMyCode\": true,                   // step through only axolotl code\n            \"env\": {\"CUDA_VISIBLE_DEVICES\": \"0\",  // Since we aren't doing distributed training, we need to limit to one GPU\n                    \"HF_HOME\": \"${workspaceFolder}/devtools/temp_debug/.hf-cache\"}, // send HF cache to a temp folder\n            \"preLaunchTask\": \"cleanup-for-dataprep\", // delete temp folders (see below)\n        }\n    ]\n}\nAdditional notes about this configuration:\n\nThe argument justMyCode is set to true such that you step through only the axolotl code. If you want to step into dependencies, set this to false.\nThe preLaunchTask: cleanup-for-dataprep is defined in .vscode/tasks.json and is used to delete the following folders before debugging, which is essential to ensure that the data pre-processing code is run from scratch:\n\n./devtools/temp_debug/axolotl_outputs\n./devtools/temp_debug/.hf-cache/datasets\n\n\n\n[!Tip] You may not want to delete these folders. For example, if you are debugging model training instead of data pre-processing, you may NOT want to delete the cache or output folders. You may also need to add additional tasks to the tasks.json file depending on your use case.\n\nBelow is the ./vscode/tasks.json file that defines the cleanup-for-dataprep task. This task is run before each debugging session when you use the above configuration. Note how there are two tasks that delete the two folders mentioned above. The third task cleanup-for-dataprep is a composite task that combines the two tasks. A composite task is necessary because VSCode does not allow you to specify multiple tasks in the preLaunchTask argument of the launch.json file.\n// .vscode/tasks.json\n// this file is used by launch.json\n{\n    \"version\": \"2.0.0\",\n    \"tasks\": [\n      // this task changes into the devtools directory and deletes the temp_debug/axolotl_outputs folder\n      {\n        \"label\": \"delete-outputs\",\n        \"type\": \"shell\",\n        \"command\": \"rm -rf temp_debug/axolotl_outputs\",\n        \"options\":{ \"cwd\": \"${workspaceFolder}/devtools\"},\n        \"problemMatcher\": []\n      },\n      // this task changes into the devtools directory and deletes the `temp_debug/.hf-cache/datasets` folder\n      {\n        \"label\": \"delete-temp-hf-dataset-cache\",\n        \"type\": \"shell\",\n        \"command\": \"rm -rf temp_debug/.hf-cache/datasets\",\n        \"options\":{ \"cwd\": \"${workspaceFolder}/devtools\"},\n        \"problemMatcher\": []\n      },\n        // this task combines the two tasks above\n      {\n       \"label\": \"cleanup-for-dataprep\",\n       \"dependsOn\": [\"delete-outputs\", \"delete-temp-hf-dataset-cache\"],\n      }\n    ]\n}\n\n\nCustomizing your debugger\nYour debugging use case may differ from the example above. The easiest thing to do is to put your own axolotl config in the devtools folder and modify the launch.json file to use your config. You may also want to modify the preLaunchTask to delete different folders or not delete anything at all.\n\n\nVideo Tutorial\nThe following video tutorial walks through the above configuration and demonstrates how to debug with VSCode, (click the image below to watch):\n\n\n\nHamel Husain’s tutorial: Debugging Axolotl w/VSCode",
+    "crumbs": [
+      "Troubleshooting",
+      "Debugging"
+    ]
+  },
+  {
+    "objectID": "docs/debugging.html#debugging-with-docker",
+    "href": "docs/debugging.html#debugging-with-docker",
+    "title": "Debugging",
+    "section": "Debugging With Docker",
+    "text": "Debugging With Docker\nUsing official Axolotl Docker images is a great way to debug your code, and is a very popular way to use Axolotl. Attaching VSCode to Docker takes a few more steps.\n\nSetup\nOn the host that is running axolotl (ex: if you are using a remote host), clone the axolotl repo and change your current directory to the root:\ngit clone https://github.com/axolotl-ai-cloud/axolotl\ncd axolotl\n\n[!Tip] If you already have axolotl cloned on your host, make sure you have the latest changes and change into the root of the project.\n\nNext, run the desired docker image and mount the current directory. Below is a docker command you can run to do this:2\ndocker run --privileged --gpus '\"all\"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src=\"${PWD}\",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface axolotlai/axolotl:main-py3.10-cu118-2.0.1\n\n[!Tip] To understand which containers are available, see the Docker section of the README and the DockerHub repo. For details of how the Docker containers are built, see axolotl’s Docker CI builds.\n\nYou will now be in the container. Next, perform an editable install of Axolotl:\npip3 install packaging\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n\n\nAttach To Container\nNext, if you are using a remote host, Remote into this host with VSCode. If you are using a local host, you can skip this step.\nNext, select Dev Containers: Attach to Running Container... using the command palette (CMD + SHIFT + P) in VSCode. You will be prompted to select a container to attach to. Select the container you just created. You will now be in the container with a working directory that is at the root of the project. Any changes you make to the code will be reflected both in the container and on the host.\nNow you are ready to debug as described above (see Debugging with VSCode).\n\n\nVideo - Attaching To Docker On Remote Host\nHere is a short video that demonstrates how to attach to a Docker container on a remote host:\n\n\n\nHamel Husain’s tutorial: Debugging Axolotl Part 2: Attaching to Docker on a Remote Host",
+    "crumbs": [
+      "Troubleshooting",
+      "Debugging"
+    ]
+  },
+  {
+    "objectID": "docs/debugging.html#footnotes",
+    "href": "docs/debugging.html#footnotes",
+    "title": "Debugging",
+    "section": "Footnotes",
+    "text": "Footnotes\n\n\nThe config actually mimics the command CUDA_VISIBLE_DEVICES=0 python -m accelerate.commands.launch -m axolotl.cli.train devtools/chat_template.yml, but this is the same thing.↩︎\nMany of the below flags are recommended best practices by Nvidia when using nvidia-container-toolkit. You can read more about these flags here.↩︎",
+    "crumbs": [
+      "Troubleshooting",
+      "Debugging"
+    ]
+  },
+  {
+    "objectID": "docs/lr_groups.html",
+    "href": "docs/lr_groups.html",
+    "title": "Learning Rate Groups",
     "section": "",
-    "text": "The below are three ways to train multi-node in Axolotl.",
+    "text": "Inspired by LoRA+, Axolotl allows practitioners to specify separate learning rates for each module or groups of modules in a model.",
     "crumbs": [
-      "Deployments",
-      "Multi Node"
+      "How To Guides",
+      "Learning Rate Groups"
     ]
   },
   {
-    "objectID": "docs/multi-node.html#accelerate",
-    "href": "docs/multi-node.html#accelerate",
-    "title": "Multi Node",
-    "section": "Accelerate",
-    "text": "Accelerate\nYou will need to create a configuration for accelerate, either by using accelerate config and follow the instructions or you can use one of the preset below:\n~/.cache/huggingface/accelerate/default_config.yaml\ncompute_environment: LOCAL_MACHINE\ndebug: false\ndistributed_type: FSDP\ndowncast_bf16: 'no'\nmachine_rank: 0 # Set to 0 for the main machine, increment by one for other machines\nmain_process_ip: 10.0.0.4 # Set to main machine's IP\nmain_process_port: 5000\nmain_training_function: main\nmixed_precision: bf16\nnum_machines: 2 # Change to the number of machines\nnum_processes: 4 # That's the total number of GPUs, (for example: if you have 2 machines with 4 GPU, put 8)\nrdzv_backend: static\nsame_network: true\ntpu_env: []\ntpu_use_cluster: false\ntpu_use_sudo: false\nuse_cpu: false\nConfigure your model to use FSDP in the Axolotl yaml. For example:\nfsdp:\n  - full_shard\n  - auto_wrap\nfsdp_config:\n  fsdp_offload_params: true\n  fsdp_state_dict_type: FULL_STATE_DICT\n  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer\nAll you have to do now is launch using accelerate as you would usually do on each machine and voila, the processes will start once you have launched accelerate on every machine.",
-    "crumbs": [
-      "Deployments",
-      "Multi Node"
-    ]
-  },
-  {
-    "objectID": "docs/multi-node.html#raytrain",
-    "href": "docs/multi-node.html#raytrain",
-    "title": "Multi Node",
-    "section": "Raytrain",
-    "text": "Raytrain\nPlease see ray train doc here.",
-    "crumbs": [
-      "Deployments",
-      "Multi Node"
-    ]
-  },
-  {
-    "objectID": "docs/multi-node.html#torchrun",
-    "href": "docs/multi-node.html#torchrun",
-    "title": "Multi Node",
-    "section": "Torchrun",
-    "text": "Torchrun\nIf you are using Infiniband, we recommend torchrun to utilize the full bandwidth.\nSet the following env (change buffersize/socketname depending on your system):\nexport NCCL_IB_DISABLE=0\nexport NCCL_SOCKET_IFNAME=\"eth0,en,eth,em,bond\"\nexport NCCL_BUFFSIZE=2097152\nRun the following on each node:\ntorchrun --nnodes $num_nodes --nproc_per_node $gpu_per_node --rdzv_id $rdzv_id --rdzv_backend c10d --rdzv_endpoint \"$head_node_ip:$head_node_port\" -m axolotl.cli.train config.yaml\nPlease make sure to substitute the placeholder variables.\n\nnum_nodes: Number of nodes (containing GPUs)\ngpu_per_node: Number of gpus per node\nhead_node_ip: IP of the head node (make sure other machines can connect to this)\nhead_node_port: Port of the head node (make sure other machines can connect to this. Default 29400)\nrdzv_id: A unique job ID that is used by the job across nodes.\n\n\n\n\n\n\n\nNote\n\n\n\nYou need to call axolotl.cli.train instead of axolotl train as the latter calls accelerate under the hood\n\n\nMore info on the available configs can be found on the Pytorch docs here",
-    "crumbs": [
-      "Deployments",
-      "Multi Node"
-    ]
-  },
-  {
-    "objectID": "FAQS.html",
-    "href": "FAQS.html",
-    "title": "FAQs",
+    "objectID": "docs/lr_groups.html#background",
+    "href": "docs/lr_groups.html#background",
+    "title": "Learning Rate Groups",
     "section": "",
-    "text": "FAQs\n\nCan you train StableLM with this? Yes, but only with a single GPU atm. Multi GPU support is coming soon! Just waiting on this PR\nWill this work with Deepspeed? That’s still a WIP, but setting export ACCELERATE_USE_DEEPSPEED=true should work in some cases\nError invalid argument at line 359 in file /workspace/bitsandbytes/csrc/pythonInterface.c /arrow/cpp/src/arrow/filesystem/s3fs.cc:2598:  arrow::fs::FinalizeS3 was not called even though S3 was initialized. This could lead to a segmentation fault at exit. Try reinstalling bitsandbytes and transformers from source."
+    "text": "Inspired by LoRA+, Axolotl allows practitioners to specify separate learning rates for each module or groups of modules in a model.",
+    "crumbs": [
+      "How To Guides",
+      "Learning Rate Groups"
+    ]
   },
   {
-    "objectID": "src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html",
-    "href": "src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html",
+    "objectID": "docs/lr_groups.html#example",
+    "href": "docs/lr_groups.html#example",
+    "title": "Learning Rate Groups",
+    "section": "Example",
+    "text": "Example\nlr_groups:\n  - name: o_proj\n    modules:\n      - self_attn.o_proj.weight\n    lr: 1e-6\n  - name: q_proj\n    modules:\n      - model.layers.2.self_attn.q_proj.weight\n    lr: 1e-5\n\nlearning_rate: 2e-5\nIn this example, we have a default learning rate of 2e-5 across the entire model, but we have a separate learning rate of 1e-6 for all the self attention o_proj modules across all layers, and a learning are of 1e-5 to the 3rd layer’s self attention q_proj module.",
+    "crumbs": [
+      "How To Guides",
+      "Learning Rate Groups"
+    ]
+  },
+  {
+    "objectID": "TODO.html",
+    "href": "TODO.html",
+    "title": "todo list",
+    "section": "",
+    "text": "[] Validation of parameters for combinations that won’t work\n\n\n\n\nFSDP offload and gradient_checkpointing - https://github.com/pytorch/pytorch/issues/82203\nadamw_bnb_8bit doesn’t play well with FSDP offload"
+  },
+  {
+    "objectID": "TODO.html#things-that-are-known-not-to-work",
+    "href": "TODO.html#things-that-are-known-not-to-work",
+    "title": "todo list",
+    "section": "",
+    "text": "FSDP offload and gradient_checkpointing - https://github.com/pytorch/pytorch/issues/82203\nadamw_bnb_8bit doesn’t play well with FSDP offload"
+  },
+  {
+    "objectID": "src/axolotl/integrations/LICENSE.html",
+    "href": "src/axolotl/integrations/LICENSE.html",
     "title": "Axolotl",
     "section": "",
-    "text": "Acknowledgements\nPortions of this Cut Cross Entropy Software may utilize the following copyrighted material, the use of which is hereby acknowledged.\n\nPyTorch\nFrom PyTorch:\n\nCopyright (c) 2016-     Facebook, Inc            (Adam Paszke)\nCopyright (c) 2014-     Facebook, Inc            (Soumith Chintala)\nCopyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)\nCopyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)\nCopyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)\nCopyright (c) 2011-2013 NYU                      (Clement Farabet)\nCopyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)\nCopyright (c) 2006      Idiap Research Institute (Samy Bengio)\nCopyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)\n\nFrom Caffe2:\n\nCopyright (c) 2016-present, Facebook Inc. All rights reserved.\n\nAll contributions by Facebook:\nCopyright (c) 2016 Facebook Inc.\n\nAll contributions by Google:\nCopyright (c) 2015 Google Inc.\nAll rights reserved.\n\nAll contributions by Yangqing Jia:\nCopyright (c) 2015 Yangqing Jia\nAll rights reserved.\n\nAll contributions by Kakao Brain:\nCopyright 2019-2020 Kakao Brain\n\nAll contributions by Cruise LLC:\nCopyright (c) 2022 Cruise LLC.\nAll rights reserved.\n\nAll contributions by Arm:\nCopyright (c) 2021, 2023-2024 Arm Limited and/or its affiliates\n\nAll contributions from Caffe:\nCopyright(c) 2013, 2014, 2015, the respective contributors\nAll rights reserved.\n\nAll other contributions:\nCopyright(c) 2015, 2016 the respective contributors\nAll rights reserved.\n\nCaffe2 uses a copyright model similar to Caffe: each contributor holds\ncopyright over their contributions to Caffe2. The project versioning records\nall such contribution and copyright details. If a contributor wants to further\nmark their specific copyright on a particular contribution, they should\nindicate their copyright solely in the commit message of the change when it is\ncommitted.\n\nAll rights reserved.\n\nRedistribution and use in source and binary forms, with or without\nmodification, are permitted provided that the following conditions are met:\n\n1. Redistributions of source code must retain the above copyright\nnotice, this list of conditions and the following disclaimer.\n\n2. Redistributions in binary form must reproduce the above copyright\nnotice, this list of conditions and the following disclaimer in the\ndocumentation and/or other materials provided with the distribution.\n\n3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories America\nand IDIAP Research Institute nor the names of its contributors may be\nused to endorse or promote products derived from this software without\nspecific prior written permission.\n\nTHIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\"\nAND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE\nIMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE\nARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE\nLIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR\nCONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF\nSUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS\nINTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN\nCONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)\nARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE\nPOSSIBILITY OF SUCH DAMAGE.\nTriton\n/*\n* Copyright 2018-2020 Philippe Tillet\n* Copyright 2020-2022 OpenAI\n*\n* Permission is hereby granted, free of charge, to any person obtaining\n* a copy of this software and associated documentation files\n* (the \"Software\"), to deal in the Software without restriction,\n* including without limitation the rights to use, copy, modify, merge,\n* publish, distribute, sublicense, and/or sell copies of the Software,\n* and to permit persons to whom the Software is furnished to do so,\n* subject to the following conditions:\n*\n* The above copyright notice and this permission notice shall be\n* included in all copies or substantial portions of the Software.\n*\n* THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND,\n* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF\n* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.\n* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY\n* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,\n* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE\n* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.\n*/\nTransformers\nCopyright 2018- The Hugging Face team. All rights reserved.\n\n                                Apache License\n                        Version 2.0, January 2004\n                        http://www.apache.org/licenses/\n\nTERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION\n\n1. Definitions.\n\n    \"License\" shall mean the terms and conditions for use, reproduction,\n    and distribution as defined by Sections 1 through 9 of this document.\n\n    \"Licensor\" shall mean the copyright owner or entity authorized by\n    the copyright owner that is granting the License.\n\n    \"Legal Entity\" shall mean the union of the acting entity and all\n    other entities that control, are controlled by, or are under common\n    control with that entity. For the purposes of this definition,\n    \"control\" means (i) the power, direct or indirect, to cause the\n    direction or management of such entity, whether by contract or\n    otherwise, or (ii) ownership of fifty percent (50%) or more of the\n    outstanding shares, or (iii) beneficial ownership of such entity.\n\n    \"You\" (or \"Your\") shall mean an individual or Legal Entity\n    exercising permissions granted by this License.\n\n    \"Source\" form shall mean the preferred form for making modifications,\n    including but not limited to software source code, documentation\n    source, and configuration files.\n\n    \"Object\" form shall mean any form resulting from mechanical\n    transformation or translation of a Source form, including but\n    not limited to compiled object code, generated documentation,\n    and conversions to other media types.\n\n    \"Work\" shall mean the work of authorship, whether in Source or\n    Object form, made available under the License, as indicated by a\n    copyright notice that is included in or attached to the work\n    (an example is provided in the Appendix below).\n\n    \"Derivative Works\" shall mean any work, whether in Source or Object\n    form, that is based on (or derived from) the Work and for which the\n    editorial revisions, annotations, elaborations, or other modifications\n    represent, as a whole, an original work of authorship. For the purposes\n    of this License, Derivative Works shall not include works that remain\n    separable from, or merely link (or bind by name) to the interfaces of,\n    the Work and Derivative Works thereof.\n\n    \"Contribution\" shall mean any work of authorship, including\n    the original version of the Work and any modifications or additions\n    to that Work or Derivative Works thereof, that is intentionally\n    submitted to Licensor for inclusion in the Work by the copyright owner\n    or by an individual or Legal Entity authorized to submit on behalf of\n    the copyright owner. For the purposes of this definition, \"submitted\"\n    means any form of electronic, verbal, or written communication sent\n    to the Licensor or its representatives, including but not limited to\n    communication on electronic mailing lists, source code control systems,\n    and issue tracking systems that are managed by, or on behalf of, the\n    Licensor for the purpose of discussing and improving the Work, but\n    excluding communication that is conspicuously marked or otherwise\n    designated in writing by the copyright owner as \"Not a Contribution.\"\n\n    \"Contributor\" shall mean Licensor and any individual or Legal Entity\n    on behalf of whom a Contribution has been received by Licensor and\n    subsequently incorporated within the Work.\n\n2. Grant of Copyright License. Subject to the terms and conditions of\n    this License, each Contributor hereby grants to You a perpetual,\n    worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n    copyright license to reproduce, prepare Derivative Works of,\n    publicly display, publicly perform, sublicense, and distribute the\n    Work and such Derivative Works in Source or Object form.\n\n3. Grant of Patent License. Subject to the terms and conditions of\n    this License, each Contributor hereby grants to You a perpetual,\n    worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n    (except as stated in this section) patent license to make, have made,\n    use, offer to sell, sell, import, and otherwise transfer the Work,\n    where such license applies only to those patent claims licensable\n    by such Contributor that are necessarily infringed by their\n    Contribution(s) alone or by combination of their Contribution(s)\n    with the Work to which such Contribution(s) was submitted. If You\n    institute patent litigation against any entity (including a\n    cross-claim or counterclaim in a lawsuit) alleging that the Work\n    or a Contribution incorporated within the Work constitutes direct\n    or contributory patent infringement, then any patent licenses\n    granted to You under this License for that Work shall terminate\n    as of the date such litigation is filed.\n\n4. Redistribution. You may reproduce and distribute copies of the\n    Work or Derivative Works thereof in any medium, with or without\n    modifications, and in Source or Object form, provided that You\n    meet the following conditions:\n\n    (a) You must give any other recipients of the Work or\n        Derivative Works a copy of this License; and\n\n    (b) You must cause any modified files to carry prominent notices\n        stating that You changed the files; and\n\n    (c) You must retain, in the Source form of any Derivative Works\n        that You distribute, all copyright, patent, trademark, and\n        attribution notices from the Source form of the Work,\n        excluding those notices that do not pertain to any part of\n        the Derivative Works; and\n\n    (d) If the Work includes a \"NOTICE\" text file as part of its\n        distribution, then any Derivative Works that You distribute must\n        include a readable copy of the attribution notices contained\n        within such NOTICE file, excluding those notices that do not\n        pertain to any part of the Derivative Works, in at least one\n        of the following places: within a NOTICE text file distributed\n        as part of the Derivative Works; within the Source form or\n        documentation, if provided along with the Derivative Works; or,\n        within a display generated by the Derivative Works, if and\n        wherever such third-party notices normally appear. The contents\n        of the NOTICE file are for informational purposes only and\n        do not modify the License. You may add Your own attribution\n        notices within Derivative Works that You distribute, alongside\n        or as an addendum to the NOTICE text from the Work, provided\n        that such additional attribution notices cannot be construed\n        as modifying the License.\n\n    You may add Your own copyright statement to Your modifications and\n    may provide additional or different license terms and conditions\n    for use, reproduction, or distribution of Your modifications, or\n    for any such Derivative Works as a whole, provided Your use,\n    reproduction, and distribution of the Work otherwise complies with\n    the conditions stated in this License.\n\n5. Submission of Contributions. Unless You explicitly state otherwise,\n    any Contribution intentionally submitted for inclusion in the Work\n    by You to the Licensor shall be under the terms and conditions of\n    this License, without any additional terms or conditions.\n    Notwithstanding the above, nothing herein shall supersede or modify\n    the terms of any separate license agreement you may have executed\n    with Licensor regarding such Contributions.\n\n6. Trademarks. This License does not grant permission to use the trade\n    names, trademarks, service marks, or product names of the Licensor,\n    except as required for reasonable and customary use in describing the\n    origin of the Work and reproducing the content of the NOTICE file.\n\n7. Disclaimer of Warranty. Unless required by applicable law or\n    agreed to in writing, Licensor provides the Work (and each\n    Contributor provides its Contributions) on an \"AS IS\" BASIS,\n    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or\n    implied, including, without limitation, any warranties or conditions\n    of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A\n    PARTICULAR PURPOSE. You are solely responsible for determining the\n    appropriateness of using or redistributing the Work and assume any\n    risks associated with Your exercise of permissions under this License.\n\n8. Limitation of Liability. In no event and under no legal theory,\n    whether in tort (including negligence), contract, or otherwise,\n    unless required by applicable law (such as deliberate and grossly\n    negligent acts) or agreed to in writing, shall any Contributor be\n    liable to You for damages, including any direct, indirect, special,\n    incidental, or consequential damages of any character arising as a\n    result of this License or out of the use or inability to use the\n    Work (including but not limited to damages for loss of goodwill,\n    work stoppage, computer failure or malfunction, or any and all\n    other commercial damages or losses), even if such Contributor\n    has been advised of the possibility of such damages.\n\n9. Accepting Warranty or Additional Liability. While redistributing\n    the Work or Derivative Works thereof, You may choose to offer,\n    and charge a fee for, acceptance of support, warranty, indemnity,\n    or other liability obligations and/or rights consistent with this\n    License. However, in accepting such obligations, You may act only\n    on Your own behalf and on Your sole responsibility, not on behalf\n    of any other Contributor, and only if You agree to indemnify,\n    defend, and hold each Contributor harmless for any liability\n    incurred by, or claims asserted against, such Contributor by reason\n    of your accepting any such warranty or additional liability.\n\nEND OF TERMS AND CONDITIONS\n\nAPPENDIX: How to apply the Apache License to your work.\n\n    To apply the Apache License to your work, attach the following\n    boilerplate notice, with the fields enclosed by brackets \"[]\"\n    replaced with your own identifying information. (Don't include\n    the brackets!)  The text should be enclosed in the appropriate\n    comment syntax for the file format. We also recommend that a\n    file or class name and description of purpose be included on the\n    same \"printed page\" as the copyright notice for easier\n    identification within third-party archives.\n\nCopyright [yyyy] [name of copyright owner]\n\nLicensed under the Apache License, Version 2.0 (the \"License\");\nyou may not use this file except in compliance with the License.\nYou may obtain a copy of the License at\n\n    http://www.apache.org/licenses/LICENSE-2.0\n\nUnless required by applicable law or agreed to in writing, software\ndistributed under the License is distributed on an \"AS IS\" BASIS,\nWITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\nSee the License for the specific language governing permissions and\nlimitations under the License."
+    "text": "AXOLOTL COMMUNITY LICENSE AGREEMENT\nThis Axolotl Community License Agreement (“Agreement”) is entered into by and between Axolotl AI Corp. (“Axolotl”) and any individual or entity (“Licensee”) who wishes to use the Software (as defined below) in accordance with the terms and conditions set forth in this Agreement.\n\nDefinitions 1.1 “Licensee” refers to any individual or entity who has obtained a copy of the Software under this Agreement. 1.2 “Plugin Integration” means independent integration software modules which may or may not be offered by Axolotl, which may be licensed separately by their respective authors and/or licensors. 1.3 “Software” refers to the specific sub-directory of the Axolotl, Inc. software located at https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations and its subdirectories which permits Plugin Integrations to integrate with the Axolotl service.\nGrant of License 2.1 Axolotl hereby grants Licensee a worldwide, non-exclusive, royalty-free, license to use, copy, modify, merge, publish, distribute, sublicense, and/or otherwise exploit the Software, subject to the following conditions: - Licensee must comply with all the terms and conditions of this Agreement. - Licensee must include the original copyright notice and disclaimer of warranty in all copies or substantial portions of the Software. 2.2 Licensee may use the Software for any lawful purpose, except as restricted in Section 3.\nRestrictions 3.1 Licensee shall not use the Software for any activity that constitutes a commercial activity of offering for free or for sale any services, platform, or equivalent to third parties for the purposes of allowing such third parties to fine-tune artificial intelligence models. 3.2 Licensee shall not: - Use the Software for any illegal or unauthorized purpose. - Reverse engineer, decompile, or disassemble the Software. - Remove or modify any copyright, trademark, or other proprietary notices contained in the Software. - Use the Software in a way that could damage, disable, overburden, or impair the functionality of the Software or interfere with any third-party use of the Software. 3.3 Axolotl reserves the right to restrict certain Plugin Integrations for use with the Software. To the extent Licensee integrates a permitted, applicable Plugin Integration with the Software, Licensee shall comply with any additional terms and conditions imposed by the licensors of such Plugin Integration for use of such Plugin Integrations. Licensee shall contact Axolotl if it has questions about whether its use of the Software falls beyond the scope of this Agreement.\nIntellectual Property Rights 4.1 Axolotl and its contributors retain all intellectual property rights in and to the Software. Licensee acknowledges that this Agreement does not transfer any ownership rights or intellectual property rights to Licensee.\nDisclaimer of Warranty 5.1 THE SOFTWARE IS PROVIDED “AS IS,” WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, AND NON-INFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT, OR OTHERWISE, ARISING FROM, OUT OF, OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.\nTermination 6.1 Axolotl may terminate this Agreement at any time if Licensee fails to comply with any of the terms and conditions set forth herein. Upon termination, Licensee shall cease all use of the Software and destroy any copies in its possession.\nGoverning Law 7.1 This Agreement shall be governed by and construed in accordance with the laws of the State of California, without regards to conflicts of laws provisions thereof.\nEntire Agreement 8.1 This Agreement constitutes the entire agreement between Axolotl and Licensee with respect to the subject matter hereof and supersedes all prior or contemporaneous understandings or agreements between the parties concerning the Software, whether written or oral. Axolotl may update the terms of this Agreement from time to time, and Licensee’s continued use of the Software after any such updates shall constitute acceptance of updated terms on a go-forward basis. Axolotl will use commercially reasonable efforts to provide Licensee notice of any material updates. By using the Software, Licensee acknowledges that it has read, understood, and agrees to be bound by the terms and conditions of this Agreement.\n\nThis Agreement was last updated on August 23, 2024."
   },
   {
     "objectID": "index.html",
@@ -648,482 +717,468 @@
     ]
   },
   {
-    "objectID": "src/axolotl/integrations/LICENSE.html",
-    "href": "src/axolotl/integrations/LICENSE.html",
+    "objectID": "src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html",
+    "href": "src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html",
     "title": "Axolotl",
     "section": "",
-    "text": "AXOLOTL COMMUNITY LICENSE AGREEMENT\nThis Axolotl Community License Agreement (“Agreement”) is entered into by and between Axolotl AI Corp. (“Axolotl”) and any individual or entity (“Licensee”) who wishes to use the Software (as defined below) in accordance with the terms and conditions set forth in this Agreement.\n\nDefinitions 1.1 “Licensee” refers to any individual or entity who has obtained a copy of the Software under this Agreement. 1.2 “Plugin Integration” means independent integration software modules which may or may not be offered by Axolotl, which may be licensed separately by their respective authors and/or licensors. 1.3 “Software” refers to the specific sub-directory of the Axolotl, Inc. software located at https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations and its subdirectories which permits Plugin Integrations to integrate with the Axolotl service.\nGrant of License 2.1 Axolotl hereby grants Licensee a worldwide, non-exclusive, royalty-free, license to use, copy, modify, merge, publish, distribute, sublicense, and/or otherwise exploit the Software, subject to the following conditions: - Licensee must comply with all the terms and conditions of this Agreement. - Licensee must include the original copyright notice and disclaimer of warranty in all copies or substantial portions of the Software. 2.2 Licensee may use the Software for any lawful purpose, except as restricted in Section 3.\nRestrictions 3.1 Licensee shall not use the Software for any activity that constitutes a commercial activity of offering for free or for sale any services, platform, or equivalent to third parties for the purposes of allowing such third parties to fine-tune artificial intelligence models. 3.2 Licensee shall not: - Use the Software for any illegal or unauthorized purpose. - Reverse engineer, decompile, or disassemble the Software. - Remove or modify any copyright, trademark, or other proprietary notices contained in the Software. - Use the Software in a way that could damage, disable, overburden, or impair the functionality of the Software or interfere with any third-party use of the Software. 3.3 Axolotl reserves the right to restrict certain Plugin Integrations for use with the Software. To the extent Licensee integrates a permitted, applicable Plugin Integration with the Software, Licensee shall comply with any additional terms and conditions imposed by the licensors of such Plugin Integration for use of such Plugin Integrations. Licensee shall contact Axolotl if it has questions about whether its use of the Software falls beyond the scope of this Agreement.\nIntellectual Property Rights 4.1 Axolotl and its contributors retain all intellectual property rights in and to the Software. Licensee acknowledges that this Agreement does not transfer any ownership rights or intellectual property rights to Licensee.\nDisclaimer of Warranty 5.1 THE SOFTWARE IS PROVIDED “AS IS,” WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, AND NON-INFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT, OR OTHERWISE, ARISING FROM, OUT OF, OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.\nTermination 6.1 Axolotl may terminate this Agreement at any time if Licensee fails to comply with any of the terms and conditions set forth herein. Upon termination, Licensee shall cease all use of the Software and destroy any copies in its possession.\nGoverning Law 7.1 This Agreement shall be governed by and construed in accordance with the laws of the State of California, without regards to conflicts of laws provisions thereof.\nEntire Agreement 8.1 This Agreement constitutes the entire agreement between Axolotl and Licensee with respect to the subject matter hereof and supersedes all prior or contemporaneous understandings or agreements between the parties concerning the Software, whether written or oral. Axolotl may update the terms of this Agreement from time to time, and Licensee’s continued use of the Software after any such updates shall constitute acceptance of updated terms on a go-forward basis. Axolotl will use commercially reasonable efforts to provide Licensee notice of any material updates. By using the Software, Licensee acknowledges that it has read, understood, and agrees to be bound by the terms and conditions of this Agreement.\n\nThis Agreement was last updated on August 23, 2024."
+    "text": "Acknowledgements\nPortions of this Cut Cross Entropy Software may utilize the following copyrighted material, the use of which is hereby acknowledged.\n\nPyTorch\nFrom PyTorch:\n\nCopyright (c) 2016-     Facebook, Inc            (Adam Paszke)\nCopyright (c) 2014-     Facebook, Inc            (Soumith Chintala)\nCopyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)\nCopyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)\nCopyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)\nCopyright (c) 2011-2013 NYU                      (Clement Farabet)\nCopyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)\nCopyright (c) 2006      Idiap Research Institute (Samy Bengio)\nCopyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)\n\nFrom Caffe2:\n\nCopyright (c) 2016-present, Facebook Inc. All rights reserved.\n\nAll contributions by Facebook:\nCopyright (c) 2016 Facebook Inc.\n\nAll contributions by Google:\nCopyright (c) 2015 Google Inc.\nAll rights reserved.\n\nAll contributions by Yangqing Jia:\nCopyright (c) 2015 Yangqing Jia\nAll rights reserved.\n\nAll contributions by Kakao Brain:\nCopyright 2019-2020 Kakao Brain\n\nAll contributions by Cruise LLC:\nCopyright (c) 2022 Cruise LLC.\nAll rights reserved.\n\nAll contributions by Arm:\nCopyright (c) 2021, 2023-2024 Arm Limited and/or its affiliates\n\nAll contributions from Caffe:\nCopyright(c) 2013, 2014, 2015, the respective contributors\nAll rights reserved.\n\nAll other contributions:\nCopyright(c) 2015, 2016 the respective contributors\nAll rights reserved.\n\nCaffe2 uses a copyright model similar to Caffe: each contributor holds\ncopyright over their contributions to Caffe2. The project versioning records\nall such contribution and copyright details. If a contributor wants to further\nmark their specific copyright on a particular contribution, they should\nindicate their copyright solely in the commit message of the change when it is\ncommitted.\n\nAll rights reserved.\n\nRedistribution and use in source and binary forms, with or without\nmodification, are permitted provided that the following conditions are met:\n\n1. Redistributions of source code must retain the above copyright\nnotice, this list of conditions and the following disclaimer.\n\n2. Redistributions in binary form must reproduce the above copyright\nnotice, this list of conditions and the following disclaimer in the\ndocumentation and/or other materials provided with the distribution.\n\n3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories America\nand IDIAP Research Institute nor the names of its contributors may be\nused to endorse or promote products derived from this software without\nspecific prior written permission.\n\nTHIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\"\nAND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE\nIMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE\nARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE\nLIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR\nCONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF\nSUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS\nINTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN\nCONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)\nARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE\nPOSSIBILITY OF SUCH DAMAGE.\nTriton\n/*\n* Copyright 2018-2020 Philippe Tillet\n* Copyright 2020-2022 OpenAI\n*\n* Permission is hereby granted, free of charge, to any person obtaining\n* a copy of this software and associated documentation files\n* (the \"Software\"), to deal in the Software without restriction,\n* including without limitation the rights to use, copy, modify, merge,\n* publish, distribute, sublicense, and/or sell copies of the Software,\n* and to permit persons to whom the Software is furnished to do so,\n* subject to the following conditions:\n*\n* The above copyright notice and this permission notice shall be\n* included in all copies or substantial portions of the Software.\n*\n* THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND,\n* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF\n* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.\n* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY\n* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,\n* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE\n* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.\n*/\nTransformers\nCopyright 2018- The Hugging Face team. All rights reserved.\n\n                                Apache License\n                        Version 2.0, January 2004\n                        http://www.apache.org/licenses/\n\nTERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION\n\n1. Definitions.\n\n    \"License\" shall mean the terms and conditions for use, reproduction,\n    and distribution as defined by Sections 1 through 9 of this document.\n\n    \"Licensor\" shall mean the copyright owner or entity authorized by\n    the copyright owner that is granting the License.\n\n    \"Legal Entity\" shall mean the union of the acting entity and all\n    other entities that control, are controlled by, or are under common\n    control with that entity. For the purposes of this definition,\n    \"control\" means (i) the power, direct or indirect, to cause the\n    direction or management of such entity, whether by contract or\n    otherwise, or (ii) ownership of fifty percent (50%) or more of the\n    outstanding shares, or (iii) beneficial ownership of such entity.\n\n    \"You\" (or \"Your\") shall mean an individual or Legal Entity\n    exercising permissions granted by this License.\n\n    \"Source\" form shall mean the preferred form for making modifications,\n    including but not limited to software source code, documentation\n    source, and configuration files.\n\n    \"Object\" form shall mean any form resulting from mechanical\n    transformation or translation of a Source form, including but\n    not limited to compiled object code, generated documentation,\n    and conversions to other media types.\n\n    \"Work\" shall mean the work of authorship, whether in Source or\n    Object form, made available under the License, as indicated by a\n    copyright notice that is included in or attached to the work\n    (an example is provided in the Appendix below).\n\n    \"Derivative Works\" shall mean any work, whether in Source or Object\n    form, that is based on (or derived from) the Work and for which the\n    editorial revisions, annotations, elaborations, or other modifications\n    represent, as a whole, an original work of authorship. For the purposes\n    of this License, Derivative Works shall not include works that remain\n    separable from, or merely link (or bind by name) to the interfaces of,\n    the Work and Derivative Works thereof.\n\n    \"Contribution\" shall mean any work of authorship, including\n    the original version of the Work and any modifications or additions\n    to that Work or Derivative Works thereof, that is intentionally\n    submitted to Licensor for inclusion in the Work by the copyright owner\n    or by an individual or Legal Entity authorized to submit on behalf of\n    the copyright owner. For the purposes of this definition, \"submitted\"\n    means any form of electronic, verbal, or written communication sent\n    to the Licensor or its representatives, including but not limited to\n    communication on electronic mailing lists, source code control systems,\n    and issue tracking systems that are managed by, or on behalf of, the\n    Licensor for the purpose of discussing and improving the Work, but\n    excluding communication that is conspicuously marked or otherwise\n    designated in writing by the copyright owner as \"Not a Contribution.\"\n\n    \"Contributor\" shall mean Licensor and any individual or Legal Entity\n    on behalf of whom a Contribution has been received by Licensor and\n    subsequently incorporated within the Work.\n\n2. Grant of Copyright License. Subject to the terms and conditions of\n    this License, each Contributor hereby grants to You a perpetual,\n    worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n    copyright license to reproduce, prepare Derivative Works of,\n    publicly display, publicly perform, sublicense, and distribute the\n    Work and such Derivative Works in Source or Object form.\n\n3. Grant of Patent License. Subject to the terms and conditions of\n    this License, each Contributor hereby grants to You a perpetual,\n    worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n    (except as stated in this section) patent license to make, have made,\n    use, offer to sell, sell, import, and otherwise transfer the Work,\n    where such license applies only to those patent claims licensable\n    by such Contributor that are necessarily infringed by their\n    Contribution(s) alone or by combination of their Contribution(s)\n    with the Work to which such Contribution(s) was submitted. If You\n    institute patent litigation against any entity (including a\n    cross-claim or counterclaim in a lawsuit) alleging that the Work\n    or a Contribution incorporated within the Work constitutes direct\n    or contributory patent infringement, then any patent licenses\n    granted to You under this License for that Work shall terminate\n    as of the date such litigation is filed.\n\n4. Redistribution. You may reproduce and distribute copies of the\n    Work or Derivative Works thereof in any medium, with or without\n    modifications, and in Source or Object form, provided that You\n    meet the following conditions:\n\n    (a) You must give any other recipients of the Work or\n        Derivative Works a copy of this License; and\n\n    (b) You must cause any modified files to carry prominent notices\n        stating that You changed the files; and\n\n    (c) You must retain, in the Source form of any Derivative Works\n        that You distribute, all copyright, patent, trademark, and\n        attribution notices from the Source form of the Work,\n        excluding those notices that do not pertain to any part of\n        the Derivative Works; and\n\n    (d) If the Work includes a \"NOTICE\" text file as part of its\n        distribution, then any Derivative Works that You distribute must\n        include a readable copy of the attribution notices contained\n        within such NOTICE file, excluding those notices that do not\n        pertain to any part of the Derivative Works, in at least one\n        of the following places: within a NOTICE text file distributed\n        as part of the Derivative Works; within the Source form or\n        documentation, if provided along with the Derivative Works; or,\n        within a display generated by the Derivative Works, if and\n        wherever such third-party notices normally appear. The contents\n        of the NOTICE file are for informational purposes only and\n        do not modify the License. You may add Your own attribution\n        notices within Derivative Works that You distribute, alongside\n        or as an addendum to the NOTICE text from the Work, provided\n        that such additional attribution notices cannot be construed\n        as modifying the License.\n\n    You may add Your own copyright statement to Your modifications and\n    may provide additional or different license terms and conditions\n    for use, reproduction, or distribution of Your modifications, or\n    for any such Derivative Works as a whole, provided Your use,\n    reproduction, and distribution of the Work otherwise complies with\n    the conditions stated in this License.\n\n5. Submission of Contributions. Unless You explicitly state otherwise,\n    any Contribution intentionally submitted for inclusion in the Work\n    by You to the Licensor shall be under the terms and conditions of\n    this License, without any additional terms or conditions.\n    Notwithstanding the above, nothing herein shall supersede or modify\n    the terms of any separate license agreement you may have executed\n    with Licensor regarding such Contributions.\n\n6. Trademarks. This License does not grant permission to use the trade\n    names, trademarks, service marks, or product names of the Licensor,\n    except as required for reasonable and customary use in describing the\n    origin of the Work and reproducing the content of the NOTICE file.\n\n7. Disclaimer of Warranty. Unless required by applicable law or\n    agreed to in writing, Licensor provides the Work (and each\n    Contributor provides its Contributions) on an \"AS IS\" BASIS,\n    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or\n    implied, including, without limitation, any warranties or conditions\n    of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A\n    PARTICULAR PURPOSE. You are solely responsible for determining the\n    appropriateness of using or redistributing the Work and assume any\n    risks associated with Your exercise of permissions under this License.\n\n8. Limitation of Liability. In no event and under no legal theory,\n    whether in tort (including negligence), contract, or otherwise,\n    unless required by applicable law (such as deliberate and grossly\n    negligent acts) or agreed to in writing, shall any Contributor be\n    liable to You for damages, including any direct, indirect, special,\n    incidental, or consequential damages of any character arising as a\n    result of this License or out of the use or inability to use the\n    Work (including but not limited to damages for loss of goodwill,\n    work stoppage, computer failure or malfunction, or any and all\n    other commercial damages or losses), even if such Contributor\n    has been advised of the possibility of such damages.\n\n9. Accepting Warranty or Additional Liability. While redistributing\n    the Work or Derivative Works thereof, You may choose to offer,\n    and charge a fee for, acceptance of support, warranty, indemnity,\n    or other liability obligations and/or rights consistent with this\n    License. However, in accepting such obligations, You may act only\n    on Your own behalf and on Your sole responsibility, not on behalf\n    of any other Contributor, and only if You agree to indemnify,\n    defend, and hold each Contributor harmless for any liability\n    incurred by, or claims asserted against, such Contributor by reason\n    of your accepting any such warranty or additional liability.\n\nEND OF TERMS AND CONDITIONS\n\nAPPENDIX: How to apply the Apache License to your work.\n\n    To apply the Apache License to your work, attach the following\n    boilerplate notice, with the fields enclosed by brackets \"[]\"\n    replaced with your own identifying information. (Don't include\n    the brackets!)  The text should be enclosed in the appropriate\n    comment syntax for the file format. We also recommend that a\n    file or class name and description of purpose be included on the\n    same \"printed page\" as the copyright notice for easier\n    identification within third-party archives.\n\nCopyright [yyyy] [name of copyright owner]\n\nLicensed under the Apache License, Version 2.0 (the \"License\");\nyou may not use this file except in compliance with the License.\nYou may obtain a copy of the License at\n\n    http://www.apache.org/licenses/LICENSE-2.0\n\nUnless required by applicable law or agreed to in writing, software\ndistributed under the License is distributed on an \"AS IS\" BASIS,\nWITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\nSee the License for the specific language governing permissions and\nlimitations under the License."
   },
   {
-    "objectID": "TODO.html",
-    "href": "TODO.html",
-    "title": "todo list",
+    "objectID": "FAQS.html",
+    "href": "FAQS.html",
+    "title": "FAQs",
     "section": "",
-    "text": "[] Validation of parameters for combinations that won’t work\n\n\n\n\nFSDP offload and gradient_checkpointing - https://github.com/pytorch/pytorch/issues/82203\nadamw_bnb_8bit doesn’t play well with FSDP offload"
+    "text": "FAQs\n\nCan you train StableLM with this? Yes, but only with a single GPU atm. Multi GPU support is coming soon! Just waiting on this PR\nWill this work with Deepspeed? That’s still a WIP, but setting export ACCELERATE_USE_DEEPSPEED=true should work in some cases\nError invalid argument at line 359 in file /workspace/bitsandbytes/csrc/pythonInterface.c /arrow/cpp/src/arrow/filesystem/s3fs.cc:2598:  arrow::fs::FinalizeS3 was not called even though S3 was initialized. This could lead to a segmentation fault at exit. Try reinstalling bitsandbytes and transformers from source."
   },
   {
-    "objectID": "TODO.html#things-that-are-known-not-to-work",
-    "href": "TODO.html#things-that-are-known-not-to-work",
-    "title": "todo list",
+    "objectID": "docs/multi-node.html",
+    "href": "docs/multi-node.html",
+    "title": "Multi Node",
     "section": "",
-    "text": "FSDP offload and gradient_checkpointing - https://github.com/pytorch/pytorch/issues/82203\nadamw_bnb_8bit doesn’t play well with FSDP offload"
-  },
-  {
-    "objectID": "docs/lr_groups.html",
-    "href": "docs/lr_groups.html",
-    "title": "Learning Rate Groups",
-    "section": "",
-    "text": "Inspired by LoRA+, Axolotl allows practitioners to specify separate learning rates for each module or groups of modules in a model.",
+    "text": "The below are three ways to train multi-node in Axolotl.",
     "crumbs": [
-      "How To Guides",
-      "Learning Rate Groups"
+      "Deployments",
+      "Multi Node"
     ]
   },
   {
-    "objectID": "docs/lr_groups.html#background",
-    "href": "docs/lr_groups.html#background",
-    "title": "Learning Rate Groups",
-    "section": "",
-    "text": "Inspired by LoRA+, Axolotl allows practitioners to specify separate learning rates for each module or groups of modules in a model.",
+    "objectID": "docs/multi-node.html#accelerate",
+    "href": "docs/multi-node.html#accelerate",
+    "title": "Multi Node",
+    "section": "Accelerate",
+    "text": "Accelerate\nYou will need to create a configuration for accelerate, either by using accelerate config and follow the instructions or you can use one of the preset below:\n~/.cache/huggingface/accelerate/default_config.yaml\ncompute_environment: LOCAL_MACHINE\ndebug: false\ndistributed_type: FSDP\ndowncast_bf16: 'no'\nmachine_rank: 0 # Set to 0 for the main machine, increment by one for other machines\nmain_process_ip: 10.0.0.4 # Set to main machine's IP\nmain_process_port: 5000\nmain_training_function: main\nmixed_precision: bf16\nnum_machines: 2 # Change to the number of machines\nnum_processes: 4 # That's the total number of GPUs, (for example: if you have 2 machines with 4 GPU, put 8)\nrdzv_backend: static\nsame_network: true\ntpu_env: []\ntpu_use_cluster: false\ntpu_use_sudo: false\nuse_cpu: false\nConfigure your model to use FSDP in the Axolotl yaml. For example:\nfsdp:\n  - full_shard\n  - auto_wrap\nfsdp_config:\n  fsdp_offload_params: true\n  fsdp_state_dict_type: FULL_STATE_DICT\n  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer\nAll you have to do now is launch using accelerate as you would usually do on each machine and voila, the processes will start once you have launched accelerate on every machine.",
     "crumbs": [
-      "How To Guides",
-      "Learning Rate Groups"
+      "Deployments",
+      "Multi Node"
     ]
   },
   {
-    "objectID": "docs/lr_groups.html#example",
-    "href": "docs/lr_groups.html#example",
-    "title": "Learning Rate Groups",
-    "section": "Example",
-    "text": "Example\nlr_groups:\n  - name: o_proj\n    modules:\n      - self_attn.o_proj.weight\n    lr: 1e-6\n  - name: q_proj\n    modules:\n      - model.layers.2.self_attn.q_proj.weight\n    lr: 1e-5\n\nlearning_rate: 2e-5\nIn this example, we have a default learning rate of 2e-5 across the entire model, but we have a separate learning rate of 1e-6 for all the self attention o_proj modules across all layers, and a learning are of 1e-5 to the 3rd layer’s self attention q_proj module.",
+    "objectID": "docs/multi-node.html#raytrain",
+    "href": "docs/multi-node.html#raytrain",
+    "title": "Multi Node",
+    "section": "Raytrain",
+    "text": "Raytrain\nPlease see ray train doc here.",
     "crumbs": [
-      "How To Guides",
-      "Learning Rate Groups"
+      "Deployments",
+      "Multi Node"
     ]
   },
   {
-    "objectID": "docs/debugging.html",
-    "href": "docs/debugging.html",
-    "title": "Debugging",
+    "objectID": "docs/multi-node.html#torchrun",
+    "href": "docs/multi-node.html#torchrun",
+    "title": "Multi Node",
+    "section": "Torchrun",
+    "text": "Torchrun\nIf you are using Infiniband, we recommend torchrun to utilize the full bandwidth.\nSet the following env (change buffersize/socketname depending on your system):\nexport NCCL_IB_DISABLE=0\nexport NCCL_SOCKET_IFNAME=\"eth0,en,eth,em,bond\"\nexport NCCL_BUFFSIZE=2097152\nRun the following on each node:\ntorchrun --nnodes $num_nodes --nproc_per_node $gpu_per_node --rdzv_id $rdzv_id --rdzv_backend c10d --rdzv_endpoint \"$head_node_ip:$head_node_port\" -m axolotl.cli.train config.yaml\nPlease make sure to substitute the placeholder variables.\n\nnum_nodes: Number of nodes (containing GPUs)\ngpu_per_node: Number of gpus per node\nhead_node_ip: IP of the head node (make sure other machines can connect to this)\nhead_node_port: Port of the head node (make sure other machines can connect to this. Default 29400)\nrdzv_id: A unique job ID that is used by the job across nodes.\n\n\n\n\n\n\n\nNote\n\n\n\nYou need to call axolotl.cli.train instead of axolotl train as the latter calls accelerate under the hood\n\n\nMore info on the available configs can be found on the Pytorch docs here",
+    "crumbs": [
+      "Deployments",
+      "Multi Node"
+    ]
+  },
+  {
+    "objectID": "docs/faq.html",
+    "href": "docs/faq.html",
+    "title": "FAQ",
     "section": "",
-    "text": "This document provides some tips and tricks for debugging Axolotl. It also provides an example configuration for debugging with VSCode. A good debugging setup is essential to understanding how Axolotl code works behind the scenes.",
+    "text": "General\nQ: The trainer stopped and hasn’t progressed in several minutes.\n\nA: Usually an issue with the GPUs communicating with each other. See the NCCL doc\n\nQ: Exitcode -9\n\nA: This usually happens when you run out of system RAM.\n\nQ: Exitcode -7 while using deepspeed\n\nA: Try upgrading deepspeed w: pip install -U deepspeed\n\nQ: AttributeError: ‘DummyOptim’ object has no attribute ‘step’\nQ: ModuleNotFoundError: No module named ‘mpi4py’ using single GPU with deepspeed\n\nA: You may be using deepspeed with single gpu. Please remove the deepspeed: section in the yaml file or --deepspeed CLI flag.\n\nQ: The codes is stuck on saving preprocessed datasets.\n\nA: This is usually an issue with the GPU. This can be resolved through setting the os environment variable CUDA_VISIBLE_DEVICES=0. If you are on runpod, this is usually a pod issue. Starting a new pod should take care of it.\n\n\n\nChat templates\nQ: jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____\n\nA: This means that the property mapping for the stated attribute does not exist when building chat_template prompt. For example, if no attribute 'content', please check you have added the correct mapping for content under message_property_mappings.\n\nQ: Empty template generated for turn ___\n\nA: The content is empty for that turn.\n\nQ: Could not find content start/end boundary for turn __\n\nA: The specific turn’s start/end could not be detected. Please ensure you have set the eos_token following your chat_template. Otherwise, this could be a chat_template which doesn’t use proper boundaries for each turn (like system). On the rare occurrence, make sure your content is not [[dummy_message]]. Please let us know about this.\n\nQ: Content end boundary is before start boundary for turn ___\n\nA: This is an edge case which should not occur. Please create an Issue if this happens.\n\nQ: Content end boundary is the same as start boundary for turn ___. This is likely an empty turn.\n\nA: This is likely an empty turn.\n\nQ: The EOS/EOT token is incorrectly being masked or not being masked.\n\nA: This is because of the mismatch between tokenizer.eos_token and EOS/EOT token in template. Please make sure to set eos_token under special_tokens to the same EOS/EOT token as in template.",
     "crumbs": [
       "Troubleshooting",
-      "Debugging"
+      "FAQ"
     ]
   },
   {
-    "objectID": "docs/debugging.html#table-of-contents",
-    "href": "docs/debugging.html#table-of-contents",
-    "title": "Debugging",
-    "section": "Table of Contents",
-    "text": "Table of Contents\n\nGeneral Tips\nDebugging with VSCode\n\nBackground\nConfiguration\nCustomizing your debugger\nVideo Tutorial\n\nDebugging With Docker\n\nSetup\nAttach To Container\nVideo - Attaching To Docker On Remote Host",
-    "crumbs": [
-      "Troubleshooting",
-      "Debugging"
-    ]
-  },
-  {
-    "objectID": "docs/debugging.html#general-tips",
-    "href": "docs/debugging.html#general-tips",
-    "title": "Debugging",
-    "section": "General Tips",
-    "text": "General Tips\nWhile debugging it’s helpful to simplify your test scenario as much as possible. Here are some tips for doing so:\n\n[!Important] All of these tips are incorporated into the example configuration for debugging with VSCode below.\n\n\nMake sure you are using the latest version of axolotl: This project changes often and bugs get fixed fast. Check your git branch and make sure you have pulled the latest changes from main.\nEliminate concurrency: Restrict the number of processes to 1 for both training and data preprocessing:\n\nSet CUDA_VISIBLE_DEVICES to a single GPU, ex: export CUDA_VISIBLE_DEVICES=0.\nSet dataset_processes: 1 in your axolotl config or run the training command with --dataset_processes=1.\n\nUse a small dataset: Construct or use a small dataset from HF Hub. When using a small dataset, you will often have to make sure sample_packing: False and eval_sample_packing: False to avoid errors. If you are in a pinch and don’t have time to construct a small dataset but want to use from the HF Hub, you can shard the data (this will still tokenize the entire dataset, but will only use a fraction of the data for training. For example, to shard the dataset into 20 pieces, add the following to your axolotl config):\ndatasets:\n    ...\n    shards: 20\nUse a small model: A good example of a small model is TinyLlama/TinyLlama-1.1B-Chat-v1.0.\nMinimize iteration time: Make sure the training loop finishes as fast as possible, with these settings.\n\nmicro_batch_size: 1\nmax_steps: 1\nval_set_size: 0\n\nClear Caches: Axolotl caches certain steps and so does the underlying HuggingFace trainer. You may want to clear some of these caches when debugging.\n\nData preprocessing: When debugging data preprocessing, which includes prompt template formation, you may want to delete the directory set in dataset_prepared_path: in your axolotl config. If you didn’t set this value, the default is last_run_prepared.\nHF Hub: If you are debugging data preprocessing, you should clear the relevant HF cache HuggingFace cache, by deleting the appropriate ~/.cache/huggingface/datasets/... folder(s).\nThe recommended approach is to redirect all outputs and caches to a temporary folder and delete selected subfolders before each run. This is demonstrated in the example configuration below.",
-    "crumbs": [
-      "Troubleshooting",
-      "Debugging"
-    ]
-  },
-  {
-    "objectID": "docs/debugging.html#debugging-with-vscode",
-    "href": "docs/debugging.html#debugging-with-vscode",
-    "title": "Debugging",
-    "section": "Debugging with VSCode",
-    "text": "Debugging with VSCode\n\nBackground\nThe below example shows how to configure VSCode to debug data preprocessing of the chat_template format. This is the format used when you have the following in your axolotl config:\ndatasets:\n  - path: &lt;path to your chat_template formatted dataset&gt; # example on HF Hub: fozziethebeat/alpaca_messages_2k_test\n    type: chat_template\n\n[!Important] If you are already familiar with advanced VSCode debugging, you can skip the below explanation and look at the files .vscode/launch.json and .vscode/tasks.json for an example configuration.\n\n\n[!Tip] If you prefer to watch a video, rather than read, you can skip to the video tutorial below (but doing both is recommended).\n\n\n\nSetup\nMake sure you have an editable install of Axolotl, which ensures that changes you make to the code are reflected at runtime. Run the following commands from the root of this project:\npip3 install packaging\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n\nRemote Hosts\nIf you developing on a remote host, you can easily use VSCode to debug remotely. To do so, you will need to follow this remote - SSH guide. You can also see the video below on Docker and Remote SSH debugging.\n\n\n\nConfiguration\nThe easiest way to get started is to modify the .vscode/launch.json file in this project. This is just an example configuration, so you may need to modify or copy it to suit your needs.\nFor example, to mimic the command cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_chat_template.yml, you would use the below configuration1. Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to devtools and set the env variable HF_HOME to a temporary folder that is later partially deleted. This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.\n// .vscode/launch.json\n{\n    \"version\": \"0.2.0\",\n    \"configurations\": [\n        {\n            \"name\": \"Debug axolotl prompt - chat_template\",\n            \"type\": \"python\",\n            \"module\": \"accelerate.commands.launch\",\n            \"request\": \"launch\",\n            \"args\": [\n                \"-m\", \"axolotl.cli.train\", \"dev_chat_template.yml\",\n                // The flags below simplify debugging by overriding the axolotl config\n                // with the debugging tips above.  Modify as needed.\n                \"--dataset_processes=1\",      // limits data preprocessing to one process\n                \"--max_steps=1\",              // limits training to just one step\n                \"--batch_size=1\",             // minimizes batch size\n                \"--micro_batch_size=1\",       // minimizes batch size\n                \"--val_set_size=0\",           // disables validation\n                \"--sample_packing=False\",     // disables sample packing which is necessary for small datasets\n                \"--eval_sample_packing=False\",// disables sample packing on eval set\n                \"--dataset_prepared_path=temp_debug/axolotl_outputs/data\", // send data outputs to a temp folder\n                \"--output_dir=temp_debug/axolotl_outputs/model\" // send model outputs to a temp folder\n                ],\n            \"console\": \"integratedTerminal\",      // show output in the integrated terminal\n            \"cwd\": \"${workspaceFolder}/devtools\", // set working directory to devtools from the root of the project\n            \"justMyCode\": true,                   // step through only axolotl code\n            \"env\": {\"CUDA_VISIBLE_DEVICES\": \"0\",  // Since we aren't doing distributed training, we need to limit to one GPU\n                    \"HF_HOME\": \"${workspaceFolder}/devtools/temp_debug/.hf-cache\"}, // send HF cache to a temp folder\n            \"preLaunchTask\": \"cleanup-for-dataprep\", // delete temp folders (see below)\n        }\n    ]\n}\nAdditional notes about this configuration:\n\nThe argument justMyCode is set to true such that you step through only the axolotl code. If you want to step into dependencies, set this to false.\nThe preLaunchTask: cleanup-for-dataprep is defined in .vscode/tasks.json and is used to delete the following folders before debugging, which is essential to ensure that the data pre-processing code is run from scratch:\n\n./devtools/temp_debug/axolotl_outputs\n./devtools/temp_debug/.hf-cache/datasets\n\n\n\n[!Tip] You may not want to delete these folders. For example, if you are debugging model training instead of data pre-processing, you may NOT want to delete the cache or output folders. You may also need to add additional tasks to the tasks.json file depending on your use case.\n\nBelow is the ./vscode/tasks.json file that defines the cleanup-for-dataprep task. This task is run before each debugging session when you use the above configuration. Note how there are two tasks that delete the two folders mentioned above. The third task cleanup-for-dataprep is a composite task that combines the two tasks. A composite task is necessary because VSCode does not allow you to specify multiple tasks in the preLaunchTask argument of the launch.json file.\n// .vscode/tasks.json\n// this file is used by launch.json\n{\n    \"version\": \"2.0.0\",\n    \"tasks\": [\n      // this task changes into the devtools directory and deletes the temp_debug/axolotl_outputs folder\n      {\n        \"label\": \"delete-outputs\",\n        \"type\": \"shell\",\n        \"command\": \"rm -rf temp_debug/axolotl_outputs\",\n        \"options\":{ \"cwd\": \"${workspaceFolder}/devtools\"},\n        \"problemMatcher\": []\n      },\n      // this task changes into the devtools directory and deletes the `temp_debug/.hf-cache/datasets` folder\n      {\n        \"label\": \"delete-temp-hf-dataset-cache\",\n        \"type\": \"shell\",\n        \"command\": \"rm -rf temp_debug/.hf-cache/datasets\",\n        \"options\":{ \"cwd\": \"${workspaceFolder}/devtools\"},\n        \"problemMatcher\": []\n      },\n        // this task combines the two tasks above\n      {\n       \"label\": \"cleanup-for-dataprep\",\n       \"dependsOn\": [\"delete-outputs\", \"delete-temp-hf-dataset-cache\"],\n      }\n    ]\n}\n\n\nCustomizing your debugger\nYour debugging use case may differ from the example above. The easiest thing to do is to put your own axolotl config in the devtools folder and modify the launch.json file to use your config. You may also want to modify the preLaunchTask to delete different folders or not delete anything at all.\n\n\nVideo Tutorial\nThe following video tutorial walks through the above configuration and demonstrates how to debug with VSCode, (click the image below to watch):\n\n\n\nHamel Husain’s tutorial: Debugging Axolotl w/VSCode",
-    "crumbs": [
-      "Troubleshooting",
-      "Debugging"
-    ]
-  },
-  {
-    "objectID": "docs/debugging.html#debugging-with-docker",
-    "href": "docs/debugging.html#debugging-with-docker",
-    "title": "Debugging",
-    "section": "Debugging With Docker",
-    "text": "Debugging With Docker\nUsing official Axolotl Docker images is a great way to debug your code, and is a very popular way to use Axolotl. Attaching VSCode to Docker takes a few more steps.\n\nSetup\nOn the host that is running axolotl (ex: if you are using a remote host), clone the axolotl repo and change your current directory to the root:\ngit clone https://github.com/axolotl-ai-cloud/axolotl\ncd axolotl\n\n[!Tip] If you already have axolotl cloned on your host, make sure you have the latest changes and change into the root of the project.\n\nNext, run the desired docker image and mount the current directory. Below is a docker command you can run to do this:2\ndocker run --privileged --gpus '\"all\"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src=\"${PWD}\",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface axolotlai/axolotl:main-py3.10-cu118-2.0.1\n\n[!Tip] To understand which containers are available, see the Docker section of the README and the DockerHub repo. For details of how the Docker containers are built, see axolotl’s Docker CI builds.\n\nYou will now be in the container. Next, perform an editable install of Axolotl:\npip3 install packaging\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n\n\nAttach To Container\nNext, if you are using a remote host, Remote into this host with VSCode. If you are using a local host, you can skip this step.\nNext, select Dev Containers: Attach to Running Container... using the command palette (CMD + SHIFT + P) in VSCode. You will be prompted to select a container to attach to. Select the container you just created. You will now be in the container with a working directory that is at the root of the project. Any changes you make to the code will be reflected both in the container and on the host.\nNow you are ready to debug as described above (see Debugging with VSCode).\n\n\nVideo - Attaching To Docker On Remote Host\nHere is a short video that demonstrates how to attach to a Docker container on a remote host:\n\n\n\nHamel Husain’s tutorial: Debugging Axolotl Part 2: Attaching to Docker on a Remote Host",
-    "crumbs": [
-      "Troubleshooting",
-      "Debugging"
-    ]
-  },
-  {
-    "objectID": "docs/debugging.html#footnotes",
-    "href": "docs/debugging.html#footnotes",
-    "title": "Debugging",
-    "section": "Footnotes",
-    "text": "Footnotes\n\n\nThe config actually mimics the command CUDA_VISIBLE_DEVICES=0 python -m accelerate.commands.launch -m axolotl.cli.train devtools/chat_template.yml, but this is the same thing.↩︎\nMany of the below flags are recommended best practices by Nvidia when using nvidia-container-toolkit. You can read more about these flags here.↩︎",
-    "crumbs": [
-      "Troubleshooting",
-      "Debugging"
-    ]
-  },
-  {
-    "objectID": "docs/multipack.html",
-    "href": "docs/multipack.html",
-    "title": "Multipack (Sample Packing)",
+    "objectID": "docs/batch_vs_grad.html",
+    "href": "docs/batch_vs_grad.html",
+    "title": "Batch size vs Gradient accumulation",
     "section": "",
-    "text": "Because Flash Attention simply drops the attention mask, we do not need to construct a 4d attention mask. We only need to concatenate the sequences into a single batch and let flash attention know where each new sequence begins.\n4k context, bsz =4, each character represents 256 tokens X represents a padding token\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B ]\n   C C C C C C C ]\n   D D D D ]]\n\n[[ E E E E E E E E ]\n [ F F F F ]\n [ G G G ]\n [ H H H H ]]\n\n[[ I I I ]\n [ J J J ]\n [ K K K K K]\n [ L L L ]]\nafter padding to longest input in each step\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B X X X X X X ]\n   C C C C C C C X X X X ]\n   D D D D X X X X X X X ]]\n\n[[ E E E E E E E E ]\n [ F F F F X X X X ]\n [ G G G X X X X X ]\n [ H H H H X X X X ]]\n\n[[ I I I X X ]\n [ J J J X X ]\n [ K K K K K ]\n [ L L L X X ]]\nw packing ( note it’s the same effective number of tokens per step, but a true bsz of 1)\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A B B B B B\n   B C C C C C C C D D D D E E E E\n   E E E E F F F F F G G G H H H H\n   I I I J J J J K K K K K L L L X ]]\ncu_seqlens: [[ 0, 11, 17, 24, 28, 36, 41 44, 48, 51, 55, 60, 64]]",
+    "text": "Gradient accumulation means accumulating gradients over several mini-batches and updating the model weights afterward. When the samples in each batch are diverse, this technique doesn’t significantly impact learning.\nThis method allows for effective training with larger effective batch sizes without needing proportionally larger memory. Here’s why:\n\nMemory Consumption with Batch Size: The primary reason increasing the batch size impacts memory is due to the storage requirements for intermediate activations. When you forward propagate a batch through a network, you have to store the activations at each layer for each sample in the batch, because these activations are used during backpropagation to compute gradients. Therefore, larger batches mean more activations, leading to greater GPU memory consumption.\nGradient Accumulation: With gradient accumulation, you’re effectively simulating a larger batch size by accumulating gradients over several smaller batches (or micro-batches). However, at any given time, you’re only forward and backward propagating a micro-batch. This means you only store activations for the micro-batch, not the full accumulated batch. As a result, you can simulate the effect of a larger batch size without the memory cost of storing activations for a large batch.\n\nExample 1: Micro batch size: 3 Gradient accumulation steps: 2 Number of GPUs: 3 Total batch size = 3 * 2 * 3 = 18\n| GPU 1          | GPU 2          | GPU 3          |\n|----------------|----------------|----------------|\n| S1, S2, S3     | S4, S5, S6     | S7, S8, S9     |\n| e1, e2, e3     | e4, e5, e6     | e7, e8, e9     |\n|----------------|----------------|----------------|\n| → (accumulate) | → (accumulate) | → (accumulate) |\n|----------------|----------------|----------------|\n| S10, S11, S12  | S13, S14, S15  | S16, S17, S18  |\n| e10, e11, e12  | e13, e14, e15  | e16, e17, e18  |\n|----------------|----------------|----------------|\n| → (apply)      | → (apply)      | → (apply)      |\n\nAccumulated gradient for the weight w1 after the second iteration (considering all GPUs):\nTotal gradient for w1 = e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8 + e9 + e10 + e11 + e12 + e13 + e14 + e15 + e16 + e17 + e18\n\nWeight update for w1:\nw1_new = w1_old - learning rate x (Total gradient for w1 / 18)\nExample 2: Micro batch size: 2 Gradient accumulation steps: 1 Number of GPUs: 3 Total batch size = 2 * 1 * 3 = 6\n| GPU 1     | GPU 2     | GPU 3     |\n|-----------|-----------|-----------|\n| S1, S2    | S3, S4    | S5, S6    |\n| e1, e2    | e3, e4    | e5, e6    |\n|-----------|-----------|-----------|\n| → (apply) | → (apply) | → (apply) |\n\nAccumulated gradient for the weight w1 (considering all GPUs):\nTotal gradient for w1 = e1 + e2 + e3 + e4 + e5 + e6\n\nWeight update for w1:\nw1_new = w1_old - learning rate × (Total gradient for w1 / 6)",
     "crumbs": [
       "Core Concepts",
-      "Multipack (Sample Packing)"
+      "Batch size vs Gradient accumulation"
     ]
   },
   {
-    "objectID": "docs/multipack.html#visualization-of-multipack-with-flash-attention",
-    "href": "docs/multipack.html#visualization-of-multipack-with-flash-attention",
-    "title": "Multipack (Sample Packing)",
+    "objectID": "docs/lora_optims.html",
+    "href": "docs/lora_optims.html",
+    "title": "LoRA Optimizations",
     "section": "",
-    "text": "Because Flash Attention simply drops the attention mask, we do not need to construct a 4d attention mask. We only need to concatenate the sequences into a single batch and let flash attention know where each new sequence begins.\n4k context, bsz =4, each character represents 256 tokens X represents a padding token\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B ]\n   C C C C C C C ]\n   D D D D ]]\n\n[[ E E E E E E E E ]\n [ F F F F ]\n [ G G G ]\n [ H H H H ]]\n\n[[ I I I ]\n [ J J J ]\n [ K K K K K]\n [ L L L ]]\nafter padding to longest input in each step\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B X X X X X X ]\n   C C C C C C C X X X X ]\n   D D D D X X X X X X X ]]\n\n[[ E E E E E E E E ]\n [ F F F F X X X X ]\n [ G G G X X X X X ]\n [ H H H H X X X X ]]\n\n[[ I I I X X ]\n [ J J J X X ]\n [ K K K K K ]\n [ L L L X X ]]\nw packing ( note it’s the same effective number of tokens per step, but a true bsz of 1)\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A B B B B B\n   B C C C C C C C D D D D E E E E\n   E E E E F F F F F G G G H H H H\n   I I I J J J J K K K K K L L L X ]]\ncu_seqlens: [[ 0, 11, 17, 24, 28, 36, 41 44, 48, 51, 55, 60, 64]]",
-    "crumbs": [
-      "Core Concepts",
-      "Multipack (Sample Packing)"
-    ]
-  },
-  {
-    "objectID": "docs/multipack.html#multipack-without-flash-attention",
-    "href": "docs/multipack.html#multipack-without-flash-attention",
-    "title": "Multipack (Sample Packing)",
-    "section": "Multipack without Flash Attention",
-    "text": "Multipack without Flash Attention\nMultipack can still be achieved without Flash attention, but with lower packing efficiency as we are not able to join multiple batches into a single batch due to context length limits without flash attention. We can use either Pytorch’s Scaled Dot Product Attention implementation or native Pytorch attention implementation along with 4d attention masks to pack sequences together and avoid cross attention.",
-    "crumbs": [
-      "Core Concepts",
-      "Multipack (Sample Packing)"
-    ]
-  },
-  {
-    "objectID": "docs/inference.html",
-    "href": "docs/inference.html",
-    "title": "Inference",
-    "section": "",
-    "text": "This guide covers how to use your trained models for inference, including model loading, interactive testing, and common troubleshooting steps.",
-    "crumbs": [
-      "Getting Started",
-      "Inference"
-    ]
-  },
-  {
-    "objectID": "docs/inference.html#sec-quickstart",
-    "href": "docs/inference.html#sec-quickstart",
-    "title": "Inference",
-    "section": "1 Quick Start",
-    "text": "1 Quick Start\n\n1.1 Basic Inference\n\nLoRA ModelsFull Fine-tuned Models\n\n\naxolotl inference your_config.yml --lora-model-dir=\"./lora-output-dir\"\n\n\naxolotl inference your_config.yml --base-model=\"./completed-model\"",
-    "crumbs": [
-      "Getting Started",
-      "Inference"
-    ]
-  },
-  {
-    "objectID": "docs/inference.html#sec-advanced",
-    "href": "docs/inference.html#sec-advanced",
-    "title": "Inference",
-    "section": "2 Advanced Usage",
-    "text": "2 Advanced Usage\n\n2.1 Gradio Interface\nLaunch an interactive web interface:\naxolotl inference your_config.yml --gradio\n\n\n2.2 File-based Prompts\nProcess prompts from a text file:\ncat /tmp/prompt.txt | axolotl inference your_config.yml \\\n  --base-model=\"./completed-model\" --prompter=None\n\n\n2.3 Memory Optimization\nFor large models or limited memory:\naxolotl inference your_config.yml --load-in-8bit=True",
-    "crumbs": [
-      "Getting Started",
-      "Inference"
-    ]
-  },
-  {
-    "objectID": "docs/inference.html#sec-merging",
-    "href": "docs/inference.html#sec-merging",
-    "title": "Inference",
-    "section": "3 Merging LoRA Weights",
-    "text": "3 Merging LoRA Weights\nMerge LoRA adapters with the base model:\naxolotl merge-lora your_config.yml --lora-model-dir=\"./completed-model\"\n\n3.1 Memory Management for Merging\n\nConfiguration OptionsForce CPU Merging\n\n\ngpu_memory_limit: 20GiB  # Adjust based on your GPU\nlora_on_cpu: true        # Process on CPU if needed\n\n\nCUDA_VISIBLE_DEVICES=\"\" axolotl merge-lora ...",
-    "crumbs": [
-      "Getting Started",
-      "Inference"
-    ]
-  },
-  {
-    "objectID": "docs/inference.html#sec-tokenization",
-    "href": "docs/inference.html#sec-tokenization",
-    "title": "Inference",
-    "section": "4 Tokenization",
-    "text": "4 Tokenization\n\n4.1 Common Issues\n\n\n\n\n\n\nWarning\n\n\n\nTokenization mismatches between training and inference are a common source of problems.\n\n\nTo debug:\n\nCheck training tokenization:\n\naxolotl preprocess your_config.yml --debug\n\nVerify inference tokenization by decoding tokens before model input\nCompare token IDs between training and inference\n\n\n\n4.2 Special Tokens\nConfigure special tokens in your YAML:\nspecial_tokens:\n  bos_token: \"&lt;s&gt;\"\n  eos_token: \"&lt;/s&gt;\"\n  unk_token: \"&lt;unk&gt;\"\ntokens:\n  - \"&lt;|im_start|&gt;\"\n  - \"&lt;|im_end|&gt;\"",
-    "crumbs": [
-      "Getting Started",
-      "Inference"
-    ]
-  },
-  {
-    "objectID": "docs/inference.html#sec-troubleshooting",
-    "href": "docs/inference.html#sec-troubleshooting",
-    "title": "Inference",
-    "section": "5 Troubleshooting",
-    "text": "5 Troubleshooting\n\n5.1 Common Problems\n\nMemory IssuesToken IssuesPerformance Issues\n\n\n\nUse 8-bit loading\nReduce batch sizes\nTry CPU offloading\n\n\n\n\nVerify special tokens\nCheck tokenizer settings\nCompare training and inference preprocessing\n\n\n\n\nVerify model loading\nCheck prompt formatting\nEnsure temperature/sampling settings\n\n\n\n\nFor more details, see our debugging guide.",
-    "crumbs": [
-      "Getting Started",
-      "Inference"
-    ]
-  },
-  {
-    "objectID": "docs/getting-started.html",
-    "href": "docs/getting-started.html",
-    "title": "Quickstart",
-    "section": "",
-    "text": "This guide will walk you through your first model fine-tuning project with Axolotl.",
-    "crumbs": [
-      "Getting Started",
-      "Quickstart"
-    ]
-  },
-  {
-    "objectID": "docs/getting-started.html#sec-quick-example",
-    "href": "docs/getting-started.html#sec-quick-example",
-    "title": "Quickstart",
-    "section": "1 Quick Example",
-    "text": "1 Quick Example\nLet’s start by fine-tuning a small language model using LoRA. This example uses a 1B parameter model to ensure it runs on most GPUs. Assuming axolotl is installed (if not, see our Installation Guide)\n\nDownload example configs:\n\naxolotl fetch examples\n\nRun the training:\n\naxolotl train examples/llama-3/lora-1b.yml\nThat’s it! Let’s understand what just happened.",
-    "crumbs": [
-      "Getting Started",
-      "Quickstart"
-    ]
-  },
-  {
-    "objectID": "docs/getting-started.html#sec-understanding",
-    "href": "docs/getting-started.html#sec-understanding",
-    "title": "Quickstart",
-    "section": "2 Understanding the Process",
-    "text": "2 Understanding the Process\n\n2.1 The Configuration File\nThe YAML configuration file controls everything about your training. Here’s what (part of) our example config looks like:\nbase_model: NousResearch/Llama-3.2-1B\n# hub_model_id: username/custom_model_name\n\ndatasets:\n  - path: teknium/GPT4-LLM-Cleaned\n    type: alpaca\ndataset_prepared_path: last_run_prepared\nval_set_size: 0.1\noutput_dir: ./outputs/lora-out\n\nadapter: lora\nlora_model_dir:\nSee our Config options for more details.\n\n\n2.2 Training\nWhen you run axolotl train, Axolotl:\n\nDownloads the base model\n(If specified) applies LoRA adapter layers\nLoads and processes the dataset\nRuns the training loop\nSaves the trained model and / or LoRA weights",
-    "crumbs": [
-      "Getting Started",
-      "Quickstart"
-    ]
-  },
-  {
-    "objectID": "docs/getting-started.html#sec-custom",
-    "href": "docs/getting-started.html#sec-custom",
-    "title": "Quickstart",
-    "section": "3 Your First Custom Training",
-    "text": "3 Your First Custom Training\nLet’s modify the example for your own data:\n\nCreate a new config file my_training.yml:\n\nbase_model: NousResearch/Nous-Hermes-llama-1b-v1\nadapter: lora\n\n# Training settings\nmicro_batch_size: 2\nnum_epochs: 3\nlearning_rate: 0.0003\n\n# Your dataset\ndatasets:\n  - path: my_data.jsonl        # Your local data file\n    type: alpaca               # Or other format\nThis specific config is for LoRA fine-tuning a model with instruction tuning data using the alpaca dataset format, which has the following format:\n{\n    \"instruction\": \"Write a description of alpacas.\",\n    \"input\": \"\",\n    \"output\": \"Alpacas are domesticated South American camelids...\"\n}\nPlease see our Dataset Formats for more dataset formats and how to format them.\n\nPrepare your JSONL data in the specified format (in this case, the expected `alpaca format):\n\n{\"instruction\": \"Classify this text\", \"input\": \"I love this!\", \"output\": \"positive\"}\n{\"instruction\": \"Classify this text\", \"input\": \"Not good at all\", \"output\": \"negative\"}\nPlease consult the supported Dataset Formats for more details.\n\nRun the training:\n\naxolotl train my_training.yml",
-    "crumbs": [
-      "Getting Started",
-      "Quickstart"
-    ]
-  },
-  {
-    "objectID": "docs/getting-started.html#sec-common-tasks",
-    "href": "docs/getting-started.html#sec-common-tasks",
-    "title": "Quickstart",
-    "section": "4 Common Tasks",
-    "text": "4 Common Tasks\n\n4.1 Testing Your Model\nAfter training, test your model:\naxolotl inference my_training.yml --lora-model-dir=\"./outputs/lora-out\"\n\n\n4.2 Preprocessing Data\nFor large datasets, preprocess first:\naxolotl preprocess my_training.yml\n\n\n4.3 Using a UI\nLaunch a Gradio interface:\naxolotl inference my_training.yml --lora-model-dir=\"./outputs/lora-out\" --gradio",
-    "crumbs": [
-      "Getting Started",
-      "Quickstart"
-    ]
-  },
-  {
-    "objectID": "docs/getting-started.html#sec-next-steps",
-    "href": "docs/getting-started.html#sec-next-steps",
-    "title": "Quickstart",
-    "section": "5 Next Steps",
-    "text": "5 Next Steps\nNow that you have the basics, you might want to:\n\nTry different model architectures\nExperiment with hyperparameters\nUse more advanced training methods\nScale up to larger models\n\nCheck our other guides for details on these topics:\n\nConfiguration Guide - Full configuration options\nDataset Formats - Working with different data formats\nMulti-GPU Training\nMulti-Node Training",
-    "crumbs": [
-      "Getting Started",
-      "Quickstart"
-    ]
-  },
-  {
-    "objectID": "docs/multimodal.html",
-    "href": "docs/multimodal.html",
-    "title": "MultiModal / Vision Language Models (BETA)",
-    "section": "",
-    "text": "MultiModal / Vision Language Models (BETA)\n\nSupported Models\n\nMllama, i.e. llama with vision models\n\n\n\nUsage\nCurrently multimodal support is limited and doesn’t have full feature parity. To finetune a multimodal Llama w/ LoRA, you’ll need to use the following in YAML in combination with the rest of the required hyperparams.\nbase_model: alpindale/Llama-3.2-11B-Vision-Instruct\nprocessor_type: AutoProcessor\nskip_prepare_dataset: true\n\nchat_template: llama3_2_vision\ndatasets:\n  - path: HuggingFaceH4/llava-instruct-mix-vsft\n    type: chat_template\n    split: train[:1%]\n    field_messages: messages\nremove_unused_columns: false\nsample_packing: false\n\n# only finetune the Language model, leave the vision model and vision tower frozen\nlora_target_modules: 'language_model.model.layers.[\\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'",
+    "text": "Inspired by Unsloth, we’ve implemented two optimizations for LoRA and QLoRA fine-tuning, supporting both single GPU and multi-GPU (in the DDP and DeepSpeed settings) training. These include (1) SwiGLU and GEGLU activation function Triton kernels, and (2) LoRA MLP and attention custom autograd functions. Our goal was to leverage operator fusion and tensor re-use in order to improve speed and reduce memory usage during the forward and backward passes of these calculations.\nWe currently support several common model architectures, including (but not limited to):",
     "crumbs": [
       "How To Guides",
-      "MultiModal / Vision Language Models (BETA)"
+      "LoRA Optimizations"
     ]
   },
   {
-    "objectID": "docs/input_output.html",
-    "href": "docs/input_output.html",
-    "title": "Template-free prompt construction",
-    "section": "",
-    "text": "The documentation moved to here."
-  },
-  {
-    "objectID": "docs/reward_modelling.html",
-    "href": "docs/reward_modelling.html",
-    "title": "Reward Modelling",
-    "section": "",
-    "text": "Overview\nReward modelling is a technique used to train models to predict the reward or value of a given input. This is particularly useful in reinforcement learning scenarios where the model needs to evaluate the quality of its actions or predictions. We support the reward modelling techniques supported by trl.\n\n\n(Outcome) Reward Models\nOutcome reward models are trained using data which contains preference annotations for an entire interaction between the user and model (e.g. rather than per-turn or per-step).\nbase_model: google/gemma-2-2b\nmodel_type: AutoModelForSequenceClassification\nnum_labels: 1\ntokenizer_type: AutoTokenizer\n\nreward_model: true\nchat_template: gemma\ndatasets:\n  - path: argilla/distilabel-intel-orca-dpo-pairs\n    type: bradley_terry.chat_template\n\nval_set_size: 0.1\neval_steps: 100\n\n\nProcess Reward Models (PRM)\nProcess reward models are trained using data which contains preference annotations for each step in a series of interactions. Typically, PRMs are trained to provide reward signals over each step of a reasoning trace and are used for downstream reinforcement learning.\nbase_model: Qwen/Qwen2.5-3B\nmodel_type: AutoModelForTokenClassification\nnum_labels: 2\n\nprocess_reward_model: true\ndatasets:\n  - path: trl-lib/math_shepherd\n    type: stepwise_supervised\n    split: train\n\nval_set_size: 0.1\neval_steps: 100",
+    "objectID": "docs/lora_optims.html#usage",
+    "href": "docs/lora_optims.html#usage",
+    "title": "LoRA Optimizations",
+    "section": "Usage",
+    "text": "Usage\nThese optimizations can be enabled in your Axolotl config YAML file. The lora_mlp_kernel option enables the optimized MLP path, while lora_qkv_kernel and lora_o_kernel enable the fused query-key-value projection and optimized output projection, respectively.\nlora_mlp_kernel: true\nlora_qkv_kernel: true\nlora_o_kernel: true",
     "crumbs": [
       "How To Guides",
-      "Reward Modelling"
+      "LoRA Optimizations"
     ]
   },
   {
-    "objectID": "docs/torchao.html",
-    "href": "docs/torchao.html",
-    "title": "PyTorch ao",
+    "objectID": "docs/lora_optims.html#requirements",
+    "href": "docs/lora_optims.html#requirements",
+    "title": "LoRA Optimizations",
+    "section": "Requirements",
+    "text": "Requirements\n\nOne or more NVIDIA or AMD GPUs (in order to use the Triton kernels)\n\nNote: Set TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 to enable memory-efficient attention on AMD GPUs\n\nTargeted LoRA adapters cannot use Dropout\n\nThis may limit model expressivity / cause overfitting\n\nTargeted LoRA adapters cannot have bias terms\n\nThis may limit model expressivity\n\n\nModels with pre-existing LoRA adapters that use Dropout or have bias terms may need to be re-finetuned without these features in order to be useful.",
+    "crumbs": [
+      "How To Guides",
+      "LoRA Optimizations"
+    ]
+  },
+  {
+    "objectID": "docs/lora_optims.html#implementation-details",
+    "href": "docs/lora_optims.html#implementation-details",
+    "title": "LoRA Optimizations",
+    "section": "Implementation details",
+    "text": "Implementation details\n\nCustom autograd functions\nThe LoRA MLP autograd function optimizes the entire MLP computation path. It fuses the LoRA and base weight computations together and provides a single, efficient backward pass for the entire MLP block.\nFor attention components, similar optimizations are provided through a function that handles the query, key, and value projections, and a function that handles the output projection. They are designed to work with the existing transformers attention implementation via some monkey-patching logic.\n\n\nTriton kernels\nTwo activation functions (SwiGLU and GeGLU) are implemented with Triton kernels for improved speed and memory performance. These kernels handle both the forward and backward passes.\n\n\nIntegration\nThe custom autograd functions and Triton kernels are designed to work together. The autograd function manages the high-level computation flow and gradient tracking, while calling the Triton kernels for the activation function computation. During the backward pass, the kernel computes both the activation output and the required gradients, which the autograd function then uses to compute the final gradients for the entire computation path.",
+    "crumbs": [
+      "How To Guides",
+      "LoRA Optimizations"
+    ]
+  },
+  {
+    "objectID": "docs/lora_optims.html#future-work",
+    "href": "docs/lora_optims.html#future-work",
+    "title": "LoRA Optimizations",
+    "section": "Future Work",
+    "text": "Future Work\n\nSupport for additional model architectures\nSupport for the FSDP setting\nSupport for dropout and bias\nAdditional operator fusions",
+    "crumbs": [
+      "How To Guides",
+      "LoRA Optimizations"
+    ]
+  },
+  {
+    "objectID": "docs/rlhf.html",
+    "href": "docs/rlhf.html",
+    "title": "RLHF (Beta)",
     "section": "",
-    "text": "To use experimental optimizers (AdamWFp8, AdamW4bit, AdamW8bit) from Pytorch Ao, please install the package as shown below.\n\n\n\n\n\n\nTip\n\n\n\nSome experimental optimizers are already present in regular Pytorch, so please re-check if you actually need this package!\n\n\n\nInstallation\nStable Release from the PyTorch index\npip install torchao --extra-index-url https://download.pytorch.org/whl/cu121 # full options are cpu/cu118/cu121/cu124\nNightly release\npip install --pre torchao-nightly --index-url https://download.pytorch.org/whl/nightly/cu121 # full options are cpu/cu118/cu121/cu124",
+    "text": "Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human feedback. Various methods include, but not limited to:\n\nDirect Preference Optimization (DPO)\nIdentity Preference Optimization (IPO)\nKahneman-Tversky Optimization (KTO)\nOdds Ratio Preference Optimization (ORPO)\nProximal Policy Optimization (PPO) (not yet supported in axolotl)",
+    "crumbs": [
+      "How To Guides",
+      "RLHF (Beta)"
+    ]
+  },
+  {
+    "objectID": "docs/rlhf.html#overview",
+    "href": "docs/rlhf.html#overview",
+    "title": "RLHF (Beta)",
+    "section": "",
+    "text": "Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human feedback. Various methods include, but not limited to:\n\nDirect Preference Optimization (DPO)\nIdentity Preference Optimization (IPO)\nKahneman-Tversky Optimization (KTO)\nOdds Ratio Preference Optimization (ORPO)\nProximal Policy Optimization (PPO) (not yet supported in axolotl)",
+    "crumbs": [
+      "How To Guides",
+      "RLHF (Beta)"
+    ]
+  },
+  {
+    "objectID": "docs/rlhf.html#rlhf-using-axolotl",
+    "href": "docs/rlhf.html#rlhf-using-axolotl",
+    "title": "RLHF (Beta)",
+    "section": "RLHF using Axolotl",
+    "text": "RLHF using Axolotl\n\n\n\n\n\n\nImportant\n\n\n\nThis is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.\n\n\nWe rely on the TRL library for implementations of various RL training methods, which we wrap around to expose in axolotl. Each method has their own supported ways of loading datasets and prompt formats.\n\n\n\n\n\n\nTip\n\n\n\nYou can find what each method supports by going into src/axolotl/prompt_strategies/{method} where {method} is one of our supported methods. The type: can be retrieved from {method}.{function_name}.\n\n\n\nDPO\nExample config:\nrl: dpo\ndatasets:\n  - path: Intel/orca_dpo_pairs\n    split: train\n    type: chatml.intel\n  - path: argilla/ultrafeedback-binarized-preferences\n    split: train\n    type: chatml\nDPO supports the following types with the following dataset format:\n\nchatml.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"chosen_response\": \"...\",\n    \"rejected_response\": \"...\"\n}\n\n\nchatml.argilla_chat\n{\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nchatml.icr\n{\n    \"system\": \"...\", // optional\n    \"input\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nchatml.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nchatml.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nchatml.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nllama3.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"chosen_response\": \"...\",\n    \"rejected_response\": \"...\"\n}\n\n\nllama3.argilla_chat\n{\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nllama3.icr\n{\n    \"system\": \"...\", // optional\n    \"input\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nllama3.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nllama3.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nllama3.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nzephyr.nectar\n{\n    \"prompt\": \"...\",\n    \"answers\": [\n        {\n            \"answer\": \"...\",\n            \"rank\": 1\n        },\n        {\n            \"answer\": \"...\",\n            \"rank\": 2\n        }\n        // ... more answers with ranks\n    ]\n}\n\n\nchat_template.default\nrl: dpo\ndatasets:\n  - path: ...\n    split: train\n    type: chat_template.default\n    field_messages: \"messages\"\n    field_chosen: \"chosen\"\n    field_rejected: \"rejected\"\n    message_property_mappings:\n      role: role\n      content: content\n    roles:\n      user: [\"user\"]\n      assistant: [\"assistant\"]\n      system: [\"system\"]\nSample input format:\n{\n    \"messages\": [\n        {\n            \"role\": \"system\",\n            \"content\": \"...\"\n        },\n        {\n            \"role\": \"user\",\n            \"content\": \"...\"\n        },\n        // ... more messages\n    ],\n    \"chosen\": {\n        \"role\": \"assistant\",\n        \"content\": \"...\"\n    },\n    \"rejected\": {\n        \"role\": \"assistant\",\n        \"content\": \"...\"\n    }\n}\n\n\nuser_defined.default\nFor custom behaviors,\nrl: dpo\ndatasets:\n  - path: ...\n    split: train\n    type: user_defined.default\n\n    field_prompt: \"prompt\"\n    field_system: \"system\"\n    field_chosen: \"chosen\"\n    field_rejected: \"rejected\"\n    prompt_format: \"{prompt}\"\n    chosen_format: \"{chosen}\"\n    rejected_format: \"{rejected}\"\nThe input format is a simple JSON input with customizable fields based on the above config.\n{\n    \"system\": \"...\",  // optional\n    \"prompt\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\n\nIPO\nAs IPO is just DPO with a different loss function, all supported options for DPO works here.\nrl: ipo\n\n\nORPO\nPaper: https://arxiv.org/abs/2403.07691\nrl: orpo\norpo_alpha: 0.1\nremove_unused_columns: false\n\nchat_template: chatml\ndatasets:\n  - path: argilla/ultrafeedback-binarized-preferences-cleaned\n    type: chat_template.argilla\nORPO supports the following types with the following dataset format:\n\nchat_template.argilla\n{\n    \"system\": \"...\",  // optional\n    \"prompt\": \"...\",  // if available, will be taken as user message for single-turn instead of from list below\n\n    // chosen/rejected should be same till last content and only even-number of alternating user/assistant turns\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\n\nKTO\nrl: kto\nrl_beta: 0.5\nkto_desirable_weight: 0.2\n\nremove_unused_columns: false\n\ndatasets:\n  - path: argilla/ultrafeedback-binarized-preferences-cleaned-kto\n    type: llama3.ultra\n    split: train\n\ngradient_checkpointing: true\ngradient_checkpointing_kwargs:\n  use_reentrant: true\nKTO supports the following types with the following dataset format:\n\nchatml.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nchatml.argilla_chat\n{\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"}\n    ],\n    \"completion\": [\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nchatml.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nchatml.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nchatml.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.argilla_chat\n{\n    \"completion\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nllama3.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nuser_defined.default\nFor custom behaviors,\nrl: kto\ndatasets:\n  - path: ...\n    split: train\n    type: user_defined.default\n\n    field_prompt: \"prompt\"\n    field_system: \"system\"\n    field_completion: \"completion\"\n    field_label: \"label\"\n    prompt_format: \"{prompt}\"\n    completion_format: \"{completion}\"\nThe input format is a simple JSON input with customizable fields based on the above config.\n{\n    \"system\": \"...\",  // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\",\n    \"label\": \"...\"\n}\n\n\n\nGRPO\nGRPO uses custom reward functions and transformations. Please have them ready locally.\nFor ex, to load OpenAI’s GSM8K and use a random reward for completions:\n# rewards.py\nimport random\n\ndef rand_reward_func(completions, **kwargs) -&gt; list[float]:\n    return [random.uniform(0, 1) for _ in completions]\n\ndef oai_gsm8k_transform(cfg, *args, **kwargs):\n    def transform_fn(example, tokenizer=None):\n        label = example[\"answer\"].split(\"####\")[-1].strip().replace(\",\", \"\")\n        return {\n            \"prompt\": [{\"role\": \"user\", \"content\": example[\"question\"]},],\n            \"answer\": label,\n        }\n    return transform_fn, {\"remove_columns\": [\"question\"]}\nrl: grpo\n\ntrl:\n    beta: 0.001\n    max_completion_length: 256\n    use_vllm: True\n    vllm_device: auto\n    vllm_gpu_memory_utilization: 0.15\n    num_generations: 4\n    reward_funcs: [\"rewards.rand_reward_func\"]    # format: '{file_name}.{fn_name}'\ndatasets:\n  - path: openai/gsm8k\n    name: main\n    type: rewards.oai_gsm8k_transform  # format: '{file_name}.{fn_name}'\nTo see other examples of custom reward functions, please see TRL GRPO Docs.\n\n\nUsing local dataset files\ndatasets:\n  - ds_type: json\n    data_files:\n      - orca_rlhf.jsonl\n    split: train\n    type: chatml.intel\n\n\nTRL auto-unwrapping for PEFT\nTRL supports auto-unwrapping PEFT models for RL training paradigms which rely on a reference model. This significantly reduces memory pressure as an additional refreference model does not need to be loaded, and reference model log-probabilities can be obtained by disabling PEFT adapters. This is enabled by default. To turn it off, pass the following config:\n# load ref model when adapter training.\nrl_adapter_ref_model: true",
+    "crumbs": [
+      "How To Guides",
+      "RLHF (Beta)"
+    ]
+  },
+  {
+    "objectID": "docs/cli.html",
+    "href": "docs/cli.html",
+    "title": "CLI Reference",
+    "section": "",
+    "text": "The Axolotl CLI provides a streamlined interface for training and fine-tuning large language models. This guide covers the CLI commands, their usage, and common examples.",
+    "crumbs": [
+      "Getting Started",
+      "CLI Reference"
+    ]
+  },
+  {
+    "objectID": "docs/cli.html#basic-commands",
+    "href": "docs/cli.html#basic-commands",
+    "title": "CLI Reference",
+    "section": "Basic Commands",
+    "text": "Basic Commands\nAll Axolotl commands follow this general structure:\naxolotl &lt;command&gt; [config.yml] [options]\nThe config file can be local or a URL to a raw YAML file.",
+    "crumbs": [
+      "Getting Started",
+      "CLI Reference"
+    ]
+  },
+  {
+    "objectID": "docs/cli.html#command-reference",
+    "href": "docs/cli.html#command-reference",
+    "title": "CLI Reference",
+    "section": "Command Reference",
+    "text": "Command Reference\n\nfetch\nDownloads example configurations and deepspeed configs to your local machine.\n# Get example YAML files\naxolotl fetch examples\n\n# Get deepspeed config files\naxolotl fetch deepspeed_configs\n\n# Specify custom destination\naxolotl fetch examples --dest path/to/folder\n\n\npreprocess\nPreprocesses and tokenizes your dataset before training. This is recommended for large datasets.\n# Basic preprocessing\naxolotl preprocess config.yml\n\n# Preprocessing with one GPU\nCUDA_VISIBLE_DEVICES=\"0\" axolotl preprocess config.yml\n\n# Debug mode to see processed examples\naxolotl preprocess config.yml --debug\n\n# Debug with limited examples\naxolotl preprocess config.yml --debug --debug-num-examples 5\nConfiguration options:\ndataset_prepared_path: Local folder for saving preprocessed data\npush_dataset_to_hub: HuggingFace repo to push preprocessed data (optional)\n\n\ntrain\nTrains or fine-tunes a model using the configuration specified in your YAML file.\n# Basic training\naxolotl train config.yml\n\n# Train and set/override specific options\naxolotl train config.yml \\\n    --learning-rate 1e-4 \\\n    --micro-batch-size 2 \\\n    --num-epochs 3\n\n# Training without accelerate\naxolotl train config.yml --no-accelerate\n\n# Resume training from checkpoint\naxolotl train config.yml --resume-from-checkpoint path/to/checkpoint\nIt is possible to run sweeps over multiple hyperparameters by passing in a sweeps config.\n# Basic training with sweeps\naxolotl train config.yml --sweep path/to/sweep.yaml\nExample sweep config:\n_:\n  # This section is for dependent variables we need to fix\n  - load_in_8bit: false\n    load_in_4bit: false\n    adapter: lora\n  - load_in_8bit: true\n    load_in_4bit: false\n    adapter: lora\n\n# These are independent variables\nlearning_rate: [0.0003, 0.0006]\nlora_r:\n  - 16\n  - 32\nlora_alpha:\n  - 16\n  - 32\n  - 64\n\n\ninference\nRuns inference using your trained model in either CLI or Gradio interface mode.\n# CLI inference with LoRA\naxolotl inference config.yml --lora-model-dir=\"./outputs/lora-out\"\n\n# CLI inference with full model\naxolotl inference config.yml --base-model=\"./completed-model\"\n\n# Gradio web interface\naxolotl inference config.yml --gradio \\\n    --lora-model-dir=\"./outputs/lora-out\"\n\n# Inference with input from file\ncat prompt.txt | axolotl inference config.yml \\\n    --base-model=\"./completed-model\"\n\n\nmerge-lora\nMerges trained LoRA adapters into the base model.\n# Basic merge\naxolotl merge-lora config.yml\n\n# Specify LoRA directory (usually used with checkpoints)\naxolotl merge-lora config.yml --lora-model-dir=\"./lora-output/checkpoint-100\"\n\n# Merge using CPU (if out of GPU memory)\nCUDA_VISIBLE_DEVICES=\"\" axolotl merge-lora config.yml\nConfiguration options:\ngpu_memory_limit: Limit GPU memory usage\nlora_on_cpu: Load LoRA weights on CPU\n\n\nmerge-sharded-fsdp-weights\nMerges sharded FSDP model checkpoints into a single combined checkpoint.\n# Basic merge\naxolotl merge-sharded-fsdp-weights config.yml\n\n\nevaluate\nEvaluates a model’s performance using metrics specified in the config.\n# Basic evaluation\naxolotl evaluate config.yml\n\n\nlm-eval\nRuns LM Evaluation Harness on your model.\n# Basic evaluation\naxolotl lm-eval config.yml\nConfiguration options:\n# List of tasks to evaluate\nlm_eval_tasks:\n  - arc_challenge\n  - hellaswag\nlm_eval_batch_size: # Batch size for evaluation\noutput_dir: # Directory to save evaluation results",
+    "crumbs": [
+      "Getting Started",
+      "CLI Reference"
+    ]
+  },
+  {
+    "objectID": "docs/cli.html#legacy-cli-usage",
+    "href": "docs/cli.html#legacy-cli-usage",
+    "title": "CLI Reference",
+    "section": "Legacy CLI Usage",
+    "text": "Legacy CLI Usage\nWhile the new Click-based CLI is preferred, Axolotl still supports the legacy module-based CLI:\n# Preprocess\npython -m axolotl.cli.preprocess config.yml\n\n# Train\naccelerate launch -m axolotl.cli.train config.yml\n\n# Inference\naccelerate launch -m axolotl.cli.inference config.yml \\\n    --lora_model_dir=\"./outputs/lora-out\"\n\n# Gradio interface\naccelerate launch -m axolotl.cli.inference config.yml \\\n    --lora_model_dir=\"./outputs/lora-out\" --gradio\n\n\n\n\n\n\nImportant\n\n\n\nWhen overriding CLI parameters in the legacy CLI, use same notation as in yaml file (e.g., --lora_model_dir).\nNote: This differs from the new Click-based CLI, which uses dash notation (e.g., --lora-model-dir). Keep this in mind if you’re referencing newer documentation or switching between CLI versions.",
+    "crumbs": [
+      "Getting Started",
+      "CLI Reference"
+    ]
+  },
+  {
+    "objectID": "docs/cli.html#remote-compute-with-modal-cloud",
+    "href": "docs/cli.html#remote-compute-with-modal-cloud",
+    "title": "CLI Reference",
+    "section": "Remote Compute with Modal Cloud",
+    "text": "Remote Compute with Modal Cloud\nAxolotl supports running training and inference workloads on Modal cloud infrastructure. This is configured using a cloud YAML file alongside your regular Axolotl config.\n\nCloud Configuration\nCreate a cloud config YAML with your Modal settings:\n# cloud_config.yml\nprovider: modal\ngpu: a100  # Supported: l40s, a100-40gb, a100-80gb, a10g, h100, t4, l4\ngpu_count: 1    # Number of GPUs to use\ntimeout: 86400  # Maximum runtime in seconds (24 hours)\nbranch: main    # Git branch to use (optional)\n\nvolumes:        # Persistent storage volumes\n  - name: axolotl-cache\n    mount: /workspace/cache\n  - name: axolotl-data\n    mount: /workspace/data\n  - name: axolotl-artifacts\n    mount: /workspace/artifacts\n\nenv:            # Environment variables\n  - WANDB_API_KEY\n  - HF_TOKEN\n\n\nRunning on Modal Cloud\nCommands that support the –cloud flag:\n# Preprocess on cloud\naxolotl preprocess config.yml --cloud cloud_config.yml\n\n# Train on cloud\naxolotl train config.yml --cloud cloud_config.yml\n\n# Train without accelerate on cloud\naxolotl train config.yml --cloud cloud_config.yml --no-accelerate\n\n# Run lm-eval on cloud\naxolotl lm-eval config.yml --cloud cloud_config.yml\n\n\nCloud Configuration Options\nprovider: # compute provider, currently only `modal` is supported\ngpu: # GPU type to use\ngpu_count: # Number of GPUs (default: 1)\nmemory: # RAM in GB (default: 128)\ntimeout: # Maximum runtime in seconds\ntimeout_preprocess: # Preprocessing timeout\nbranch: # Git branch to use\ndocker_tag: # Custom Docker image tag\nvolumes: # List of persistent storage volumes\nenv: # Environment variables to pass\nsecrets: # Secrets to inject",
+    "crumbs": [
+      "Getting Started",
+      "CLI Reference"
+    ]
+  },
+  {
+    "objectID": "docs/unsloth.html",
+    "href": "docs/unsloth.html",
+    "title": "Unsloth",
+    "section": "",
+    "text": "Overview\nUnsloth provides hand-written optimized kernels for LLM finetuning that slightly improve speed and VRAM over standard industry baselines.\n\n\n\n\n\n\nImportant\n\n\n\nDue to breaking changes in transformers v4.48.0, users will need to downgrade to &lt;=v4.47.1 to use this patch.\nThis will later be deprecated in favor of LoRA Optimizations.\n\n\n\n\nInstallation\nThe following will install the correct unsloth and extras from source.\npython scripts/unsloth_install.py | sh\n\n\nUsage\nAxolotl exposes a few configuration options to try out unsloth and get most of the performance gains.\nOur unsloth integration is currently limited to the following model architectures: - llama\nThese options are specific to LoRA finetuning and cannot be used for multi-GPU finetuning\nunsloth_lora_mlp: true\nunsloth_lora_qkv: true\nunsloth_lora_o: true\nThese options are composable and can be used with multi-gpu finetuning\nunsloth_cross_entropy_loss: true\nunsloth_rms_norm: true\nunsloth_rope: true\n\n\nLimitations\n\nSingle GPU only; e.g. no multi-gpu support\nNo deepspeed or FSDP support (requires multi-gpu)\nLoRA + QLoRA support only. No full fine tunes or fp8 support.\nLimited model architecture support. Llama, Phi, Gemma, Mistral only\nNo MoE support.",
     "crumbs": [
       "Advanced Features",
-      "PyTorch ao"
+      "Unsloth"
     ]
   },
   {
-    "objectID": "docs/installation.html",
-    "href": "docs/installation.html",
-    "title": "Installation",
+    "objectID": "docs/fsdp_qlora.html",
+    "href": "docs/fsdp_qlora.html",
+    "title": "FDSP + QLoRA",
     "section": "",
-    "text": "This guide covers all the ways you can install and set up Axolotl for your environment.",
+    "text": "Using FSDP with QLoRA is essential for fine-tuning larger (70b+ parameter) LLMs on consumer GPUs. For example, you can use FSDP + QLoRA to train a 70b model on two 24GB GPUs1.\nBelow, we describe how to use this feature in Axolotl.",
     "crumbs": [
-      "Getting Started",
-      "Installation"
+      "Advanced Features",
+      "FDSP + QLoRA"
     ]
   },
   {
-    "objectID": "docs/installation.html#sec-requirements",
-    "href": "docs/installation.html#sec-requirements",
-    "title": "Installation",
-    "section": "1 Requirements",
-    "text": "1 Requirements\n\nNVIDIA GPU (Ampere architecture or newer for bf16 and Flash Attention) or AMD GPU\nPython ≥3.10\nPyTorch ≥2.4.1",
-    "crumbs": [
-      "Getting Started",
-      "Installation"
-    ]
-  },
-  {
-    "objectID": "docs/installation.html#sec-installation-methods",
-    "href": "docs/installation.html#sec-installation-methods",
-    "title": "Installation",
-    "section": "2 Installation Methods",
-    "text": "2 Installation Methods\n\n2.1 PyPI Installation (Recommended)\npip3 install --no-build-isolation axolotl[flash-attn,deepspeed]\nWe use --no-build-isolation in order to detect the installed PyTorch version (if installed) in order not to clobber it, and so that we set the correct version of dependencies that are specific to the PyTorch version or other installed co-dependencies.\n\n\n2.2 Edge/Development Build\nFor the latest features between releases:\ngit clone https://github.com/axolotl-ai-cloud/axolotl.git\ncd axolotl\npip3 install packaging ninja\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n\n\n2.3 Docker\ndocker run --gpus '\"all\"' --rm -it axolotlai/axolotl:main-latest\nFor development with Docker:\ndocker compose up -d\n\n\n\n\n\n\nAdvanced Docker Configuration\n\n\n\ndocker run --privileged --gpus '\"all\"' --shm-size 10g --rm -it \\\n  --name axolotl --ipc=host \\\n  --ulimit memlock=-1 --ulimit stack=67108864 \\\n  --mount type=bind,src=\"${PWD}\",target=/workspace/axolotl \\\n  -v ${HOME}/.cache/huggingface:/root/.cache/huggingface \\\n  axolotlai/axolotl:main-latest",
-    "crumbs": [
-      "Getting Started",
-      "Installation"
-    ]
-  },
-  {
-    "objectID": "docs/installation.html#sec-cloud",
-    "href": "docs/installation.html#sec-cloud",
-    "title": "Installation",
-    "section": "3 Cloud Environments",
-    "text": "3 Cloud Environments\n\n3.1 Cloud GPU Providers\nFor providers supporting Docker:\n\nUse axolotlai/axolotl-cloud:main-latest\nAvailable on:\n\nLatitude.sh\nJarvisLabs.ai\nRunPod\n\n\n\n\n3.2 Google Colab\nUse our example notebook.",
-    "crumbs": [
-      "Getting Started",
-      "Installation"
-    ]
-  },
-  {
-    "objectID": "docs/installation.html#sec-platform-specific",
-    "href": "docs/installation.html#sec-platform-specific",
-    "title": "Installation",
-    "section": "4 Platform-Specific Instructions",
-    "text": "4 Platform-Specific Instructions\n\n4.1 macOS\npip3 install --no-build-isolation -e '.'\nSee Section 6 for Mac-specific issues.\n\n\n4.2 Windows\n\n\n\n\n\n\nImportant\n\n\n\nWe recommend using WSL2 (Windows Subsystem for Linux) or Docker.",
-    "crumbs": [
-      "Getting Started",
-      "Installation"
-    ]
-  },
-  {
-    "objectID": "docs/installation.html#sec-env-managers",
-    "href": "docs/installation.html#sec-env-managers",
-    "title": "Installation",
-    "section": "5 Environment Managers",
-    "text": "5 Environment Managers\n\n5.1 Conda/Pip venv\n\nInstall Python ≥3.10\nInstall PyTorch: https://pytorch.org/get-started/locally/\nInstall Axolotl:\npip3 install packaging\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n(Optional) Login to Hugging Face:\nhuggingface-cli login",
-    "crumbs": [
-      "Getting Started",
-      "Installation"
-    ]
-  },
-  {
-    "objectID": "docs/installation.html#sec-troubleshooting",
-    "href": "docs/installation.html#sec-troubleshooting",
-    "title": "Installation",
-    "section": "6 Troubleshooting",
-    "text": "6 Troubleshooting\nIf you encounter installation issues, see our FAQ and Debugging Guide.",
-    "crumbs": [
-      "Getting Started",
-      "Installation"
-    ]
-  },
-  {
-    "objectID": "docs/multi-gpu.html",
-    "href": "docs/multi-gpu.html",
-    "title": "Multi-GPU",
+    "objectID": "docs/fsdp_qlora.html#background",
+    "href": "docs/fsdp_qlora.html#background",
+    "title": "FDSP + QLoRA",
     "section": "",
-    "text": "This guide covers advanced training configurations for multi-GPU setups using Axolotl.",
+    "text": "Using FSDP with QLoRA is essential for fine-tuning larger (70b+ parameter) LLMs on consumer GPUs. For example, you can use FSDP + QLoRA to train a 70b model on two 24GB GPUs1.\nBelow, we describe how to use this feature in Axolotl.",
     "crumbs": [
-      "Deployments",
-      "Multi-GPU"
+      "Advanced Features",
+      "FDSP + QLoRA"
     ]
   },
   {
-    "objectID": "docs/multi-gpu.html#sec-overview",
-    "href": "docs/multi-gpu.html#sec-overview",
-    "title": "Multi-GPU",
-    "section": "1 Overview",
-    "text": "1 Overview\nAxolotl supports several methods for multi-GPU training:\n\nDeepSpeed (recommended)\nFSDP (Fully Sharded Data Parallel)\nFSDP + QLoRA",
+    "objectID": "docs/fsdp_qlora.html#usage",
+    "href": "docs/fsdp_qlora.html#usage",
+    "title": "FDSP + QLoRA",
+    "section": "Usage",
+    "text": "Usage\nTo enable QLoRA with FSDP, you need to perform the following steps:\n\n![Tip] See the example config file in addition to reading these instructions.\n\n\nSet adapter: qlora in your axolotl config file.\nEnable FSDP in your axolotl config, as described here.\nUse one of the supported model types: llama, mistral or mixtral.",
     "crumbs": [
-      "Deployments",
-      "Multi-GPU"
+      "Advanced Features",
+      "FDSP + QLoRA"
     ]
   },
   {
-    "objectID": "docs/multi-gpu.html#sec-deepspeed",
-    "href": "docs/multi-gpu.html#sec-deepspeed",
-    "title": "Multi-GPU",
-    "section": "2 DeepSpeed",
-    "text": "2 DeepSpeed\nDeepSpeed is the recommended approach for multi-GPU training due to its stability and performance. It provides various optimization levels through ZeRO stages.\n\n2.1 Configuration\nAdd to your YAML config:\ndeepspeed: deepspeed_configs/zero1.json\n\n\n2.2 Usage\n# Passing arg via config\naxolotl train config.yml\n\n# Passing arg via cli\naxolotl train config.yml --deepspeed deepspeed_configs/zero1.json\n\n\n2.3 ZeRO Stages\nWe provide default configurations for:\n\nZeRO Stage 1 (zero1.json)\nZeRO Stage 2 (zero2.json)\nZeRO Stage 3 (zero3.json)\n\nChoose based on your memory requirements and performance needs.",
+    "objectID": "docs/fsdp_qlora.html#example-config",
+    "href": "docs/fsdp_qlora.html#example-config",
+    "title": "FDSP + QLoRA",
+    "section": "Example Config",
+    "text": "Example Config\nexamples/llama-2/qlora-fsdp.yml contains an example of how to enable QLoRA + FSDP in axolotl.",
     "crumbs": [
-      "Deployments",
-      "Multi-GPU"
+      "Advanced Features",
+      "FDSP + QLoRA"
     ]
   },
   {
-    "objectID": "docs/multi-gpu.html#sec-fsdp",
-    "href": "docs/multi-gpu.html#sec-fsdp",
-    "title": "Multi-GPU",
-    "section": "3 FSDP",
-    "text": "3 FSDP\n\n3.1 Basic FSDP Configuration\nfsdp:\n  - full_shard\n  - auto_wrap\nfsdp_config:\n  fsdp_offload_params: true\n  fsdp_state_dict_type: FULL_STATE_DICT\n  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer\n\n\n3.2 FSDP + QLoRA\nFor combining FSDP with QLoRA, see our dedicated guide.",
+    "objectID": "docs/fsdp_qlora.html#references",
+    "href": "docs/fsdp_qlora.html#references",
+    "title": "FDSP + QLoRA",
+    "section": "References",
+    "text": "References\n\nPR #1378 enabling QLoRA in FSDP in Axolotl.\nBlog Post from the Answer.AI team describing the work that enabled QLoRA in FSDP.\nRelated HuggingFace PRs Enabling FDSP + QLoRA:\n\nAccelerate PR#2544\nTransformers PR#29587\nTRL PR#1416\nPEFT PR#1550",
     "crumbs": [
-      "Deployments",
-      "Multi-GPU"
+      "Advanced Features",
+      "FDSP + QLoRA"
     ]
   },
   {
-    "objectID": "docs/multi-gpu.html#sec-performance",
-    "href": "docs/multi-gpu.html#sec-performance",
-    "title": "Multi-GPU",
-    "section": "4 Performance Optimization",
-    "text": "4 Performance Optimization\n\n4.1 Liger Kernel Integration\nPlease see docs for more info.",
+    "objectID": "docs/fsdp_qlora.html#footnotes",
+    "href": "docs/fsdp_qlora.html#footnotes",
+    "title": "FDSP + QLoRA",
+    "section": "Footnotes",
+    "text": "Footnotes\n\n\nThis was enabled by this work from the Answer.AI team.↩︎",
     "crumbs": [
-      "Deployments",
-      "Multi-GPU"
+      "Advanced Features",
+      "FDSP + QLoRA"
     ]
   },
   {
-    "objectID": "docs/multi-gpu.html#sec-troubleshooting",
-    "href": "docs/multi-gpu.html#sec-troubleshooting",
-    "title": "Multi-GPU",
-    "section": "5 Troubleshooting",
-    "text": "5 Troubleshooting\n\n5.1 NCCL Issues\nFor NCCL-related problems, see our NCCL troubleshooting guide.\n\n\n5.2 Common Problems\n\nMemory IssuesTraining Instability\n\n\n\nReduce micro_batch_size\nReduce eval_batch_size\nAdjust gradient_accumulation_steps\nConsider using a higher ZeRO stage\n\n\n\n\nStart with DeepSpeed ZeRO-2\nMonitor loss values\nCheck learning rates\n\n\n\n\nFor more detailed troubleshooting, see our debugging guide.",
-    "crumbs": [
-      "Deployments",
-      "Multi-GPU"
-    ]
-  },
-  {
-    "objectID": "docs/config.html",
-    "href": "docs/config.html",
-    "title": "Config options",
+    "objectID": "docs/dataset_preprocessing.html",
+    "href": "docs/dataset_preprocessing.html",
+    "title": "Dataset Preprocessing",
     "section": "",
-    "text": "# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files\n# This can also be a relative path to a model on disk\nbase_model: ./llama-7b-hf\n# You can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)\nbase_model_ignore_patterns:\n# If the base_model repo on hf hub doesn't include configuration .json files,\n# You can set that here, or leave this empty to default to base_model\nbase_model_config: ./llama-7b-hf\n# You can specify to choose a specific model revision from huggingface hub\nrevision_of_model:\n# Optional tokenizer configuration path in case you want to use a different tokenizer\n# than the one defined in the base model\ntokenizer_config:\n# If you want to specify the type of model to load, AutoModelForCausalLM is a good choice too\nmodel_type: AutoModelForCausalLM\n# Corresponding tokenizer for the model AutoTokenizer is a good choice\ntokenizer_type: AutoTokenizer\n# Trust remote code for untrusted source\ntrust_remote_code:\n# use_fast option for tokenizer loading from_pretrained, default to True\ntokenizer_use_fast:\n# Whether to use the legacy tokenizer setting, defaults to True\ntokenizer_legacy:\n# Resize the model embeddings when new tokens are added to multiples of 32\n# This is reported to improve training speed on some models\nresize_token_embeddings_to_32x:\n\n# (Internal use only)\n# Used to identify which the model is based on\nis_falcon_derived_model:\nis_llama_derived_model:\nis_qwen_derived_model:\n# Please note that if you set this to true, `padding_side` will be set to \"left\" by default\nis_mistral_derived_model:\n\n# optional overrides to the base model configuration\noverrides_of_model_config:\n  # RoPE Scaling https://github.com/huggingface/transformers/pull/24653\n  rope_scaling:\n    type: # linear | dynamic\n    factor: # float\n\n# optional overrides the base model loading from_pretrained\noverrides_of_model_kwargs:\n  # use_cache: False\n\n# optional overrides to the bnb 4bit quantization configuration\n# https://huggingface.co/docs/transformers/main/main_classes/quantization#transformers.BitsAndBytesConfig\nbnb_config_kwargs:\n  # These are default values\n  llm_int8_has_fp16_weight: false\n  bnb_4bit_quant_type: nf4\n  bnb_4bit_use_double_quant: true\n\n\n# Whether you are training a 4-bit GPTQ quantized model\ngptq: true\n\n# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer\nload_in_8bit: true\n# Use bitsandbytes 4 bit\nload_in_4bit:\n\n# Use CUDA bf16\nbf16: true # bool or 'full' for `bf16_full_eval`. require &gt;=ampere\n# Use CUDA fp16\nfp16: true\n# Use CUDA tf32\ntf32: true # require &gt;=ampere\n\n# No AMP (automatic mixed precision)\nbfloat16: true # require &gt;=ampere\nfloat16: true\n\n# Limit the memory for all available GPUs to this amount (if an integer, expressed in gigabytes); default: unset\ngpu_memory_limit: 20GiB\n# Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge\nlora_on_cpu: true\n\n# A list of one or more datasets to finetune the model with\ndatasets:\n  # HuggingFace dataset repo | s3://,gs:// path | \"json\" for local dataset, make sure to fill data_files\n  - path: vicgalle/alpaca-gpt4\n    # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]\n    type: alpaca # format | format:&lt;prompt_style&gt; (chat/instruct) | &lt;prompt_strategies&gt;.load_&lt;load_fn&gt;\n    ds_type: # Optional[str] (json|arrow|parquet|text|csv) defines the datatype when path is a file\n    data_files: # Optional[str] path to source data files\n\n    shards: # Optional[int] split dataset into N pieces (use with shards_idx)\n    shards_idx: # Optional[int] = 0 the index of sharded dataset to use\n\n    preprocess_shards: # Optional[int] process dataset in N sequential chunks for memory efficiency (exclusive with `shards`)\n\n    name: # Optional[str] name of dataset configuration to load\n    train_on_split: train # Optional[str] name of dataset split to load from\n    revision: # Optional[str] The specific revision of the dataset to use when loading from the Hugging Face Hub. This can be a commit hash, tag, or branch name. If not specified, the latest version will be used. This parameter is ignored for local datasets.\n    trust_remote_code: # Optional[bool] Trust remote code for untrusted source\n\n  # Custom user instruction prompt\n  - path: repo\n    type:\n      # The below are defaults. only set what's needed if you use a different column name.\n      system_prompt: \"\"\n      system_format: \"{system}\"\n      field_system: system\n      field_instruction: instruction\n      field_input: input\n      field_output: output\n\n      # Customizable to be single line or multi-line\n      # Use {instruction}/{input} as key to be replaced\n      # 'format' can include {input}\n      format: |-\n        User: {instruction} {input}\n        Assistant:\n      # 'no_input_format' cannot include {input}\n      no_input_format: \"{instruction} \"\n\n      # For `completion` datsets only, uses the provided field instead of `text` column\n      field:\n\n  # Using chat template\n  - path: ...\n    # Set type to `chat_template` to use this strategy\n    type: chat_template\n    # Specify the name of the chat template to use\n    # The name of the chat template to use for training, following values are supported:\n    # - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default.\n    # - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py\n    # - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to if the tokenizer does not have a chat template else default to tokenizer. E.g. tokenizer_default_fallback_chatml.\n    # - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.\n    chat_template: tokenizer_default\n\n    # Custom jinja chat template. Used only if `chat_template: jinja` or empty.\n    chat_template_jinja:\n\n    # Key containing the messages (default: \"messages\")\n    field_messages: messages\n\n    # Mapping of properties from the input dataset to the chat template.\n    # (default: message_property_mappings={'role':'role', 'content':'content'})\n    # If a property exists in the template but not in this mapping, the system will attempt\n    # to load it directly from the message using the property name as the key.\n    # Example: In the mapping below, 'from' is loaded from input dataset and used as 'role',\n    # while 'value' is loaded and used as 'content' in the chat template.\n    message_property_mappings:\n      role: from\n      content: value\n      # ...\n\n    message_property_mappings:\n\n    # Optional[Dict[str, List]]. Roles mapping in the messages. The default is:\n    roles:\n      user: [\"human\", \"user\"]\n      assistant: [\"gpt\", \"assistant\"]\n      system: [\"system\"]\n      tool: [\"tool\"]\n\n    # IMPORTANT: The following fields determine which parts of the conversation to train on.\n    # Priority order: message_field_training &gt; message_field_training_detail &gt; train_on_inputs or role in roles_to_train\n    # See examples at `docs/dataset-formats/conversation.qmd`\n    # Note: If the below 4 fields are set to empty, defaults to training only on the last message.\n\n    # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.\n    roles_to_train: [\"assistant\"]  # default\n    # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:\n    # - all: train on all EOS tokens\n    # - turn (default): train on the EOS token at the end of each trainable turn\n    # - last: train on the last EOS token in the conversation\n    # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.\n    train_on_eos: last\n    # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.\n    message_field_training: training\n    # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.\n    # The value of the key is a List[Dict] containing `begin_offset` (start character index in content), `end_offset` (end character index in content), and `train` (boolean whether to train).\n    message_field_training_detail: train_detail\n\n\n# If false, the datasets will not be shuffled and will keep their original order in `datasets`.\n# The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.\nshuffle_merged_datasets: true\n\nDeduplicates datasets and test_datasets with identical entries.\ndataset_exact_deduplication: true\n\n# A list of one or more datasets to eval the model with.\n# You can use either test_datasets, or val_set_size, but not both.\ntest_datasets:\n  - path: /workspace/data/eval.jsonl\n    ds_type: json\n    # You need to specify a split. For \"json\" datasets the default split is called \"train\".\n    split: train\n    type: completion\n    data_files:\n      - /workspace/data/eval.jsonl\n\n# use RL training: 'dpo', 'ipo', 'kto'\nrl:\n# whether to perform weighting if doing DPO training. Boolean.\ndpo_use_weighting:\n\n# reward modelling: `True` or `False`\nreward_model:\n\n# process reward modelling: `True` or `False`\nprocess_reward_model:\n\n# The name of the chat template to use for training, following values are supported:\n# - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.\n# - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py\n# - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to. E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not available in the tokenizer.\n# - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.\n# The selected chat template will be saved to the tokenizer_config.json for easier inferencing\n# Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template.\nchat_template: tokenizer_default\n# custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null.\nchat_template_jinja: null\n# Changes the default system message\ndefault_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.\n# Axolotl attempts to save the dataset as an arrow after packing the data together so\n# subsequent training attempts load faster, relative path\ndataset_prepared_path: data/last_run_prepared\n# Push prepared dataset to hub\npush_dataset_to_hub: # repo path\n# The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`\n# if not set.\ndataset_processes: # defaults to os.cpu_count() if not set\n# Keep dataset in memory while preprocessing\n# Only needed if cached dataset is taking too much storage\ndataset_keep_in_memory:\n# push checkpoints to hub\nhub_model_id: # private repo path to push finetuned model\n# how to push checkpoints to hub\n# https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy\nhub_strategy:\n# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets\n# Required to be true when used in combination with `push_dataset_to_hub`\nhf_use_auth_token: # boolean\n# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.\nval_set_size: 0.04\n# Num shards for whole dataset\ndataset_shard_num:\n# Index of shard to use for whole dataset\ndataset_shard_idx:\n\n# The maximum length of an input to train with, this should typically be less than 2048\n# as most models have a token/context limit of 2048\nsequence_len: 2048\n# Pad inputs so each step uses constant sized buffers\n# This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently\npad_to_sequence_len:\n# Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'\nsample_packing:\n# Set to 'false' if getting errors during eval with sample_packing on.\neval_sample_packing:\n# You can set these packing optimizations AFTER starting a training at least once.\n# The trainer will provide recommended values for these values.\nsample_packing_eff_est:\ntotal_num_tokens:\n# Increasing the following values helps with packing, but usually only slightly (&lt;%1.)\n# The number of samples packed at a time.\nsample_packing_group_size: 100000\n# The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.\nsample_packing_bin_size: 200\n# whether to concatenate samples during pretraining\npretraining_sample_concatenation:\n\n# Use batch flattening for speedups when not using sample_packing\nbatch_flattening:\n\n# Passed through to transformers when loading the model when launched without accelerate\n# Use `sequential` when training w/ model parallelism to limit memory\ndevice_map:\n# Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model.\nmax_memory:\n\n# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model\nadapter: lora\n# If you already have a lora model trained that you want to load, put that here.\n# This means after training, if you want to test the model, you should set this to the value of `output_dir`.\n# Note that if you merge an adapter to the base model, a new subdirectory `merged` will be created under the `output_dir`.\nlora_model_dir:\n\n# LoRA hyperparameters\n# For more details about the following options, see:\n# https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2\nlora_r: 8\nlora_alpha: 16\nlora_dropout: 0.05\nlora_target_modules:\n  - q_proj\n  - v_proj\n#  - k_proj\n#  - o_proj\n#  - gate_proj\n#  - down_proj\n#  - up_proj\nlora_target_linear: # If true, will target all linear modules\npeft_layers_to_transform: # The layer indices to transform, otherwise, apply to all layers\n\n# If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.\n# For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.\n# `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.\n# https://github.com/huggingface/peft/issues/334#issuecomment-1561727994\nlora_modules_to_save:\n#  - embed_tokens\n#  - lm_head\n\nlora_fan_in_fan_out: false\n\n# Apply custom LoRA autograd functions and activation function Triton kernels for\n# speed and memory savings\n# See: https://axolotl-ai-cloud.github.io/axolotl/docs/lora_optims.html\nlora_mlp_kernel: true\nlora_qkv_kernel: true\nlora_o_kernel: true\n\n# LoRA+ hyperparameters\n# For more details about the following options, see:\n# https://arxiv.org/abs/2402.12354  and `src/axolotl/core/train_builder.py`\nloraplus_lr_ratio: # loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4.\nloraplus_lr_embedding: #  loraplus learning rate for lora embedding layers. Default value is 1e-6.\n\npeft:\n  # Configuration options for loftq initialization for LoRA\n  # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization\n  loftq_config:\n    loftq_bits:  # typically 4 bits\n\n# ReLoRA configuration\n# Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed\nrelora_steps: # Number of steps per ReLoRA restart\nrelora_warmup_steps: # Number of per-restart warmup steps\nrelora_anneal_steps: # Number of anneal steps for each relora cycle\nrelora_prune_ratio: # threshold for optimizer magnitude when pruning\nrelora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings\n\n# wandb configuration if you're using it\n# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.\nwandb_mode: # \"offline\" to save run metadata locally and not sync to the server, \"disabled\" to turn off wandb\nwandb_project: # Your wandb project name\nwandb_entity: # A wandb Team name if using a Team\nwandb_watch:\nwandb_name: # Set the name of your wandb run\nwandb_run_id: # Set the ID of your wandb run\nwandb_log_model: # \"checkpoint\" to log model to wandb Artifacts every `save_steps` or \"end\" to log only at the end of training\n\n# mlflow configuration if you're using it\nmlflow_tracking_uri: # URI to mlflow\nmlflow_experiment_name: # Your experiment name\nmlflow_run_name: # Your run name\nhf_mlflow_log_artifacts:  # set to true to copy each saved checkpoint on each save to mlflow artifact registry\n\n# Comet configuration if you're using it\n# Make sure your `COMET_API_KEY` environment variable is set (recommended) or you login to Comet with `comet login`.\n# Check out our documentation for more details https://www.comet.com/docs/v2/api-and-sdk/python-sdk/reference/Experiment-Creation/#comet_ml.start\nuse_comet: # Enable or disable Comet integration.\ncomet_api_key: # API key for Comet. Recommended to set via `comet login`.\ncomet_workspace: # Workspace name in Comet. Defaults to the user's default workspace.\ncomet_project_name: # Project name in Comet. Defaults to Uncategorized.\ncomet_experiment_key: # Identifier for the experiment. Used to append data to an existing experiment or control the key of new experiments. Default to a random key.\ncomet_mode: # Create a new experiment (\"create\") or log to an existing one (\"get\"). Default (\"get_or_create\") auto-selects based on configuration.\ncomet_online: # Set to True to log data to Comet server, or False for offline storage. Default is True.\ncomet_experiment_config: # Dictionary for additional configuration settings, see the doc for more details.\n\n# Tensorboard\nuse_tensorboard: # Optional[bool]\n\n# Where to save the full-finetuned model to\noutput_dir: ./completed-model\n\n# Whether to use torch.compile and which backend to use\n# setting to `auto` will enable torch compile when torch&gt;=2.5.1\ntorch_compile:  # Optional[Union[Literal[\"auto\"], bool]]\ntorch_compile_backend:  # Optional[str]\n\n# Training hyperparameters\n\n# If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.\ngradient_accumulation_steps: 1\n# The number of samples to include in each batch. This is the number of samples sent to each GPU.\n# Batch size per gpu = micro_batch_size * gradient_accumulation_steps\nmicro_batch_size: 2\neval_batch_size:\nnum_epochs: 4\nwarmup_steps: 100  # cannot use with warmup_ratio\nwarmup_ratio: 0.05  # cannot use with warmup_steps\nlearning_rate: 0.00003\nlr_quadratic_warmup:\nlogging_steps:\neval_steps: # Leave empty to eval at each epoch, integer for every N steps. float for fraction of total steps\nevals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps\neval_strategy: # Set to `\"no\"` to skip evaluation, `\"epoch\"` at end of each epoch, leave empty to infer from `eval_steps`.\nsave_strategy: # Set to `\"no\"` to skip checkpoint saves, `\"epoch\"` at end of each epoch, `\"best\"` when better result is achieved, leave empty to infer from `save_steps`.\nsave_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps\nsaves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps\nsave_total_limit: # Checkpoints saved at a time\n# Maximum number of iterations to train for. It precedes num_epochs which means that\n# if both are set, num_epochs will not be guaranteed.\n# e.g., when 1 epoch is 1000 steps =&gt; `num_epochs: 2` and `max_steps: 100` will train for 100 steps\nmax_steps:\n\n# bool of whether to include tokens trainer per second in the training metrics. This iterates over the entire dataset once, so it takes some time.\ninclude_tokens_per_second: # Optional[bool]\n\n# whether to find batch size that fits in memory. Passed to underlying transformers Trainer\nauto_find_batch_size: # Optional[bool]\n\neval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0\neval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128\neval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is [\"sacrebleu\", \"comet\", \"ter\", \"chrf\", \"perplexity\"]\n\nprofiler_steps: # enable the pytorch profiler to capture the first N steps of training to the output_dir.\n                # see https://pytorch.org/blog/understanding-gpu-memory-1/ for more information\n                # snapshots can be visualized @ https://pytorch.org/memory_viz\n\nloss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)\nloss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)\n\n# Save model as safetensors (require safetensors package)\nsave_safetensors:\n\n# Whether to mask out or include the human's prompt from the training labels\ntrain_on_inputs: false\n# Group similarly sized data to minimize padding.\n# May be slower to start, as it must download and sort the entire dataset.\n# Note that training loss may have an oscillating pattern with this enabled.\ngroup_by_length: false\n\n# Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing\ngradient_checkpointing: false\n# additional kwargs to pass to the trainer for gradient checkpointing\n# gradient_checkpointing_kwargs:\n#   use_reentrant: true\n\n# Stop training after this many evaluation losses have increased in a row\n# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback\nearly_stopping_patience: 3\n\n# Specify a scheduler and kwargs to use with the optimizer\nlr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine\nlr_scheduler_kwargs:\ncosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr\ncosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)\n\n# For one_cycle optim\nlr_div_factor: # Learning rate div factor\n\n# Specify optimizer\n# Valid values are driven by the Transformers OptimizerNames class, see:\n# https://github.com/huggingface/transformers/blob/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134\n#\n# Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of\n# torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used\n# in the examples/ for your model and fine-tuning use case.\n#\n# Valid values for 'optimizer' include:\n# - adamw_hf\n# - adamw_torch\n# - adamw_torch_fused\n# - adamw_torch_xla\n# - adamw_apex_fused\n# - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version &gt;= 2.5.1)\n# - adafactor\n# - adamw_anyprecision\n# - sgd\n# - adagrad\n# - adamw_bnb_8bit\n# - lion_8bit\n# - lion_32bit\n# - paged_adamw_32bit\n# - paged_adamw_8bit\n# - paged_lion_32bit\n# - paged_lion_8bit\n# - galore_adamw\n# - galore_adamw_8bit\n# - galore_adafactor\n# - galore_adamw_layerwise\n# - galore_adamw_8bit_layerwise\n# - galore_adafactor_layerwise\noptimizer:\n# Dictionary of arguments to pass to the optimizer\noptim_args:\n# For Galore Optimizers the following optim_args are available\n# rank:  # type: int\n# update_proj_gap  # type: int\n# scale  # type: float\n# proj_type:  # type: str, default = std\n\n# The target modules to optimize, i.e. the module names that you would like to train, right now this is used only for GaLore algorithm\noptim_target_modules:\n# - self_attn  # for llama\n# - mlp\n\n# Specify weight decay\nweight_decay:\n# adamw hyperparams\nadam_beta1:\nadam_beta2:\nadam_epsilon:\n# Gradient clipping max norm\nmax_grad_norm:\n\n# Augmentation techniques\n# NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings\n# currently only supported on Llama and Mistral\nneftune_noise_alpha:\n\n# Whether to bettertransformers\nflash_optimum:\n# Whether to use xformers attention patch https://github.com/facebookresearch/xformers:\nxformers_attention:\n# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:\nflash_attention:\nflash_attn_cross_entropy:  # Whether to use flash-attention cross entropy implementation - advanced use only\nflash_attn_rms_norm:  # Whether to use flash-attention rms norm implementation - advanced use only\nflash_attn_fuse_qkv: # Whether to fuse QKV into a single operation\nflash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation\n# Whether to use scaled-dot-product attention\n# https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html\nsdp_attention:\n# Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf\ns2_attention:\n# Resume from a specific checkpoint dir\nresume_from_checkpoint:\n# If resume_from_checkpoint isn't set and you simply want it to start where it left off.\n# Be careful with this being turned on between different models.\nauto_resume_from_checkpoints: false\n\n# Don't mess with this, it's here for accelerate and torchrun\nlocal_rank:\n\n# Add or change special tokens.\n# If you add tokens here, you don't need to add them to the `tokens` list.\nspecial_tokens:\n  # bos_token: \"&lt;s&gt;\"\n  # eos_token: \"&lt;/s&gt;\"\n  # unk_token: \"&lt;unk&gt;\"\n  # pad_token: \"[PAD]\"\n\n# Add extra tokens.\ntokens:\n\n# FSDP\nfsdp:\nfsdp_config:\n\n# Deepspeed config path. e.g., deepspeed_configs/zero3.json\ndeepspeed:\n\n# Advanced DDP Arguments\nddp_timeout:\nddp_bucket_cap_mb:\nddp_broadcast_buffers:\n\n# Path to torch distx for optim 'adamw_anyprecision'\ntorchdistx_path:\n\n# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize\npretraining_dataset:\n\n# Debug mode\ndebug:\n\n# Seed\nseed:\n\n# Allow overwrite yml config using from cli\nstrict:",
+    "text": "Dataset pre-processing is the step where Axolotl takes each dataset you’ve configured alongside the dataset format and prompt strategies to:\n\nparse the dataset based on the dataset format\ntransform the dataset to how you would interact with the model based on the prompt strategy\ntokenize the dataset based on the configured model & tokenizer\nshuffle and merge multiple datasets together if using more than one\n\nThe processing of the datasets can happen one of two ways:\n\nBefore kicking off training by calling axolotl preprocess config.yaml --debug\nWhen training is started\n\n\n\nWhen training interactively or for sweeps (e.g. you are restarting the trainer often), processing the datasets can oftentimes be frustratingly slow. Pre-processing will cache the tokenized/formatted datasets according to a hash of dependent training parameters so that it will intelligently pull from its cache when possible.\nThe path of the cache is controlled by dataset_prepared_path: and is often left blank in example YAMLs as this leads to a more robust solution that prevents unexpectedly reusing cached data.\nIf dataset_prepared_path: is left empty, when training, the processed dataset will be cached in a default path of ./last_run_prepared/, but will ignore anything already cached there. By explicitly setting dataset_prepared_path: ./last_run_prepared, the trainer will use whatever pre-processed data is in the cache.\n\n\n\nLet’s say you are writing a custom prompt strategy or using a user-defined prompt template. Because the trainer cannot readily detect these changes, we cannot change the calculated hash value for the pre-processed dataset.\nIf you have dataset_prepared_path: ... set and change your prompt templating logic, it may not pick up the changes you made and you will be training over the old prompt.",
     "crumbs": [
-      "Reference",
-      "Config options"
+      "Core Concepts",
+      "Dataset Preprocessing"
+    ]
+  },
+  {
+    "objectID": "docs/dataset_preprocessing.html#overview",
+    "href": "docs/dataset_preprocessing.html#overview",
+    "title": "Dataset Preprocessing",
+    "section": "",
+    "text": "Dataset pre-processing is the step where Axolotl takes each dataset you’ve configured alongside the dataset format and prompt strategies to:\n\nparse the dataset based on the dataset format\ntransform the dataset to how you would interact with the model based on the prompt strategy\ntokenize the dataset based on the configured model & tokenizer\nshuffle and merge multiple datasets together if using more than one\n\nThe processing of the datasets can happen one of two ways:\n\nBefore kicking off training by calling axolotl preprocess config.yaml --debug\nWhen training is started\n\n\n\nWhen training interactively or for sweeps (e.g. you are restarting the trainer often), processing the datasets can oftentimes be frustratingly slow. Pre-processing will cache the tokenized/formatted datasets according to a hash of dependent training parameters so that it will intelligently pull from its cache when possible.\nThe path of the cache is controlled by dataset_prepared_path: and is often left blank in example YAMLs as this leads to a more robust solution that prevents unexpectedly reusing cached data.\nIf dataset_prepared_path: is left empty, when training, the processed dataset will be cached in a default path of ./last_run_prepared/, but will ignore anything already cached there. By explicitly setting dataset_prepared_path: ./last_run_prepared, the trainer will use whatever pre-processed data is in the cache.\n\n\n\nLet’s say you are writing a custom prompt strategy or using a user-defined prompt template. Because the trainer cannot readily detect these changes, we cannot change the calculated hash value for the pre-processed dataset.\nIf you have dataset_prepared_path: ... set and change your prompt templating logic, it may not pick up the changes you made and you will be training over the old prompt.",
+    "crumbs": [
+      "Core Concepts",
+      "Dataset Preprocessing"
+    ]
+  },
+  {
+    "objectID": "docs/custom_integrations.html",
+    "href": "docs/custom_integrations.html",
+    "title": "Custom Integrations",
+    "section": "",
+    "text": "Axolotl adds custom features through integrations. They are located within the src/axolotl/integrations directory.\nTo enable them, please check the respective documentations.",
+    "crumbs": [
+      "Advanced Features",
+      "Custom Integrations"
+    ]
+  },
+  {
+    "objectID": "docs/custom_integrations.html#cut-cross-entropy",
+    "href": "docs/custom_integrations.html#cut-cross-entropy",
+    "title": "Custom Integrations",
+    "section": "Cut Cross Entropy",
+    "text": "Cut Cross Entropy\nCut Cross Entropy reduces VRAM usage through optimization on the cross-entropy operation during loss calculation.\nSee https://github.com/apple/ml-cross-entropy\n\nRequirements\n\nPyTorch 2.4.0 or higher\n\n\n\nInstallation\nRun the following command to install cut_cross_entropy[transformers] if you don’t have it already.\npython scripts/cutcrossentropy_install.py | sh\n\npip3 uninstall -y cut-cross-entropy && pip3 install \"cut-cross-entropy @ git+https://github.com/apple/ml-cross-entropy.git@9c297c905f55b73594b5d650722d1e78183b77bd\"'\n\n\nUsage\nplugins:\n  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin\n\ncut_cross_entropy: true\n\n\nCitation\n@article{wijmans2024cut,\n  author       = {Erik Wijmans and\n                  Brody Huval and\n                  Alexander Hertzberg and\n                  Vladlen Koltun and\n                  Philipp Kr\\\"ahenb\\\"uhl},\n  title        = {Cut Your Losses in Large-Vocabulary Language Models},\n  journal      = {arXiv},\n  year         = {2024},\n  url          = {https://arxiv.org/abs/2411.09009},\n}\nPlease see reference here",
+    "crumbs": [
+      "Advanced Features",
+      "Custom Integrations"
+    ]
+  },
+  {
+    "objectID": "docs/custom_integrations.html#grokfast",
+    "href": "docs/custom_integrations.html#grokfast",
+    "title": "Custom Integrations",
+    "section": "Grokfast",
+    "text": "Grokfast\nSee https://github.com/ironjr/grokfast\n\nUsage\nplugins:\n  - axolotl.integrations.grokfast.GrokfastPlugin\n\ngrokfast_alpha: 2.0\ngrokfast_lamb: 0.98\n\n\nCitation\n@article{lee2024grokfast,\n    title={{Grokfast}: Accelerated Grokking by Amplifying Slow Gradients},\n    author={Lee, Jaerin and Kang, Bong Gyun and Kim, Kihoon and Lee, Kyoung Mu},\n    journal={arXiv preprint arXiv:2405.20233},\n    year={2024}\n}\nPlease see reference here",
+    "crumbs": [
+      "Advanced Features",
+      "Custom Integrations"
+    ]
+  },
+  {
+    "objectID": "docs/custom_integrations.html#knowledge-distillation-kd",
+    "href": "docs/custom_integrations.html#knowledge-distillation-kd",
+    "title": "Custom Integrations",
+    "section": "Knowledge Distillation (KD)",
+    "text": "Knowledge Distillation (KD)\n\nUsage\nplugins:\n  - \"axolotl.integrations.kd.KDPlugin\"\n\nkd_trainer: True\nkd_ce_alpha: 0.1\nkd_alpha: 0.9\nkd_temperature: 1.0\n\ntorch_compile: True  # torch&gt;=2.5.1, recommended to reduce vram\n\ndatasets:\n  - path: ...\n    type: \"axolotl.integrations.kd.chat_template\"\n    field_messages: \"messages_combined\"\n    logprobs_field: \"llm_text_generation_vllm_logprobs\"  # for kd only, field of logprobs\nAn example dataset can be found at axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample\nPlease see reference here",
+    "crumbs": [
+      "Advanced Features",
+      "Custom Integrations"
+    ]
+  },
+  {
+    "objectID": "docs/custom_integrations.html#liger-kernels",
+    "href": "docs/custom_integrations.html#liger-kernels",
+    "title": "Custom Integrations",
+    "section": "Liger Kernels",
+    "text": "Liger Kernels\nLiger Kernel provides efficient Triton kernels for LLM training, offering:\n\n20% increase in multi-GPU training throughput\n60% reduction in memory usage\nCompatibility with both FSDP and DeepSpeed\n\nSee https://github.com/linkedin/Liger-Kernel\n\nUsage\nplugins:\n  - axolotl.integrations.liger.LigerPlugin\nliger_rope: true\nliger_rms_norm: true\nliger_glu_activation: true\nliger_layer_norm: true\nliger_fused_linear_cross_entropy: true\n\n\nCitation\n@article{hsu2024ligerkernelefficienttriton,\n      title={Liger Kernel: Efficient Triton Kernels for LLM Training},\n      author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},\n      year={2024},\n      eprint={2410.10989},\n      archivePrefix={arXiv},\n      primaryClass={cs.LG},\n      url={https://arxiv.org/abs/2410.10989},\n      journal={arXiv preprint arXiv:2410.10989},\n}\nPlease see reference here",
+    "crumbs": [
+      "Advanced Features",
+      "Custom Integrations"
+    ]
+  },
+  {
+    "objectID": "docs/custom_integrations.html#language-model-evaluation-harness-lm-eval",
+    "href": "docs/custom_integrations.html#language-model-evaluation-harness-lm-eval",
+    "title": "Custom Integrations",
+    "section": "Language Model Evaluation Harness (LM Eval)",
+    "text": "Language Model Evaluation Harness (LM Eval)\nRun evaluation on model using the popular lm-evaluation-harness library.\nSee https://github.com/EleutherAI/lm-evaluation-harness\n\nUsage\nplugins:\n  - axolotl.integrations.lm_eval.LMEvalPlugin\n\nlm_eval_tasks:\n  - gsm8k\n  - hellaswag\n  - arc_easy\n\nlm_eval_batch_size: # Batch size for evaluation\noutput_dir: # Directory to save evaluation results\n\n\nCitation\n@misc{eval-harness,\n  author       = {Gao, Leo and Tow, Jonathan and Abbasi, Baber and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and Le Noac'h, Alain and Li, Haonan and McDonell, Kyle and Muennighoff, Niklas and Ociepa, Chris and Phang, Jason and Reynolds, Laria and Schoelkopf, Hailey and Skowron, Aviya and Sutawika, Lintang and Tang, Eric and Thite, Anish and Wang, Ben and Wang, Kevin and Zou, Andy},\n  title        = {A framework for few-shot language model evaluation},\n  month        = 07,\n  year         = 2024,\n  publisher    = {Zenodo},\n  version      = {v0.4.3},\n  doi          = {10.5281/zenodo.12608602},\n  url          = {https://zenodo.org/records/12608602}\n}\nPlease see reference here",
+    "crumbs": [
+      "Advanced Features",
+      "Custom Integrations"
+    ]
+  },
+  {
+    "objectID": "docs/custom_integrations.html#spectrum",
+    "href": "docs/custom_integrations.html#spectrum",
+    "title": "Custom Integrations",
+    "section": "Spectrum",
+    "text": "Spectrum\nby Eric Hartford, Lucas Atkins, Fernando Fernandes, David Golchinfar\nThis plugin contains code to freeze the bottom fraction of modules in a model, based on the Signal-to-Noise Ratio (SNR).\nSee https://github.com/cognitivecomputations/spectrum\n\nOverview\nSpectrum is a tool for scanning and evaluating the Signal-to-Noise Ratio (SNR) of layers in large language models. By identifying the top n% of layers with the highest SNR, you can optimize training efficiency.\n\n\nUsage\nplugins:\n  - axolotl.integrations.spectrum.SpectrumPlugin\n\nspectrum_top_fraction: 0.5\nspectrum_model_name: meta-llama/Meta-Llama-3.1-8B\n\n\nCitation\n@misc{hartford2024spectrumtargetedtrainingsignal,\n      title={Spectrum: Targeted Training on Signal to Noise Ratio},\n      author={Eric Hartford and Lucas Atkins and Fernando Fernandes Neto and David Golchinfar},\n      year={2024},\n      eprint={2406.06623},\n      archivePrefix={arXiv},\n      primaryClass={cs.LG},\n      url={https://arxiv.org/abs/2406.06623},\n}\nPlease see reference here",
+    "crumbs": [
+      "Advanced Features",
+      "Custom Integrations"
+    ]
+  },
+  {
+    "objectID": "docs/mac.html",
+    "href": "docs/mac.html",
+    "title": "Mac M-series",
+    "section": "",
+    "text": "Currently Axolotl on Mac is partially usable, many of the dependencies of Axolotl including Pytorch do not support MPS or have incomplete support.\nCurrent support:\n\nSupport for all models\nFull training of models\nLoRA training\nSample packing\nFP16 and BF16 (awaiting AMP support for MPS in Pytorch)\nTri-dao’s flash-attn (until it is supported use spd_attention as an alternative)\nxformers\nbitsandbytes (meaning no 4/8 bits loading and bnb optimizers)\nqlora\nDeepSpeed\n\nUntested:\n\nFSDP",
+    "crumbs": [
+      "Deployments",
+      "Mac M-series"
+    ]
+  },
+  {
+    "objectID": "docs/docker.html",
+    "href": "docs/docker.html",
+    "title": "Docker",
+    "section": "",
+    "text": "This section describes the different Docker images that are released by AxolotlAI at Docker Hub.",
+    "crumbs": [
+      "Deployments",
+      "Docker"
+    ]
+  },
+  {
+    "objectID": "docs/docker.html#base",
+    "href": "docs/docker.html#base",
+    "title": "Docker",
+    "section": "Base",
+    "text": "Base\nThe base image is the most minimal image that can install Axolotl. It is based on the nvidia/cuda image. It includes python, torch, git, git-lfs, awscli, pydantic, and more.\n\nImage\naxolotlai/axolotl-base\nLink: Docker Hub\n\n\nTags format\nmain-base-py{python_version}-cu{cuda_version}-{pytorch_version}\nTags examples:\n\nmain-base-py3.11-cu124-2.6.0\nmain-base-py3.11-cu124-2.5.1\nmain-base-py3.11-cu124-2.4.1",
+    "crumbs": [
+      "Deployments",
+      "Docker"
+    ]
+  },
+  {
+    "objectID": "docs/docker.html#main",
+    "href": "docs/docker.html#main",
+    "title": "Docker",
+    "section": "Main",
+    "text": "Main\nThe main image is the image that is used to run Axolotl. It is based on the axolotlai/axolotl-base image and includes the Axolotl codebase, dependencies, and more.\n\nImage\naxolotlai/axolotl\nLink: Docker Hub\n\n\nTags format\n# on push to main\nmain-py{python_version}-cu{cuda_version}-{pytorch_version}\n\n# latest main (currently torch 2.5.1, python 3.11, cuda 12.4)\nmain-latest\n\n# nightly build\n{branch}-{date_in_YYYYMMDD}-py{python_version}-cu{cuda_version}-{pytorch_version}\n\n# tagged release\n{version}\n\n\n\n\n\n\nTip\n\n\n\nThere may be some extra tags appended to the image, like -vllm which installs those packages.\n\n\nTags examples:\n\nmain-py3.11-cu124-2.6.0\nmain-py3.11-cu124-2.5.1\nmain-py3.11-cu124-2.4.1\nmain-latest\nmain-20250303-py3.11-cu124-2.6.0\nmain-20250303-py3.11-cu124-2.5.1\nmain-20250303-py3.11-cu124-2.4.1\n0.7.1",
+    "crumbs": [
+      "Deployments",
+      "Docker"
+    ]
+  },
+  {
+    "objectID": "docs/docker.html#cloud",
+    "href": "docs/docker.html#cloud",
+    "title": "Docker",
+    "section": "Cloud",
+    "text": "Cloud\nThe cloud image is the image that is used to run Axolotl in the cloud. It is based on the axolotlai/axolotl image and sets ENV variables like HuggingFace cache directories for volume mounts, tmux, and more for different cloud providers.\n\n\n\n\n\n\nTip\n\n\n\nJupyter lab is run by default. Set JUPYTER_DISABLE=1 in the environment variables to disable it.\n\n\n\nImage\naxolotlai/axolotl-cloud\nLink: Docker Hub\n\n\nTags format\nThis uses the same tags as the main image.\n\n\nEnvironment variables\n\nJUPYTER_DISABLE: Disable Jupyter lab.\nJUPYTER_PASSWORD: Set a password for the Jupyter lab.\nPUBLIC_KEY: Add a public key for the SSH service.\nSSH_KEY: Add a private key for the SSH service.\n\n\n\nVolume mounts\n\n\n\n\n\n\nTip\n\n\n\nWe recommend mounting volumes to /workspace/data for data persistence. /workspace/axolotl contains the source code and is ephemeral.\n\n\n\n/workspace/data/axolotl-artifacts: Directory to store Axolotl artifacts.\n/workspace/data/huggingface-cache: Directory to store HuggingFace cache.",
+    "crumbs": [
+      "Deployments",
+      "Docker"
+    ]
+  },
+  {
+    "objectID": "docs/docker.html#cloud-no-tmux",
+    "href": "docs/docker.html#cloud-no-tmux",
+    "title": "Docker",
+    "section": "Cloud-no-tmux",
+    "text": "Cloud-no-tmux\nThis is the same as the cloud image but without tmux.\n\nImage\naxolotlai/axolotl-cloud-term\nLink: Docker Hub\n\n\n\n\n\n\nNote\n\n\n\nThe naming may be a bit confusing as it has -term appended to the end.\n\n\n\n\nTags format\nThis uses the same tags as the cloud image.",
+    "crumbs": [
+      "Deployments",
+      "Docker"
     ]
   },
   {
@@ -1206,7 +1261,7 @@
     "href": "docs/dataset-formats/index.html#supervised-fine-tuning-sft",
     "title": "Dataset Formats",
     "section": "Supervised fine-tuning (SFT)",
-    "text": "Supervised fine-tuning (SFT)\nSupervised fine-tuning is the process of training models to respond to an instruction or chat input.\nAs there are a wide variety of dataset formats, Axolotl tries to support a majority of the formats available in public datasets.\nAxolotl provides four approaches for loading datasets, however, it’s easier to work backwards from the dataset you have available to figure out which approach to use.\nA flow chart is as follows:\n\nDo you already have the dataset tokenized? If yes, check Pre-Tokenized Dataset.\nDo you want to format the dataset yourself and manually choose each section to mask? If yes, check Template Free Dataset\nIs your dataset in a “conversation” format, containing a list[messages]? If yes, check Conversation Dataset\nIs your dataset in an “instruct” format, containing { instruction, response }? If yes, check Instruction Dataset\n\nIf you went through the flow chart and did not find one that matches, it is recommended to preprocess your dataset into one of the above or create a thread on Github Discussion.\n\n\n\n\n\n\nTip\n\n\n\nYou can mix and match within each approach or across approaches to train a model on a variety of datasets.\n\n\n\nPre-Tokenized Dataset\nWe suggest this approach when you want to bring your own tokenized dataset.\nAxolotl expects the dataset to have three keys: - input_ids: from tokenizing formatted prompt - attention_mask: for masking padding. If you don’t add padding, it would be equal to len(input_ids) * [1] - labels: this is the same as input_ids, however, if you want to mask certain tokens, you would set those indices to -100.\n\n\n\n\n\n\nTip\n\n\n\nMake sure to add BOS/EOS tokens to your prompt and mask it appropriately.\n\n\nA config for this would look like:\ndatasets:\n  - path: A.jsonl\n    type:\n\n\n\n\n\n\nNote\n\n\n\ntype: is empty!\n\n\nReference: Pre-Tokenized Dataset Documentation.\n\n\nTemplate Free Dataset\nWe reccomend this approach when you want granular control over the prompt formatting, special tokens, and masking, whilst letting Axolotl handle the tokenization. This is very useful if your dataset has unique prompts that differ across samples and where one single general template wouldn’t suffice.\nIn the example below, you could see that there is no proper structure. At the same time, it’s very flexible as there are no constraints on how your prompt can look.\n{\n    \"segments\": [\n        {\n            \"label\": true,\n            \"text\": \"&lt;s&gt;Hello\\n\"\n        },\n        {\n            \"label\": true,\n            \"text\": \"hi there!. \"\n        },\n        {\n            \"label\": false,\n            \"text\": \"goodbye \"\n        },\n        {\n            \"label\": true,\n            \"text\": \"farewell&lt;/s&gt;\"\n        }\n    ]\n}\nEach prompt must be have a key called segments which is a list of { text, label }.\ndatasets:\n  - path: A.jsonl\n    type: input_output\nReference: Template Free Documentation.\n\n\nConversation Dataset\nconversation messages are a list of messages which usually contain a role and content key.\n\n\n\n\n\n\nTip\n\n\n\nFun fact: Axolotl synonymously refers to “chat” messages as conversation messages due to how FastChat initially used this term to build a widely used fastchat conversation method for formatting chat messages prior to the creation of chat_templates.\n\n\n\nWhat are chat_templates?\nThe current most popular and convenient method for inference is to use chat_templates for formatting prompts. Axolotl supports using chat_templates for training to ensure that the model performs in the same environment as in inference.\nHere’s a quick rundown on chat_template: A chat_template is a Jinja2 template which formats a list of messages into a prompt.\nAn example of a prompt formatted into a popular template called ChatML can be seen below:\nSingle prompt (pretty-printed):\n{\n    \"messages\": [\n        {\n            \"role\": \"user\",\n            \"content\": \"Hi\"\n        },\n        {\n            \"role\": \"assistant\",\n            \"content\": \"How can I help you?\"\n        },\n        {\n            \"role\": \"user\",\n            \"content\": \"Can you add 3+5?\"\n        },\n        {\n            \"role\": \"assistant\",\n            \"content\": \"The answer is 8.\"\n        }\n    ]\n}\nThe ChatML template is as follows:\n{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'&lt;|im_start|&gt;' + message['role'] + '\\n' + message['content'] + '&lt;|im_end|&gt;' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '&lt;|im_start|&gt;assistant\\n' }}{% endif %}\nThe above prompt formatted into this template will result in:\n&lt;|im_start|&gt;user\nHi&lt;|im_end|&gt;\n&lt;|im_start|&gt;assistant\nHow can I help you?&lt;|im_end|&gt;\n&lt;|im_start|&gt;user\nCan you add 3+5?&lt;|im_end|&gt;\n&lt;|im_start|&gt;assistant\nThe answer is 8.&lt;|im_end|&gt;\nBy using delimiters (&lt;|im_start|&gt; and &lt;|im_end|&gt;), a prompt separates different speakers which helps the model identify which portion belongs to whom.\n\n\nCommon Conversation Dataset formats\nOlder conversation datasets with the following format are colloquially called sharegpt datasets.\n{\"conversations\": [{\"from\": \"...\", \"value\": \"...\"}]}\nNewer conversation datasets usually follow the OpenAI format.\n{\"messages\": [{\"role\": \"...\", \"content\": \"...\"}]}\nAxolotl supports both as well as allowing customization of any kind of key.\n\n\nChat Template Usage\nTo properly use this method, it is important to identify three things:\n\nWhich chat_template would you use?\nWhat are the keys in your dataset, and what are the possible roles? For example, in OpenAI format, the keys would be messages, role, and content, respectively, whereas the possible roles are system, user, and assistant.\nWhat do you want to mask? For instance, only assistant messages, only last message, or nothing.\n\n\nChoosing a chat_template\nThere are a lot of chat_templates out there. Axolotl supports the common ones: supported chat templates. For example, to use ChatML, it would be chat_template: chatml.\nHowever, it is also possible to use the already configured template within the tokenizer by specifying chat_template: tokenizer_default. If you want a fallback (in case some tokenizer does not have it pre-configured), you can do chat_template: tokenizer_default_fallback_chatml to fallback to the ChatML template if a tokenizer template was not found.\nOne last but powerful approach is to bring your own template. This can be set via:\nchat_template_jinja: # your template\n\n\nSetting chat_template dataset keys\nWe currently default to OpenAI format for dataset keys, so if that’s your current dataset format, there’s nothing to do here.\nIf your dataset format is different, here are the keys you should check (with their defaults):\ndatasets:\n    ...\n    field_messages: messages  # this should point to the key containing the list of conversations\n    message_property_mappings:  # this is a mapping from keys in your dataset to keys in chat_template\n      role: role\n      content: content\nIn some chat_templates (e.g. Gemma), the roles are hardcoded to user and assistant. Consequently, you may find it necessary to map the roles in your dataset to these above. We currently have some defaults that should work for common datasets, but if you get a KeyError, it would be necessary to add mapping for your roles. Here is an example of how it would look like:\ndatasets:\n    ...\n    roles:\n      assistant:\n        - gpt\n        - model\n      user:\n        - human\nIn the example above, all gpt and model values are converted to assistant. All human values are converted to user.\n\n\nHandling masking\nThe common use case for chat_template is for chat messages, therefore, it is common to mask all non-assistant messages. Assistant messages refer to the bot messages that you want the model to learn on.\nTo train on all assistant messages, you would set the following configs.\ndatasets:\n    ...\n    roles_to_train: [\"assistant\"]\n    train_on_eos: \"turn\"\nThe train_on_eos config means that it would mask all EOS tokens for turns that aren’t assistant-turns. The other options are: all and last to choose which EOS to train on.\nPerhaps, you want to train on assistant and narrator roles, you can simply add narrator to the list of roles_to_train. You would also need to add it to the mapping of roles above.\ndatasets:\n    ...\n    roles_to_train: [\"assistant\", \"narrator\"]\n    roles:\n      assistant:\n        - gpt\n        - model\n      user:\n        - human\n      narrator: [\"narrator\"]\n\n\n\n\n\n\nTip\n\n\n\nAs chat_templates may use hardcoded EOS/EOT tokens that are different from the tokenizer’s EOS, it is highly recommended to set them. For example, ChatML uses &lt;|im_end|&gt; to end turns.\nspecial_tokens:\n  eos_token: &lt;|im_end|&gt;\n\n\n\n\nApplying chat_template\nOnce all the above steps are completed, you could combine all these configs together to form a bespoke configuration for your custom dataset.\ndatasets:\n  - path: A.jsonl\n    type: chat_template\n\n    # step 1\n    chat_template: chatml\n\n    # step 2\n    field_messages: messages\n    message_property_mappings:\n      role: role\n      content: content\n\n    roles:\n      assistant:\n        - gpt\n        - model\n        - assistant\n      user:\n        - human\n        - user\n\n    # step 3\n    roles_to_train: [\"assistant\"]\n    train_on_eos: \"turn\"\n\nspecial_tokens:\n  eos_token: &lt;|im_end|&gt;\nIf this config were to be applied to the sample dataset above, the output would look as such (which can be retrieved via axolotl preprocess config.yaml --debug):\n&lt;|im_start|&gt;(-100, 128256) user(-100, 882)\n(-100, 198) Hi(-100, 13347) &lt;|im_end|&gt;(-100, 128257)\n(-100, 198) &lt;|im_start|&gt;(-100, 128256) assistant(-100, 78191)\n(-100, 198) How(4438, 4438)  can(649, 649)  I(358, 358)  help(1520, 1520)  you(499, 499) ?(30, 30) &lt;|im_end|&gt;(128257, 128257)\n(-100, 198) &lt;|im_start|&gt;(-100, 128256) user(-100, 882)\n(-100, 198) Can(-100, 6854)  you(-100, 499)  add(-100, 923)  (-100, 220) 3(-100, 18) +(-100, 10) 5(-100, 20) ?(-100, 30) &lt;|im_end|&gt;(-100, 128257)\n(-100, 198) &lt;|im_start|&gt;(-100, 128256) assistant(-100, 78191)\n(-100, 198) The(791, 791)  answer(4320, 4320)  is(374, 374)  (220, 220) 8(23, 23) .(13, 13) &lt;|im_end|&gt;(128257, 128257)\n(-100, 198)\nThe first number refers to the label, the second refers to the token_id. For example, -100 labels appear on non-assistant portions, meaning that they are masked during. For assistant portions, the label is the same as the token_id.\n\n\n\n\n\n\nNote\n\n\n\nIf during preprocess, there are a lot of warnings of Could not find content __ boundary, please check the FAQ section for chat_templates.\n\n\n\n\n\nReference\nPlease see docs here.\n\n\n\nInstruction Dataset\nInstruction datasets are used to train instruction-following models and comprise a prompt, containing an instruction, and a single response. In contrast to chat datasets which may be multi-turn, instruct datasets are typically single-turn.\nAn example is of a common format called Alpaca:\n{\"instruction\": \"...\", \"input\": \"...\", \"output\": \"...\"}\nUsing those keys, a prompt can be built based on it.\nBelow is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n{output}\nThis can be configured as such:\ndatasets:\n  - path: A.jsonl\n    type: alpaca\nAxolotl supports many kinds of instruction dataset. All of them can be found here (https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/inst_tune.html) with their respective type and sample row format.\nReference: Instruction Dataset Documentation.\n\nCustom Instruct Prompt Format\nDue to the myriad possibilities of instruction formats, Axolotl allows customizing your own instruction format without having to dive into the code directly.\nIn the example below, a sample row is used to output in mistral_v1 format.\n{\"input\": \"...\", \"output\": \"...\"}\ndatasets:\n  - path: repo\n    type:\n      system_prompt: \"\"\n\n      field_system:\n      field_instruction: input\n      field_input:\n      field_output: output\n\n      # multi-line example with input\n      format: |-\n        [INST] {instruction} {input} [/INST]\n\n      # single-line example without input\n      no_input_format: \"[INST] {instruction} [/INST]\"\nThe config sets that the field_instruction is actually named input, and the field_input is empty as we don’t have an input in this sample. Generally, instruction can be thought as the question to the model, and input as the additional information with output being the response. It is not necessary to have an input nor system. In the end, the most important part is to understand what format you want it to look like and how you can customize this to your use case.\nReference: Custom Instruct Prompt Format Documentation.",
+    "text": "Supervised fine-tuning (SFT)\nSupervised fine-tuning is the process of training models to respond to an instruction or chat input.\nAs there are a wide variety of dataset formats, Axolotl tries to support a majority of the formats available in public datasets.\nAxolotl provides four approaches for loading datasets, however, it’s easier to work backwards from the dataset you have available to figure out which approach to use.\nA flow chart is as follows:\n\nDo you already have the dataset tokenized? If yes, check Pre-Tokenized Dataset.\nDo you want to format the dataset yourself and manually choose each section to mask? If yes, check Template Free Dataset\nIs your dataset in a “conversation” format, containing a list[messages]? If yes, check Conversation Dataset\nIs your dataset in an “instruct” format, containing { instruction, response }? If yes, check Instruction Dataset\n\nIf you went through the flow chart and did not find one that matches, it is recommended to preprocess your dataset into one of the above or create a thread on Github Discussion.\n\n\n\n\n\n\nTip\n\n\n\nYou can mix and match within each approach or across approaches to train a model on a variety of datasets.\n\n\n\nPre-Tokenized Dataset\nWe suggest this approach when you want to bring your own tokenized dataset.\nAxolotl expects the dataset to have three keys:\n\ninput_ids: from tokenizing formatted prompt\nattention_mask: for masking padding. If you don’t add padding, it would be equal to len(input_ids) * [1]\nlabels: this is the same as input_ids, however, if you want to mask certain tokens, you would set those indices to -100.\n\n\n\n\n\n\n\nTip\n\n\n\nMake sure to add BOS/EOS tokens to your prompt and mask it appropriately.\n\n\nA config for this would look like:\ndatasets:\n  - path: A.jsonl\n    type:\n\n\n\n\n\n\nNote\n\n\n\ntype: is empty!\n\n\nReference: Pre-Tokenized Dataset Documentation.\n\n\nTemplate Free Dataset\nWe reccomend this approach when you want granular control over the prompt formatting, special tokens, and masking, whilst letting Axolotl handle the tokenization. This is very useful if your dataset has unique prompts that differ across samples and where one single general template wouldn’t suffice.\nIn the example below, you could see that there is no proper structure. At the same time, it’s very flexible as there are no constraints on how your prompt can look.\n{\n    \"segments\": [\n        {\n            \"label\": true,\n            \"text\": \"&lt;s&gt;Hello\\n\"\n        },\n        {\n            \"label\": true,\n            \"text\": \"hi there!. \"\n        },\n        {\n            \"label\": false,\n            \"text\": \"goodbye \"\n        },\n        {\n            \"label\": true,\n            \"text\": \"farewell&lt;/s&gt;\"\n        }\n    ]\n}\nEach prompt must be have a key called segments which is a list of { text, label }.\ndatasets:\n  - path: A.jsonl\n    type: input_output\nReference: Template Free Documentation.\n\n\nConversation Dataset\nconversation messages are a list of messages which usually contain a role and content key.\n\n\n\n\n\n\nTip\n\n\n\nFun fact: Axolotl synonymously refers to “chat” messages as conversation messages due to how FastChat initially used this term to build a widely used fastchat conversation method for formatting chat messages prior to the creation of chat_templates.\n\n\n\nWhat are chat_templates?\nThe current most popular and convenient method for inference is to use chat_templates for formatting prompts. Axolotl supports using chat_templates for training to ensure that the model performs in the same environment as in inference.\nHere’s a quick rundown on chat_template: A chat_template is a Jinja2 template which formats a list of messages into a prompt.\nAn example of a prompt formatted into a popular template called ChatML can be seen below:\nSingle prompt (pretty-printed):\n{\n    \"messages\": [\n        {\n            \"role\": \"user\",\n            \"content\": \"Hi\"\n        },\n        {\n            \"role\": \"assistant\",\n            \"content\": \"How can I help you?\"\n        },\n        {\n            \"role\": \"user\",\n            \"content\": \"Can you add 3+5?\"\n        },\n        {\n            \"role\": \"assistant\",\n            \"content\": \"The answer is 8.\"\n        }\n    ]\n}\nThe ChatML template is as follows:\n{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'&lt;|im_start|&gt;' + message['role'] + '\\n' + message['content'] + '&lt;|im_end|&gt;' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '&lt;|im_start|&gt;assistant\\n' }}{% endif %}\nThe above prompt formatted into this template will result in:\n&lt;|im_start|&gt;user\nHi&lt;|im_end|&gt;\n&lt;|im_start|&gt;assistant\nHow can I help you?&lt;|im_end|&gt;\n&lt;|im_start|&gt;user\nCan you add 3+5?&lt;|im_end|&gt;\n&lt;|im_start|&gt;assistant\nThe answer is 8.&lt;|im_end|&gt;\nBy using delimiters (&lt;|im_start|&gt; and &lt;|im_end|&gt;), a prompt separates different speakers which helps the model identify which portion belongs to whom.\n\n\nCommon Conversation Dataset formats\nOlder conversation datasets with the following format are colloquially called sharegpt datasets.\n{\"conversations\": [{\"from\": \"...\", \"value\": \"...\"}]}\nNewer conversation datasets usually follow the OpenAI format.\n{\"messages\": [{\"role\": \"...\", \"content\": \"...\"}]}\nAxolotl supports both as well as allowing customization of any kind of key.\n\n\nChat Template Usage\nTo properly use this method, it is important to identify three things:\n\nWhich chat_template would you use?\nWhat are the keys in your dataset, and what are the possible roles? For example, in OpenAI format, the keys would be messages, role, and content, respectively, whereas the possible roles are system, user, and assistant.\nWhat do you want to mask? For instance, only assistant messages, only last message, or nothing.\n\n\nChoosing a chat_template\nThere are a lot of chat_templates out there. Axolotl supports the common ones: supported chat templates. For example, to use ChatML, it would be chat_template: chatml.\nHowever, it is also possible to use the already configured template within the tokenizer by specifying chat_template: tokenizer_default. If you want a fallback (in case some tokenizer does not have it pre-configured), you can do chat_template: tokenizer_default_fallback_chatml to fallback to the ChatML template if a tokenizer template was not found.\nOne last but powerful approach is to bring your own template. This can be set via:\nchat_template_jinja: # your template\n\n\nSetting chat_template dataset keys\nWe currently default to OpenAI format for dataset keys, so if that’s your current dataset format, there’s nothing to do here.\nIf your dataset format is different, here are the keys you should check (with their defaults):\ndatasets:\n    ...\n    field_messages: messages  # this should point to the key containing the list of conversations\n    message_property_mappings:  # this is a mapping from keys in your dataset to keys in chat_template\n      role: role\n      content: content\nIn some chat_templates (e.g. Gemma), the roles are hardcoded to user and assistant. Consequently, you may find it necessary to map the roles in your dataset to these above. We currently have some defaults that should work for common datasets, but if you get a KeyError, it would be necessary to add mapping for your roles. Here is an example of how it would look like:\ndatasets:\n    ...\n    roles:\n      assistant:\n        - gpt\n        - model\n      user:\n        - human\nIn the example above, all gpt and model values are converted to assistant. All human values are converted to user.\n\n\nHandling masking\nThe common use case for chat_template is for chat messages, therefore, it is common to mask all non-assistant messages. Assistant messages refer to the bot messages that you want the model to learn on.\nTo train on all assistant messages, you would set the following configs.\ndatasets:\n    ...\n    roles_to_train: [\"assistant\"]\n    train_on_eos: \"turn\"\nThe train_on_eos config means that it would mask all EOS tokens for turns that aren’t assistant-turns. The other options are: all and last to choose which EOS to train on.\nPerhaps, you want to train on assistant and narrator roles, you can simply add narrator to the list of roles_to_train. You would also need to add it to the mapping of roles above.\ndatasets:\n    ...\n    roles_to_train: [\"assistant\", \"narrator\"]\n    roles:\n      assistant:\n        - gpt\n        - model\n      user:\n        - human\n      narrator: [\"narrator\"]\n\n\n\n\n\n\nTip\n\n\n\nAs chat_templates may use hardcoded EOS/EOT tokens that are different from the tokenizer’s EOS, it is highly recommended to set them. For example, ChatML uses &lt;|im_end|&gt; to end turns.\nspecial_tokens:\n  eos_token: &lt;|im_end|&gt;\n\n\n\n\nApplying chat_template\nOnce all the above steps are completed, you could combine all these configs together to form a bespoke configuration for your custom dataset.\ndatasets:\n  - path: A.jsonl\n    type: chat_template\n\n    # step 1\n    chat_template: chatml\n\n    # step 2\n    field_messages: messages\n    message_property_mappings:\n      role: role\n      content: content\n\n    roles:\n      assistant:\n        - gpt\n        - model\n        - assistant\n      user:\n        - human\n        - user\n\n    # step 3\n    roles_to_train: [\"assistant\"]\n    train_on_eos: \"turn\"\n\nspecial_tokens:\n  eos_token: &lt;|im_end|&gt;\nIf this config were to be applied to the sample dataset above, the output would look as such (which can be retrieved via axolotl preprocess config.yaml --debug):\n&lt;|im_start|&gt;(-100, 128256) user(-100, 882)\n(-100, 198) Hi(-100, 13347) &lt;|im_end|&gt;(-100, 128257)\n(-100, 198) &lt;|im_start|&gt;(-100, 128256) assistant(-100, 78191)\n(-100, 198) How(4438, 4438)  can(649, 649)  I(358, 358)  help(1520, 1520)  you(499, 499) ?(30, 30) &lt;|im_end|&gt;(128257, 128257)\n(-100, 198) &lt;|im_start|&gt;(-100, 128256) user(-100, 882)\n(-100, 198) Can(-100, 6854)  you(-100, 499)  add(-100, 923)  (-100, 220) 3(-100, 18) +(-100, 10) 5(-100, 20) ?(-100, 30) &lt;|im_end|&gt;(-100, 128257)\n(-100, 198) &lt;|im_start|&gt;(-100, 128256) assistant(-100, 78191)\n(-100, 198) The(791, 791)  answer(4320, 4320)  is(374, 374)  (220, 220) 8(23, 23) .(13, 13) &lt;|im_end|&gt;(128257, 128257)\n(-100, 198)\nThe first number refers to the label, the second refers to the token_id. For example, -100 labels appear on non-assistant portions, meaning that they are masked during. For assistant portions, the label is the same as the token_id.\n\n\n\n\n\n\nNote\n\n\n\nIf during preprocess, there are a lot of warnings of Could not find content __ boundary, please check the FAQ section for chat_templates.\n\n\n\n\n\nReference\nPlease see docs here.\n\n\n\nInstruction Dataset\nInstruction datasets are used to train instruction-following models and comprise a prompt, containing an instruction, and a single response. In contrast to chat datasets which may be multi-turn, instruct datasets are typically single-turn.\nAn example is of a common format called Alpaca:\n{\"instruction\": \"...\", \"input\": \"...\", \"output\": \"...\"}\nUsing those keys, a prompt can be built based on it.\nBelow is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n{output}\nThis can be configured as such:\ndatasets:\n  - path: A.jsonl\n    type: alpaca\nAxolotl supports many kinds of instruction dataset. All of them can be found here (https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/inst_tune.html) with their respective type and sample row format.\nReference: Instruction Dataset Documentation.\n\nCustom Instruct Prompt Format\nDue to the myriad possibilities of instruction formats, Axolotl allows customizing your own instruction format without having to dive into the code directly.\nIn the example below, a sample row is used to output in mistral_v1 format.\n{\"input\": \"...\", \"output\": \"...\"}\ndatasets:\n  - path: repo\n    type:\n      system_prompt: \"\"\n\n      field_system:\n      field_instruction: input\n      field_input:\n      field_output: output\n\n      # multi-line example with input\n      format: |-\n        [INST] {instruction} {input} [/INST]\n\n      # single-line example without input\n      no_input_format: \"[INST] {instruction} [/INST]\"\nThe config sets that the field_instruction is actually named input, and the field_input is empty as we don’t have an input in this sample. Generally, instruction can be thought as the question to the model, and input as the additional information with output being the response. It is not necessary to have an input nor system. In the end, the most important part is to understand what format you want it to look like and how you can customize this to your use case.\nReference: Custom Instruct Prompt Format Documentation.",
     "crumbs": [
       "Dataset Formats"
     ]
diff --git a/sitemap.xml b/sitemap.xml
index 0b694452d..ffa75d659 100644
--- a/sitemap.xml
+++ b/sitemap.xml
@@ -2,158 +2,162 @@
 <urlset xmlns="http://www.sitemaps.org/schemas/sitemap/0.9">
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/examples/colab-notebooks/colab-axolotl-example.html</loc>
-    <lastmod>2025-03-05T13:58:47.518Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.716Z</lastmod>
   </url>
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/stepwise_supervised.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
   </url>
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/template_free.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
   </url>
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/tokenized.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
   </url>
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/nccl.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
   </url>
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/amd_hpc.html</loc>
-    <lastmod>2025-03-05T13:58:47.513Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/mac.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/custom_integrations.html</loc>
-    <lastmod>2025-03-05T13:58:47.513Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/dataset_preprocessing.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/fsdp_qlora.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/unsloth.html</loc>
-    <lastmod>2025-03-05T13:58:47.518Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/cli.html</loc>
-    <lastmod>2025-03-05T13:58:47.513Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/rlhf.html</loc>
-    <lastmod>2025-03-05T13:58:47.518Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/lora_optims.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/batch_vs_grad.html</loc>
-    <lastmod>2025-03-05T13:58:47.513Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/faq.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/multi-node.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/FAQS.html</loc>
-    <lastmod>2025-03-05T13:58:47.512Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html</loc>
-    <lastmod>2025-03-05T13:58:47.534Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/index.html</loc>
-    <lastmod>2025-03-05T13:58:47.530Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/src/axolotl/integrations/LICENSE.html</loc>
-    <lastmod>2025-03-05T13:58:47.534Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/TODO.html</loc>
-    <lastmod>2025-03-05T13:58:47.512Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/lr_groups.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/debugging.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/multipack.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/inference.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/getting-started.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/multimodal.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/input_output.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/reward_modelling.html</loc>
-    <lastmod>2025-03-05T13:58:47.518Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/torchao.html</loc>
-    <lastmod>2025-03-05T13:58:47.518Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/installation.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
-  </url>
-  <url>
-    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/multi-gpu.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.711Z</lastmod>
   </url>
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/config.html</loc>
-    <lastmod>2025-03-05T13:58:47.513Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.711Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/multi-gpu.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/installation.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/torchao.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/reward_modelling.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/input_output.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/multimodal.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/getting-started.html</loc>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/inference.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/multipack.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/debugging.html</loc>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/lr_groups.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/TODO.html</loc>
+    <lastmod>2025-03-05T15:02:20.710Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/src/axolotl/integrations/LICENSE.html</loc>
+    <lastmod>2025-03-05T15:02:20.730Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/index.html</loc>
+    <lastmod>2025-03-05T15:02:20.727Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html</loc>
+    <lastmod>2025-03-05T15:02:20.730Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/FAQS.html</loc>
+    <lastmod>2025-03-05T15:02:20.710Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/multi-node.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/faq.html</loc>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/batch_vs_grad.html</loc>
+    <lastmod>2025-03-05T15:02:20.711Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/lora_optims.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/rlhf.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/cli.html</loc>
+    <lastmod>2025-03-05T15:02:20.711Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/unsloth.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/fsdp_qlora.html</loc>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/dataset_preprocessing.html</loc>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/custom_integrations.html</loc>
+    <lastmod>2025-03-05T15:02:20.711Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/mac.html</loc>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
+  </url>
+  <url>
+    <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/docker.html</loc>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
   </url>
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/ray-integration.html</loc>
-    <lastmod>2025-03-05T13:58:47.517Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.715Z</lastmod>
   </url>
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/index.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
   </url>
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.711Z</lastmod>
   </url>
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/pretraining.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
   </url>
   <url>
     <loc>https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/inst_tune.html</loc>
-    <lastmod>2025-03-05T13:58:47.514Z</lastmod>
+    <lastmod>2025-03-05T15:02:20.712Z</lastmod>
   </url>
 </urlset>
diff --git a/src/axolotl/integrations/LICENSE.html b/src/axolotl/integrations/LICENSE.html
index 60c049b26..aafa36874 100644
--- a/src/axolotl/integrations/LICENSE.html
+++ b/src/axolotl/integrations/LICENSE.html
@@ -206,6 +206,12 @@ ul.task-list li input[type="checkbox"] {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../../../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../../../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>
diff --git a/src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html b/src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html
index 691938981..79a4f0959 100644
--- a/src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html
+++ b/src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html
@@ -206,6 +206,12 @@ ul.task-list li input[type="checkbox"] {
       <ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth1 show">  
           <li class="sidebar-item">
   <div class="sidebar-item-container"> 
+  <a href="../../../../docs/docker.html" class="sidebar-item-text sidebar-link">
+ <span class="menu-text">Docker</span></a>
+  </div>
+</li>
+          <li class="sidebar-item">
+  <div class="sidebar-item-container"> 
   <a href="../../../../docs/multi-gpu.html" class="sidebar-item-text sidebar-link">
  <span class="menu-text">Multi-GPU</span></a>
   </div>