Built site for gh-pages
This commit is contained in:
Quarto GHA Workflow Runner
257
docs/config.html
257
docs/config.html
@@ -75,7 +75,7 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
|
||||
<link href="../site_libs/quarto-html/quarto-syntax-highlighting-dark-2b3e328b71be8d25427581baeb23079b.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-653e373a27bf50c3d267316c2b2b59fb.min.css" rel="stylesheet" append-hash="true" id="quarto-bootstrap" data-mode="dark">
|
||||
<link href="../site_libs/bootstrap/bootstrap-ce762b396f898894284bb8eeee180359.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,
|
||||
@@ -1080,133 +1080,134 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin
|
||||
<span id="cb1-606"><a href="#cb1-606" aria-hidden="true" tabindex="-1"></a><span class="co"># - optimi_adamw</span></span>
|
||||
<span id="cb1-607"><a href="#cb1-607" aria-hidden="true" tabindex="-1"></a><span class="co"># - ao_adamw_8bit</span></span>
|
||||
<span id="cb1-608"><a href="#cb1-608" aria-hidden="true" tabindex="-1"></a><span class="co"># - ao_adamw_fp8</span></span>
|
||||
<span id="cb1-609"><a href="#cb1-609" aria-hidden="true" tabindex="-1"></a><span class="fu">optimizer</span><span class="kw">:</span></span>
|
||||
<span id="cb1-610"><a href="#cb1-610" aria-hidden="true" tabindex="-1"></a><span class="co"># Dictionary of arguments to pass to the optimizer</span></span>
|
||||
<span id="cb1-611"><a href="#cb1-611" aria-hidden="true" tabindex="-1"></a><span class="fu">optim_args</span><span class="kw">:</span></span>
|
||||
<span id="cb1-612"><a href="#cb1-612" aria-hidden="true" tabindex="-1"></a><span class="co"># For Galore Optimizers the following optim_args are available</span></span>
|
||||
<span id="cb1-613"><a href="#cb1-613" aria-hidden="true" tabindex="-1"></a><span class="co"># rank: # type: int</span></span>
|
||||
<span id="cb1-614"><a href="#cb1-614" aria-hidden="true" tabindex="-1"></a><span class="co"># update_proj_gap # type: int</span></span>
|
||||
<span id="cb1-615"><a href="#cb1-615" aria-hidden="true" tabindex="-1"></a><span class="co"># scale # type: float</span></span>
|
||||
<span id="cb1-616"><a href="#cb1-616" aria-hidden="true" tabindex="-1"></a><span class="co"># proj_type: # type: str, default = std</span></span>
|
||||
<span id="cb1-617"><a href="#cb1-617" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-618"><a href="#cb1-618" 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="cb1-619"><a href="#cb1-619" aria-hidden="true" tabindex="-1"></a><span class="fu">optim_target_modules</span><span class="kw">:</span></span>
|
||||
<span id="cb1-620"><a href="#cb1-620" aria-hidden="true" tabindex="-1"></a><span class="co"># - self_attn # for llama</span></span>
|
||||
<span id="cb1-621"><a href="#cb1-621" aria-hidden="true" tabindex="-1"></a><span class="co"># - mlp</span></span>
|
||||
<span id="cb1-622"><a href="#cb1-622" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-623"><a href="#cb1-623" aria-hidden="true" tabindex="-1"></a><span class="co"># Specify weight decay</span></span>
|
||||
<span id="cb1-624"><a href="#cb1-624" aria-hidden="true" tabindex="-1"></a><span class="fu">weight_decay</span><span class="kw">:</span></span>
|
||||
<span id="cb1-625"><a href="#cb1-625" aria-hidden="true" tabindex="-1"></a><span class="co"># adamw hyperparams</span></span>
|
||||
<span id="cb1-626"><a href="#cb1-626" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_beta1</span><span class="kw">:</span></span>
|
||||
<span id="cb1-627"><a href="#cb1-627" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_beta2</span><span class="kw">:</span></span>
|
||||
<span id="cb1-628"><a href="#cb1-628" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_epsilon</span><span class="kw">:</span></span>
|
||||
<span id="cb1-629"><a href="#cb1-629" aria-hidden="true" tabindex="-1"></a><span class="co"># Gradient clipping max norm</span></span>
|
||||
<span id="cb1-630"><a href="#cb1-630" aria-hidden="true" tabindex="-1"></a><span class="fu">max_grad_norm</span><span class="kw">:</span></span>
|
||||
<span id="cb1-631"><a href="#cb1-631" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-632"><a href="#cb1-632" aria-hidden="true" tabindex="-1"></a><span class="co"># Augmentation techniques</span></span>
|
||||
<span id="cb1-633"><a href="#cb1-633" 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="cb1-634"><a href="#cb1-634" aria-hidden="true" tabindex="-1"></a><span class="co"># currently only supported on Llama and Mistral</span></span>
|
||||
<span id="cb1-635"><a href="#cb1-635" aria-hidden="true" tabindex="-1"></a><span class="fu">neftune_noise_alpha</span><span class="kw">:</span></span>
|
||||
<span id="cb1-636"><a href="#cb1-636" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-637"><a href="#cb1-637" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to bettertransformers</span></span>
|
||||
<span id="cb1-638"><a href="#cb1-638" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_optimum</span><span class="kw">:</span></span>
|
||||
<span id="cb1-639"><a href="#cb1-639" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-640"><a href="#cb1-640" aria-hidden="true" tabindex="-1"></a><span class="co"># Note: Only one of the following attention patches can be used at a time.</span></span>
|
||||
<span id="cb1-641"><a href="#cb1-641" aria-hidden="true" tabindex="-1"></a><span class="co"># For example, if you set `xformers_attention` to `true`, do not set `flash_attention` to `true`.</span></span>
|
||||
<span id="cb1-642"><a href="#cb1-642" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-643"><a href="#cb1-643" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use xformers attention patch https://github.com/facebookresearch/xformers:</span></span>
|
||||
<span id="cb1-644"><a href="#cb1-644" aria-hidden="true" tabindex="-1"></a><span class="fu">xformers_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-645"><a href="#cb1-645" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:</span></span>
|
||||
<span id="cb1-646"><a href="#cb1-646" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-647"><a href="#cb1-647" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_cross_entropy</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to use flash-attention cross entropy implementation - advanced use only</span></span>
|
||||
<span id="cb1-648"><a href="#cb1-648" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_rms_norm</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to use flash-attention rms norm implementation - advanced use only</span></span>
|
||||
<span id="cb1-649"><a href="#cb1-649" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_qkv</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to fuse QKV into a single operation</span></span>
|
||||
<span id="cb1-650"><a href="#cb1-650" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_mlp</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to fuse part of the MLP into a single operation</span></span>
|
||||
<span id="cb1-651"><a href="#cb1-651" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use scaled-dot-product attention</span></span>
|
||||
<span id="cb1-652"><a href="#cb1-652" 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="cb1-653"><a href="#cb1-653" aria-hidden="true" tabindex="-1"></a><span class="fu">sdp_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-654"><a href="#cb1-654" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf</span></span>
|
||||
<span id="cb1-655"><a href="#cb1-655" aria-hidden="true" tabindex="-1"></a><span class="fu">s2_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-656"><a href="#cb1-656" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-657"><a href="#cb1-657" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use low_cpu_mem_usage</span></span>
|
||||
<span id="cb1-658"><a href="#cb1-658" aria-hidden="true" tabindex="-1"></a><span class="fu">low_cpu_mem_usage</span><span class="kw">:</span></span>
|
||||
<span id="cb1-659"><a href="#cb1-659" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[str]. Resume from a specific checkpoint dir</span></span>
|
||||
<span id="cb1-660"><a href="#cb1-660" aria-hidden="true" tabindex="-1"></a><span class="fu">resume_from_checkpoint</span><span class="kw">:</span></span>
|
||||
<span id="cb1-661"><a href="#cb1-661" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. If resume_from_checkpoint isn't set and you simply want it to start where it left off.</span></span>
|
||||
<span id="cb1-662"><a href="#cb1-662" aria-hidden="true" tabindex="-1"></a><span class="co"># Be careful with this being turned on between different models.</span></span>
|
||||
<span id="cb1-663"><a href="#cb1-663" 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="cb1-664"><a href="#cb1-664" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-665"><a href="#cb1-665" aria-hidden="true" tabindex="-1"></a><span class="co">## Multimodal section</span></span>
|
||||
<span id="cb1-666"><a href="#cb1-666" aria-hidden="true" tabindex="-1"></a><span class="co"># int | tuple[int, int] | None . Size to resize images to, width x height.</span></span>
|
||||
<span id="cb1-667"><a href="#cb1-667" aria-hidden="true" tabindex="-1"></a><span class="co"># Will read from model/processor config if not set.</span></span>
|
||||
<span id="cb1-668"><a href="#cb1-668" aria-hidden="true" tabindex="-1"></a><span class="fu">image_size</span><span class="kw">:</span></span>
|
||||
<span id="cb1-669"><a href="#cb1-669" aria-hidden="true" tabindex="-1"></a><span class="co"># str. Algorithm to use for image resizing. "bilinear", "bicubic", "lanczos". Default is "bilinear".</span></span>
|
||||
<span id="cb1-670"><a href="#cb1-670" aria-hidden="true" tabindex="-1"></a><span class="fu">image_resize_algorithm</span><span class="kw">:</span><span class="at"> </span><span class="st">'bilinear'</span></span>
|
||||
<span id="cb1-671"><a href="#cb1-671" aria-hidden="true" tabindex="-1"></a><span class="co">## End of multimodal section</span></span>
|
||||
<span id="cb1-672"><a href="#cb1-672" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-673"><a href="#cb1-673" 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="cb1-674"><a href="#cb1-674" aria-hidden="true" tabindex="-1"></a><span class="fu">local_rank</span><span class="kw">:</span></span>
|
||||
<span id="cb1-675"><a href="#cb1-675" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-676"><a href="#cb1-676" aria-hidden="true" tabindex="-1"></a><span class="co"># Add or change special tokens.</span></span>
|
||||
<span id="cb1-677"><a href="#cb1-677" 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="cb1-678"><a href="#cb1-678" aria-hidden="true" tabindex="-1"></a><span class="fu">special_tokens</span><span class="kw">:</span></span>
|
||||
<span id="cb1-679"><a href="#cb1-679" aria-hidden="true" tabindex="-1"></a><span class="co"> # bos_token: "<s>"</span></span>
|
||||
<span id="cb1-680"><a href="#cb1-680" aria-hidden="true" tabindex="-1"></a><span class="co"> # eos_token: "</s>"</span></span>
|
||||
<span id="cb1-681"><a href="#cb1-681" aria-hidden="true" tabindex="-1"></a><span class="co"> # unk_token: "<unk>"</span></span>
|
||||
<span id="cb1-682"><a href="#cb1-682" aria-hidden="true" tabindex="-1"></a><span class="co"> # pad_token: "[PAD]"</span></span>
|
||||
<span id="cb1-683"><a href="#cb1-683" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-684"><a href="#cb1-684" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[list[str]]. Add extra tokens to the tokenizer.</span></span>
|
||||
<span id="cb1-685"><a href="#cb1-685" aria-hidden="true" tabindex="-1"></a><span class="fu">tokens</span><span class="kw">:</span></span>
|
||||
<span id="cb1-686"><a href="#cb1-686" aria-hidden="true" tabindex="-1"></a><span class="co"> # - "<|startoftext|>"</span></span>
|
||||
<span id="cb1-687"><a href="#cb1-687" aria-hidden="true" tabindex="-1"></a><span class="co"> # - "<|endoftext|>"</span></span>
|
||||
<span id="cb1-688"><a href="#cb1-688" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-689"><a href="#cb1-689" aria-hidden="true" tabindex="-1"></a><span class="co"># Mapping token_id to new_token_string to override reserved added_tokens in the tokenizer.</span></span>
|
||||
<span id="cb1-690"><a href="#cb1-690" aria-hidden="true" tabindex="-1"></a><span class="co"># Only works for tokens that are not part of the base vocab (aka are added_tokens).</span></span>
|
||||
<span id="cb1-691"><a href="#cb1-691" aria-hidden="true" tabindex="-1"></a><span class="co"># Can be checked if they exist in tokenizer.json added_tokens.</span></span>
|
||||
<span id="cb1-692"><a href="#cb1-692" aria-hidden="true" tabindex="-1"></a><span class="fu">added_tokens_overrides</span><span class="kw">:</span><span class="co"> # Dict[int, str]</span></span>
|
||||
<span id="cb1-693"><a href="#cb1-693" aria-hidden="true" tabindex="-1"></a><span class="co"># 128041: "<|im_start|>"</span></span>
|
||||
<span id="cb1-694"><a href="#cb1-694" aria-hidden="true" tabindex="-1"></a><span class="co"># 128042: "<|im_end|>"</span></span>
|
||||
<span id="cb1-695"><a href="#cb1-695" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-696"><a href="#cb1-696" aria-hidden="true" tabindex="-1"></a><span class="co"># FSDP</span></span>
|
||||
<span id="cb1-697"><a href="#cb1-697" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp</span><span class="kw">:</span></span>
|
||||
<span id="cb1-698"><a href="#cb1-698" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp_config</span><span class="kw">:</span></span>
|
||||
<span id="cb1-699"><a href="#cb1-699" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-700"><a href="#cb1-700" aria-hidden="true" tabindex="-1"></a><span class="co"># Deepspeed config path. e.g., deepspeed_configs/zero3.json</span></span>
|
||||
<span id="cb1-701"><a href="#cb1-701" aria-hidden="true" tabindex="-1"></a><span class="fu">deepspeed</span><span class="kw">:</span></span>
|
||||
<span id="cb1-702"><a href="#cb1-702" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-703"><a href="#cb1-703" aria-hidden="true" tabindex="-1"></a><span class="co"># Advanced DDP Arguments</span></span>
|
||||
<span id="cb1-704"><a href="#cb1-704" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_timeout</span><span class="kw">:</span></span>
|
||||
<span id="cb1-705"><a href="#cb1-705" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_bucket_cap_mb</span><span class="kw">:</span></span>
|
||||
<span id="cb1-706"><a href="#cb1-706" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_broadcast_buffers</span><span class="kw">:</span></span>
|
||||
<span id="cb1-707"><a href="#cb1-707" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-708"><a href="#cb1-708" aria-hidden="true" tabindex="-1"></a><span class="co"># Sequence parallelism</span></span>
|
||||
<span id="cb1-709"><a href="#cb1-709" aria-hidden="true" tabindex="-1"></a><span class="co"># Set to a divisor of the number of GPUs available to split sequences into chunks of equal size.</span></span>
|
||||
<span id="cb1-710"><a href="#cb1-710" aria-hidden="true" tabindex="-1"></a><span class="co"># Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM.</span></span>
|
||||
<span id="cb1-711"><a href="#cb1-711" aria-hidden="true" tabindex="-1"></a><span class="co"># E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized</span></span>
|
||||
<span id="cb1-712"><a href="#cb1-712" aria-hidden="true" tabindex="-1"></a><span class="co"># subsequences, or set to 4 to split into four equal-sized subsequences.</span></span>
|
||||
<span id="cb1-713"><a href="#cb1-713" aria-hidden="true" tabindex="-1"></a><span class="co"># See https://docs.axolotl.ai/docs/sequence_parallelism.html for more details.</span></span>
|
||||
<span id="cb1-714"><a href="#cb1-714" aria-hidden="true" tabindex="-1"></a><span class="fu">sequence_parallel_degree</span><span class="kw">:</span></span>
|
||||
<span id="cb1-715"><a href="#cb1-715" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional; strides across the key dimension. Larger values use more memory but should make training faster.</span></span>
|
||||
<span id="cb1-716"><a href="#cb1-716" aria-hidden="true" tabindex="-1"></a><span class="co"># Must evenly divide the number of KV heads in your model.</span></span>
|
||||
<span id="cb1-717"><a href="#cb1-717" aria-hidden="true" tabindex="-1"></a><span class="fu">heads_k_stride</span><span class="kw">:</span><span class="at"> </span><span class="dv">1</span></span>
|
||||
<span id="cb1-718"><a href="#cb1-718" aria-hidden="true" tabindex="-1"></a><span class="co"># One of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to "varlen_llama3"</span></span>
|
||||
<span id="cb1-719"><a href="#cb1-719" aria-hidden="true" tabindex="-1"></a><span class="co"># in the sample packing case, and "batch_ring" in the non-sample packing case.</span></span>
|
||||
<span id="cb1-720"><a href="#cb1-720" aria-hidden="true" tabindex="-1"></a><span class="fu">ring_attn_func</span><span class="kw">:</span></span>
|
||||
<span id="cb1-721"><a href="#cb1-721" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-722"><a href="#cb1-722" aria-hidden="true" tabindex="-1"></a><span class="co"># Path to torch distx for optim 'adamw_anyprecision'</span></span>
|
||||
<span id="cb1-723"><a href="#cb1-723" aria-hidden="true" tabindex="-1"></a><span class="fu">torchdistx_path</span><span class="kw">:</span></span>
|
||||
<span id="cb1-724"><a href="#cb1-724" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-725"><a href="#cb1-725" 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="cb1-726"><a href="#cb1-726" aria-hidden="true" tabindex="-1"></a><span class="fu">pretraining_dataset</span><span class="kw">:</span></span>
|
||||
<span id="cb1-727"><a href="#cb1-727" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-728"><a href="#cb1-728" aria-hidden="true" tabindex="-1"></a><span class="co"># Debug mode</span></span>
|
||||
<span id="cb1-729"><a href="#cb1-729" aria-hidden="true" tabindex="-1"></a><span class="fu">debug</span><span class="kw">:</span></span>
|
||||
<span id="cb1-730"><a href="#cb1-730" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-731"><a href="#cb1-731" aria-hidden="true" tabindex="-1"></a><span class="co"># Seed</span></span>
|
||||
<span id="cb1-732"><a href="#cb1-732" aria-hidden="true" tabindex="-1"></a><span class="fu">seed</span><span class="kw">:</span></span>
|
||||
<span id="cb1-733"><a href="#cb1-733" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-734"><a href="#cb1-734" aria-hidden="true" tabindex="-1"></a><span class="co"># Allow overwrite yml config using from cli</span></span>
|
||||
<span id="cb1-735"><a href="#cb1-735" 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="cb1-609"><a href="#cb1-609" aria-hidden="true" tabindex="-1"></a><span class="co"># - came_pytorch</span></span>
|
||||
<span id="cb1-610"><a href="#cb1-610" aria-hidden="true" tabindex="-1"></a><span class="fu">optimizer</span><span class="kw">:</span></span>
|
||||
<span id="cb1-611"><a href="#cb1-611" aria-hidden="true" tabindex="-1"></a><span class="co"># Dictionary of arguments to pass to the optimizer</span></span>
|
||||
<span id="cb1-612"><a href="#cb1-612" aria-hidden="true" tabindex="-1"></a><span class="fu">optim_args</span><span class="kw">:</span></span>
|
||||
<span id="cb1-613"><a href="#cb1-613" aria-hidden="true" tabindex="-1"></a><span class="co"># For Galore Optimizers the following optim_args are available</span></span>
|
||||
<span id="cb1-614"><a href="#cb1-614" aria-hidden="true" tabindex="-1"></a><span class="co"># rank: # type: int</span></span>
|
||||
<span id="cb1-615"><a href="#cb1-615" aria-hidden="true" tabindex="-1"></a><span class="co"># update_proj_gap # type: int</span></span>
|
||||
<span id="cb1-616"><a href="#cb1-616" aria-hidden="true" tabindex="-1"></a><span class="co"># scale # type: float</span></span>
|
||||
<span id="cb1-617"><a href="#cb1-617" aria-hidden="true" tabindex="-1"></a><span class="co"># proj_type: # type: str, default = std</span></span>
|
||||
<span id="cb1-618"><a href="#cb1-618" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-619"><a href="#cb1-619" 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="cb1-620"><a href="#cb1-620" aria-hidden="true" tabindex="-1"></a><span class="fu">optim_target_modules</span><span class="kw">:</span></span>
|
||||
<span id="cb1-621"><a href="#cb1-621" aria-hidden="true" tabindex="-1"></a><span class="co"># - self_attn # for llama</span></span>
|
||||
<span id="cb1-622"><a href="#cb1-622" aria-hidden="true" tabindex="-1"></a><span class="co"># - mlp</span></span>
|
||||
<span id="cb1-623"><a href="#cb1-623" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-624"><a href="#cb1-624" aria-hidden="true" tabindex="-1"></a><span class="co"># Specify weight decay</span></span>
|
||||
<span id="cb1-625"><a href="#cb1-625" aria-hidden="true" tabindex="-1"></a><span class="fu">weight_decay</span><span class="kw">:</span></span>
|
||||
<span id="cb1-626"><a href="#cb1-626" aria-hidden="true" tabindex="-1"></a><span class="co"># adamw hyperparams</span></span>
|
||||
<span id="cb1-627"><a href="#cb1-627" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_beta1</span><span class="kw">:</span></span>
|
||||
<span id="cb1-628"><a href="#cb1-628" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_beta2</span><span class="kw">:</span></span>
|
||||
<span id="cb1-629"><a href="#cb1-629" aria-hidden="true" tabindex="-1"></a><span class="fu">adam_epsilon</span><span class="kw">:</span></span>
|
||||
<span id="cb1-630"><a href="#cb1-630" aria-hidden="true" tabindex="-1"></a><span class="co"># Gradient clipping max norm</span></span>
|
||||
<span id="cb1-631"><a href="#cb1-631" aria-hidden="true" tabindex="-1"></a><span class="fu">max_grad_norm</span><span class="kw">:</span></span>
|
||||
<span id="cb1-632"><a href="#cb1-632" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-633"><a href="#cb1-633" aria-hidden="true" tabindex="-1"></a><span class="co"># Augmentation techniques</span></span>
|
||||
<span id="cb1-634"><a href="#cb1-634" 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="cb1-635"><a href="#cb1-635" aria-hidden="true" tabindex="-1"></a><span class="co"># currently only supported on Llama and Mistral</span></span>
|
||||
<span id="cb1-636"><a href="#cb1-636" aria-hidden="true" tabindex="-1"></a><span class="fu">neftune_noise_alpha</span><span class="kw">:</span></span>
|
||||
<span id="cb1-637"><a href="#cb1-637" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-638"><a href="#cb1-638" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to bettertransformers</span></span>
|
||||
<span id="cb1-639"><a href="#cb1-639" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_optimum</span><span class="kw">:</span></span>
|
||||
<span id="cb1-640"><a href="#cb1-640" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-641"><a href="#cb1-641" aria-hidden="true" tabindex="-1"></a><span class="co"># Note: Only one of the following attention patches can be used at a time.</span></span>
|
||||
<span id="cb1-642"><a href="#cb1-642" aria-hidden="true" tabindex="-1"></a><span class="co"># For example, if you set `xformers_attention` to `true`, do not set `flash_attention` to `true`.</span></span>
|
||||
<span id="cb1-643"><a href="#cb1-643" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-644"><a href="#cb1-644" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use xformers attention patch https://github.com/facebookresearch/xformers:</span></span>
|
||||
<span id="cb1-645"><a href="#cb1-645" aria-hidden="true" tabindex="-1"></a><span class="fu">xformers_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-646"><a href="#cb1-646" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:</span></span>
|
||||
<span id="cb1-647"><a href="#cb1-647" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-648"><a href="#cb1-648" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_cross_entropy</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to use flash-attention cross entropy implementation - advanced use only</span></span>
|
||||
<span id="cb1-649"><a href="#cb1-649" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_rms_norm</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to use flash-attention rms norm implementation - advanced use only</span></span>
|
||||
<span id="cb1-650"><a href="#cb1-650" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_qkv</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to fuse QKV into a single operation</span></span>
|
||||
<span id="cb1-651"><a href="#cb1-651" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attn_fuse_mlp</span><span class="kw">:</span><span class="co"> # Optional[bool]. Whether to fuse part of the MLP into a single operation</span></span>
|
||||
<span id="cb1-652"><a href="#cb1-652" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use scaled-dot-product attention</span></span>
|
||||
<span id="cb1-653"><a href="#cb1-653" 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="cb1-654"><a href="#cb1-654" aria-hidden="true" tabindex="-1"></a><span class="fu">sdp_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-655"><a href="#cb1-655" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf</span></span>
|
||||
<span id="cb1-656"><a href="#cb1-656" aria-hidden="true" tabindex="-1"></a><span class="fu">s2_attention</span><span class="kw">:</span></span>
|
||||
<span id="cb1-657"><a href="#cb1-657" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-658"><a href="#cb1-658" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. Whether to use low_cpu_mem_usage</span></span>
|
||||
<span id="cb1-659"><a href="#cb1-659" aria-hidden="true" tabindex="-1"></a><span class="fu">low_cpu_mem_usage</span><span class="kw">:</span></span>
|
||||
<span id="cb1-660"><a href="#cb1-660" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[str]. Resume from a specific checkpoint dir</span></span>
|
||||
<span id="cb1-661"><a href="#cb1-661" aria-hidden="true" tabindex="-1"></a><span class="fu">resume_from_checkpoint</span><span class="kw">:</span></span>
|
||||
<span id="cb1-662"><a href="#cb1-662" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[bool]. If resume_from_checkpoint isn't set and you simply want it to start where it left off.</span></span>
|
||||
<span id="cb1-663"><a href="#cb1-663" aria-hidden="true" tabindex="-1"></a><span class="co"># Be careful with this being turned on between different models.</span></span>
|
||||
<span id="cb1-664"><a href="#cb1-664" 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="cb1-665"><a href="#cb1-665" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-666"><a href="#cb1-666" aria-hidden="true" tabindex="-1"></a><span class="co">## Multimodal section</span></span>
|
||||
<span id="cb1-667"><a href="#cb1-667" aria-hidden="true" tabindex="-1"></a><span class="co"># int | tuple[int, int] | None . Size to resize images to, width x height.</span></span>
|
||||
<span id="cb1-668"><a href="#cb1-668" aria-hidden="true" tabindex="-1"></a><span class="co"># Will read from model/processor config if not set.</span></span>
|
||||
<span id="cb1-669"><a href="#cb1-669" aria-hidden="true" tabindex="-1"></a><span class="fu">image_size</span><span class="kw">:</span></span>
|
||||
<span id="cb1-670"><a href="#cb1-670" aria-hidden="true" tabindex="-1"></a><span class="co"># str. Algorithm to use for image resizing. "bilinear", "bicubic", "lanczos". Default is "bilinear".</span></span>
|
||||
<span id="cb1-671"><a href="#cb1-671" aria-hidden="true" tabindex="-1"></a><span class="fu">image_resize_algorithm</span><span class="kw">:</span><span class="at"> </span><span class="st">'bilinear'</span></span>
|
||||
<span id="cb1-672"><a href="#cb1-672" aria-hidden="true" tabindex="-1"></a><span class="co">## End of multimodal section</span></span>
|
||||
<span id="cb1-673"><a href="#cb1-673" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-674"><a href="#cb1-674" 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="cb1-675"><a href="#cb1-675" aria-hidden="true" tabindex="-1"></a><span class="fu">local_rank</span><span class="kw">:</span></span>
|
||||
<span id="cb1-676"><a href="#cb1-676" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-677"><a href="#cb1-677" aria-hidden="true" tabindex="-1"></a><span class="co"># Add or change special tokens.</span></span>
|
||||
<span id="cb1-678"><a href="#cb1-678" 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="cb1-679"><a href="#cb1-679" aria-hidden="true" tabindex="-1"></a><span class="fu">special_tokens</span><span class="kw">:</span></span>
|
||||
<span id="cb1-680"><a href="#cb1-680" aria-hidden="true" tabindex="-1"></a><span class="co"> # bos_token: "<s>"</span></span>
|
||||
<span id="cb1-681"><a href="#cb1-681" aria-hidden="true" tabindex="-1"></a><span class="co"> # eos_token: "</s>"</span></span>
|
||||
<span id="cb1-682"><a href="#cb1-682" aria-hidden="true" tabindex="-1"></a><span class="co"> # unk_token: "<unk>"</span></span>
|
||||
<span id="cb1-683"><a href="#cb1-683" aria-hidden="true" tabindex="-1"></a><span class="co"> # pad_token: "[PAD]"</span></span>
|
||||
<span id="cb1-684"><a href="#cb1-684" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-685"><a href="#cb1-685" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional[list[str]]. Add extra tokens to the tokenizer.</span></span>
|
||||
<span id="cb1-686"><a href="#cb1-686" aria-hidden="true" tabindex="-1"></a><span class="fu">tokens</span><span class="kw">:</span></span>
|
||||
<span id="cb1-687"><a href="#cb1-687" aria-hidden="true" tabindex="-1"></a><span class="co"> # - "<|startoftext|>"</span></span>
|
||||
<span id="cb1-688"><a href="#cb1-688" aria-hidden="true" tabindex="-1"></a><span class="co"> # - "<|endoftext|>"</span></span>
|
||||
<span id="cb1-689"><a href="#cb1-689" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-690"><a href="#cb1-690" aria-hidden="true" tabindex="-1"></a><span class="co"># Mapping token_id to new_token_string to override reserved added_tokens in the tokenizer.</span></span>
|
||||
<span id="cb1-691"><a href="#cb1-691" aria-hidden="true" tabindex="-1"></a><span class="co"># Only works for tokens that are not part of the base vocab (aka are added_tokens).</span></span>
|
||||
<span id="cb1-692"><a href="#cb1-692" aria-hidden="true" tabindex="-1"></a><span class="co"># Can be checked if they exist in tokenizer.json added_tokens.</span></span>
|
||||
<span id="cb1-693"><a href="#cb1-693" aria-hidden="true" tabindex="-1"></a><span class="fu">added_tokens_overrides</span><span class="kw">:</span><span class="co"> # Dict[int, str]</span></span>
|
||||
<span id="cb1-694"><a href="#cb1-694" aria-hidden="true" tabindex="-1"></a><span class="co"># 128041: "<|im_start|>"</span></span>
|
||||
<span id="cb1-695"><a href="#cb1-695" aria-hidden="true" tabindex="-1"></a><span class="co"># 128042: "<|im_end|>"</span></span>
|
||||
<span id="cb1-696"><a href="#cb1-696" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-697"><a href="#cb1-697" aria-hidden="true" tabindex="-1"></a><span class="co"># FSDP</span></span>
|
||||
<span id="cb1-698"><a href="#cb1-698" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp</span><span class="kw">:</span></span>
|
||||
<span id="cb1-699"><a href="#cb1-699" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp_config</span><span class="kw">:</span></span>
|
||||
<span id="cb1-700"><a href="#cb1-700" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-701"><a href="#cb1-701" aria-hidden="true" tabindex="-1"></a><span class="co"># Deepspeed config path. e.g., deepspeed_configs/zero3.json</span></span>
|
||||
<span id="cb1-702"><a href="#cb1-702" aria-hidden="true" tabindex="-1"></a><span class="fu">deepspeed</span><span class="kw">:</span></span>
|
||||
<span id="cb1-703"><a href="#cb1-703" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-704"><a href="#cb1-704" aria-hidden="true" tabindex="-1"></a><span class="co"># Advanced DDP Arguments</span></span>
|
||||
<span id="cb1-705"><a href="#cb1-705" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_timeout</span><span class="kw">:</span></span>
|
||||
<span id="cb1-706"><a href="#cb1-706" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_bucket_cap_mb</span><span class="kw">:</span></span>
|
||||
<span id="cb1-707"><a href="#cb1-707" aria-hidden="true" tabindex="-1"></a><span class="fu">ddp_broadcast_buffers</span><span class="kw">:</span></span>
|
||||
<span id="cb1-708"><a href="#cb1-708" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-709"><a href="#cb1-709" aria-hidden="true" tabindex="-1"></a><span class="co"># Sequence parallelism</span></span>
|
||||
<span id="cb1-710"><a href="#cb1-710" aria-hidden="true" tabindex="-1"></a><span class="co"># Set to a divisor of the number of GPUs available to split sequences into chunks of equal size.</span></span>
|
||||
<span id="cb1-711"><a href="#cb1-711" aria-hidden="true" tabindex="-1"></a><span class="co"># Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM.</span></span>
|
||||
<span id="cb1-712"><a href="#cb1-712" aria-hidden="true" tabindex="-1"></a><span class="co"># E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized</span></span>
|
||||
<span id="cb1-713"><a href="#cb1-713" aria-hidden="true" tabindex="-1"></a><span class="co"># subsequences, or set to 4 to split into four equal-sized subsequences.</span></span>
|
||||
<span id="cb1-714"><a href="#cb1-714" aria-hidden="true" tabindex="-1"></a><span class="co"># See https://docs.axolotl.ai/docs/sequence_parallelism.html for more details.</span></span>
|
||||
<span id="cb1-715"><a href="#cb1-715" aria-hidden="true" tabindex="-1"></a><span class="fu">sequence_parallel_degree</span><span class="kw">:</span></span>
|
||||
<span id="cb1-716"><a href="#cb1-716" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional; strides across the key dimension. Larger values use more memory but should make training faster.</span></span>
|
||||
<span id="cb1-717"><a href="#cb1-717" aria-hidden="true" tabindex="-1"></a><span class="co"># Must evenly divide the number of KV heads in your model.</span></span>
|
||||
<span id="cb1-718"><a href="#cb1-718" aria-hidden="true" tabindex="-1"></a><span class="fu">heads_k_stride</span><span class="kw">:</span><span class="at"> </span><span class="dv">1</span></span>
|
||||
<span id="cb1-719"><a href="#cb1-719" aria-hidden="true" tabindex="-1"></a><span class="co"># One of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to "varlen_llama3"</span></span>
|
||||
<span id="cb1-720"><a href="#cb1-720" aria-hidden="true" tabindex="-1"></a><span class="co"># in the sample packing case, and "batch_ring" in the non-sample packing case.</span></span>
|
||||
<span id="cb1-721"><a href="#cb1-721" aria-hidden="true" tabindex="-1"></a><span class="fu">ring_attn_func</span><span class="kw">:</span></span>
|
||||
<span id="cb1-722"><a href="#cb1-722" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-723"><a href="#cb1-723" aria-hidden="true" tabindex="-1"></a><span class="co"># Path to torch distx for optim 'adamw_anyprecision'</span></span>
|
||||
<span id="cb1-724"><a href="#cb1-724" aria-hidden="true" tabindex="-1"></a><span class="fu">torchdistx_path</span><span class="kw">:</span></span>
|
||||
<span id="cb1-725"><a href="#cb1-725" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-726"><a href="#cb1-726" 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="cb1-727"><a href="#cb1-727" aria-hidden="true" tabindex="-1"></a><span class="fu">pretraining_dataset</span><span class="kw">:</span></span>
|
||||
<span id="cb1-728"><a href="#cb1-728" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-729"><a href="#cb1-729" aria-hidden="true" tabindex="-1"></a><span class="co"># Debug mode</span></span>
|
||||
<span id="cb1-730"><a href="#cb1-730" aria-hidden="true" tabindex="-1"></a><span class="fu">debug</span><span class="kw">:</span></span>
|
||||
<span id="cb1-731"><a href="#cb1-731" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-732"><a href="#cb1-732" aria-hidden="true" tabindex="-1"></a><span class="co"># Seed</span></span>
|
||||
<span id="cb1-733"><a href="#cb1-733" aria-hidden="true" tabindex="-1"></a><span class="fu">seed</span><span class="kw">:</span></span>
|
||||
<span id="cb1-734"><a href="#cb1-734" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb1-735"><a href="#cb1-735" aria-hidden="true" tabindex="-1"></a><span class="co"># Allow overwrite yml config using from cli</span></span>
|
||||
<span id="cb1-736"><a href="#cb1-736" 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>
|
||||
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user