Built site for gh-pages
This commit is contained in:
Quarto GHA Workflow Runner
@@ -572,15 +572,7 @@ Tip
|
||||
<a href="https://github.com/zhuzilin/ring-flash-attention">ring-flash-attention</a> project. This
|
||||
allows one to split up sequences across GPUs, which is useful in the event that a
|
||||
single sequence causes OOM errors during model training.</p>
|
||||
<p>First, install <code>ring-flash-attn</code>, recommended via <code>pip install axolotl[ring-flash-attn]</code>,
|
||||
or from source with <code>pip install .[ring-flash-attn]</code>.</p>
|
||||
<p>Your Axolotl YAML config should contain the following lines:</p>
|
||||
<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">sequence_parallel_degree</span><span class="kw">:</span><span class="at"> </span><span class="dv">4</span><span class="co"> # Split each sequence into 4 parts, one per GPU</span></span>
|
||||
<span id="cb4-2"><a href="#cb4-2" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attention</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span><span class="co"> # Required with sequence parallelism</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"># Optional; strides across the key dimension. Larger values use more memory but will make training faster.</span></span>
|
||||
<span id="cb4-5"><a href="#cb4-5" 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></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
|
||||
<p>See our <a href="../docs/sequence_parallelism.html">dedicated guide</a> for more details.</p>
|
||||
<p>See our <a href="../docs/sequence_parallelism.html">dedicated guide</a> for more information.</p>
|
||||
<section id="sec-fsdp-qlora" class="level3" data-number="4.1">
|
||||
<h3 data-number="4.1" class="anchored" data-anchor-id="sec-fsdp-qlora"><span class="header-section-number">4.1</span> FSDP + QLoRA</h3>
|
||||
<p>For combining FSDP with QLoRA, see our <a href="../docs/fsdp_qlora.html">dedicated guide</a>.</p>
|
||||
@@ -1080,146 +1072,133 @@ or from source with <code>pip install .[ring-flash-attn]</code>.</p>
|
||||
}
|
||||
});
|
||||
</script><div class="modal fade" id="quarto-embedded-source-code-modal" tabindex="-1" aria-labelledby="quarto-embedded-source-code-modal-label" aria-hidden="true"><div class="modal-dialog modal-dialog-scrollable"><div class="modal-content"><div class="modal-header"><h5 class="modal-title" id="quarto-embedded-source-code-modal-label">Source Code</h5><button class="btn-close" data-bs-dismiss="modal"></button></div><div class="modal-body"><div class="">
|
||||
<div class="sourceCode" id="cb5" data-shortcodes="false"><pre class="sourceCode markdown code-with-copy"><code class="sourceCode markdown"><span id="cb5-1"><a href="#cb5-1" aria-hidden="true" tabindex="-1"></a><span class="co">---</span></span>
|
||||
<span id="cb5-2"><a href="#cb5-2" aria-hidden="true" tabindex="-1"></a><span class="an">title:</span><span class="co"> "Multi-GPU"</span></span>
|
||||
<span id="cb5-3"><a href="#cb5-3" aria-hidden="true" tabindex="-1"></a><span class="an">format:</span></span>
|
||||
<span id="cb5-4"><a href="#cb5-4" aria-hidden="true" tabindex="-1"></a><span class="co"> html:</span></span>
|
||||
<span id="cb5-5"><a href="#cb5-5" aria-hidden="true" tabindex="-1"></a><span class="co"> toc: true</span></span>
|
||||
<span id="cb5-6"><a href="#cb5-6" aria-hidden="true" tabindex="-1"></a><span class="co"> toc-depth: 3</span></span>
|
||||
<span id="cb5-7"><a href="#cb5-7" aria-hidden="true" tabindex="-1"></a><span class="co"> number-sections: true</span></span>
|
||||
<span id="cb5-8"><a href="#cb5-8" aria-hidden="true" tabindex="-1"></a><span class="co"> code-tools: true</span></span>
|
||||
<span id="cb5-9"><a href="#cb5-9" aria-hidden="true" tabindex="-1"></a><span class="an">execute:</span></span>
|
||||
<span id="cb5-10"><a href="#cb5-10" aria-hidden="true" tabindex="-1"></a><span class="co"> enabled: false</span></span>
|
||||
<span id="cb5-11"><a href="#cb5-11" aria-hidden="true" tabindex="-1"></a><span class="co">---</span></span>
|
||||
<span id="cb5-12"><a href="#cb5-12" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-13"><a href="#cb5-13" aria-hidden="true" tabindex="-1"></a>This guide covers advanced training configurations for multi-GPU setups using Axolotl.</span>
|
||||
<span id="cb5-14"><a href="#cb5-14" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-15"><a href="#cb5-15" aria-hidden="true" tabindex="-1"></a><span class="fu">## Overview {#sec-overview}</span></span>
|
||||
<span id="cb5-16"><a href="#cb5-16" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-17"><a href="#cb5-17" aria-hidden="true" tabindex="-1"></a>Axolotl supports several methods for multi-GPU training:</span>
|
||||
<span id="cb5-18"><a href="#cb5-18" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-19"><a href="#cb5-19" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>DeepSpeed (recommended)</span>
|
||||
<span id="cb5-20"><a href="#cb5-20" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>FSDP (Fully Sharded Data Parallel)</span>
|
||||
<span id="cb5-21"><a href="#cb5-21" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Sequence parallelism</span>
|
||||
<span id="cb5-22"><a href="#cb5-22" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>FSDP + QLoRA</span>
|
||||
<span id="cb5-23"><a href="#cb5-23" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-24"><a href="#cb5-24" aria-hidden="true" tabindex="-1"></a><span class="fu">## DeepSpeed {#sec-deepspeed}</span></span>
|
||||
<span id="cb5-25"><a href="#cb5-25" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-26"><a href="#cb5-26" aria-hidden="true" tabindex="-1"></a>DeepSpeed is the recommended approach for multi-GPU training due to its stability and performance. It provides various optimization levels through ZeRO stages.</span>
|
||||
<span id="cb5-27"><a href="#cb5-27" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-28"><a href="#cb5-28" aria-hidden="true" tabindex="-1"></a><span class="fu">### Configuration {#sec-deepspeed-config}</span></span>
|
||||
<span id="cb5-29"><a href="#cb5-29" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-30"><a href="#cb5-30" aria-hidden="true" tabindex="-1"></a>Add to your YAML config:</span>
|
||||
<span id="cb5-31"><a href="#cb5-31" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-32"><a href="#cb5-32" aria-hidden="true" tabindex="-1"></a><span class="in">```{.yaml}</span></span>
|
||||
<span id="cb5-33"><a href="#cb5-33" aria-hidden="true" tabindex="-1"></a><span class="fu">deepspeed</span><span class="kw">:</span><span class="at"> deepspeed_configs/zero1.json</span></span>
|
||||
<span id="cb5-34"><a href="#cb5-34" aria-hidden="true" tabindex="-1"></a><span class="in">```</span></span>
|
||||
<span id="cb5-35"><a href="#cb5-35" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-36"><a href="#cb5-36" aria-hidden="true" tabindex="-1"></a><span class="fu">### Usage {#sec-deepspeed-usage}</span></span>
|
||||
<span id="cb5-37"><a href="#cb5-37" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-38"><a href="#cb5-38" aria-hidden="true" tabindex="-1"></a><span class="in">```{.bash}</span></span>
|
||||
<span id="cb5-39"><a href="#cb5-39" aria-hidden="true" tabindex="-1"></a><span class="co"># Fetch deepspeed configs (if not already present)</span></span>
|
||||
<span id="cb5-40"><a href="#cb5-40" aria-hidden="true" tabindex="-1"></a><span class="ex">axolotl</span> fetch deepspeed_configs</span>
|
||||
<span id="cb5-41"><a href="#cb5-41" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-42"><a href="#cb5-42" aria-hidden="true" tabindex="-1"></a><span class="co"># Passing arg via config</span></span>
|
||||
<span id="cb5-43"><a href="#cb5-43" aria-hidden="true" tabindex="-1"></a><span class="ex">axolotl</span> train config.yml</span>
|
||||
<span id="cb5-44"><a href="#cb5-44" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-45"><a href="#cb5-45" aria-hidden="true" tabindex="-1"></a><span class="co"># Passing arg via cli</span></span>
|
||||
<span id="cb5-46"><a href="#cb5-46" aria-hidden="true" tabindex="-1"></a><span class="ex">axolotl</span> train config.yml <span class="at">--deepspeed</span> deepspeed_configs/zero1.json</span>
|
||||
<span id="cb5-47"><a href="#cb5-47" aria-hidden="true" tabindex="-1"></a><span class="in">```</span></span>
|
||||
<span id="cb5-48"><a href="#cb5-48" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-49"><a href="#cb5-49" aria-hidden="true" tabindex="-1"></a><span class="fu">### ZeRO Stages {#sec-zero-stages}</span></span>
|
||||
<span id="cb5-50"><a href="#cb5-50" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-51"><a href="#cb5-51" aria-hidden="true" tabindex="-1"></a>We provide default configurations for:</span>
|
||||
<span id="cb5-52"><a href="#cb5-52" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-53"><a href="#cb5-53" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 1 (<span class="in">`zero1.json`</span>)</span>
|
||||
<span id="cb5-54"><a href="#cb5-54" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 1 with torch compile (<span class="in">`zero1_torch_compile.json`</span>)</span>
|
||||
<span id="cb5-55"><a href="#cb5-55" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 2 (<span class="in">`zero2.json`</span>)</span>
|
||||
<span id="cb5-56"><a href="#cb5-56" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 3 (<span class="in">`zero3.json`</span>)</span>
|
||||
<span id="cb5-57"><a href="#cb5-57" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 3 with bf16 (<span class="in">`zero3_bf16.json`</span>)</span>
|
||||
<span id="cb5-58"><a href="#cb5-58" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 3 with bf16 and CPU offload params(<span class="in">`zero3_bf16_cpuoffload_params.json`</span>)</span>
|
||||
<span id="cb5-59"><a href="#cb5-59" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 3 with bf16 and CPU offload params and optimizer (<span class="in">`zero3_bf16_cpuoffload_all.json`</span>)</span>
|
||||
<span id="cb5-60"><a href="#cb5-60" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-61"><a href="#cb5-61" aria-hidden="true" tabindex="-1"></a>::: {.callout-tip}</span>
|
||||
<span id="cb5-62"><a href="#cb5-62" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-63"><a href="#cb5-63" aria-hidden="true" tabindex="-1"></a>Choose the configuration that offloads the least amount to memory while still being able to fit on VRAM for best performance.</span>
|
||||
<span id="cb5-64"><a href="#cb5-64" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-65"><a href="#cb5-65" aria-hidden="true" tabindex="-1"></a>Start from Stage 1 -> Stage 2 -> Stage 3.</span>
|
||||
<span id="cb5-66"><a href="#cb5-66" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-67"><a href="#cb5-67" aria-hidden="true" tabindex="-1"></a>:::</span>
|
||||
<span id="cb5-68"><a href="#cb5-68" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-69"><a href="#cb5-69" aria-hidden="true" tabindex="-1"></a><span class="fu">## FSDP {#sec-fsdp}</span></span>
|
||||
<span id="cb5-70"><a href="#cb5-70" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-71"><a href="#cb5-71" aria-hidden="true" tabindex="-1"></a><span class="fu">### Basic FSDP Configuration {#sec-fsdp-config}</span></span>
|
||||
<span id="cb5-72"><a href="#cb5-72" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-73"><a href="#cb5-73" aria-hidden="true" tabindex="-1"></a><span class="in">```{.yaml}</span></span>
|
||||
<span id="cb5-74"><a href="#cb5-74" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp</span><span class="kw">:</span></span>
|
||||
<span id="cb5-75"><a href="#cb5-75" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="kw">-</span><span class="at"> </span>full_shard</span>
|
||||
<span id="cb5-76"><a href="#cb5-76" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="kw">-</span><span class="at"> </span>auto_wrap</span>
|
||||
<span id="cb5-77"><a href="#cb5-77" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp_config</span><span class="kw">:</span></span>
|
||||
<span id="cb5-78"><a href="#cb5-78" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">fsdp_offload_params</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
|
||||
<span id="cb5-79"><a href="#cb5-79" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">fsdp_state_dict_type</span><span class="kw">:</span><span class="at"> FULL_STATE_DICT</span></span>
|
||||
<span id="cb5-80"><a href="#cb5-80" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">fsdp_transformer_layer_cls_to_wrap</span><span class="kw">:</span><span class="at"> LlamaDecoderLayer</span></span>
|
||||
<span id="cb5-81"><a href="#cb5-81" aria-hidden="true" tabindex="-1"></a><span class="in">```</span></span>
|
||||
<span id="cb5-82"><a href="#cb5-82" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-83"><a href="#cb5-83" aria-hidden="true" tabindex="-1"></a><span class="fu">## Sequence parallelism {#sec-sequence-parallelism}</span></span>
|
||||
<span id="cb5-84"><a href="#cb5-84" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-85"><a href="#cb5-85" aria-hidden="true" tabindex="-1"></a>We support sequence parallelism (SP) via the</span>
|
||||
<span id="cb5-86"><a href="#cb5-86" aria-hidden="true" tabindex="-1"></a><span class="co">[</span><span class="ot">ring-flash-attention</span><span class="co">](https://github.com/zhuzilin/ring-flash-attention)</span> project. This</span>
|
||||
<span id="cb5-87"><a href="#cb5-87" aria-hidden="true" tabindex="-1"></a>allows one to split up sequences across GPUs, which is useful in the event that a</span>
|
||||
<span id="cb5-88"><a href="#cb5-88" aria-hidden="true" tabindex="-1"></a>single sequence causes OOM errors during model training.</span>
|
||||
<span id="cb5-89"><a href="#cb5-89" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-90"><a href="#cb5-90" aria-hidden="true" tabindex="-1"></a>First, install <span class="in">`ring-flash-attn`</span>, recommended via <span class="in">`pip install axolotl[ring-flash-attn]`</span>,</span>
|
||||
<span id="cb5-91"><a href="#cb5-91" aria-hidden="true" tabindex="-1"></a>or from source with <span class="in">`pip install .[ring-flash-attn]`</span>.</span>
|
||||
<span id="cb5-92"><a href="#cb5-92" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-93"><a href="#cb5-93" aria-hidden="true" tabindex="-1"></a>Your Axolotl YAML config should contain the following lines:</span>
|
||||
<span id="cb5-94"><a href="#cb5-94" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-95"><a href="#cb5-95" aria-hidden="true" tabindex="-1"></a><span class="in">```{.yaml}</span></span>
|
||||
<span id="cb5-96"><a href="#cb5-96" aria-hidden="true" tabindex="-1"></a><span class="fu">sequence_parallel_degree</span><span class="kw">:</span><span class="at"> </span><span class="dv">4</span><span class="co"> # Split each sequence into 4 parts, one per GPU</span></span>
|
||||
<span id="cb5-97"><a href="#cb5-97" aria-hidden="true" tabindex="-1"></a><span class="fu">flash_attention</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span><span class="co"> # Required with sequence parallelism</span></span>
|
||||
<span id="cb5-98"><a href="#cb5-98" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-99"><a href="#cb5-99" aria-hidden="true" tabindex="-1"></a><span class="co"># Optional; strides across the key dimension. Larger values use more memory but will make training faster.</span></span>
|
||||
<span id="cb5-100"><a href="#cb5-100" 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="cb5-101"><a href="#cb5-101" aria-hidden="true" tabindex="-1"></a><span class="in">```</span></span>
|
||||
<span id="cb5-102"><a href="#cb5-102" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-103"><a href="#cb5-103" aria-hidden="true" tabindex="-1"></a>See our <span class="co">[</span><span class="ot">dedicated guide</span><span class="co">](sequence_parallelism.qmd)</span> for more details.</span>
|
||||
<span id="cb5-104"><a href="#cb5-104" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-105"><a href="#cb5-105" aria-hidden="true" tabindex="-1"></a><span class="fu">### FSDP + QLoRA {#sec-fsdp-qlora}</span></span>
|
||||
<span id="cb5-106"><a href="#cb5-106" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-107"><a href="#cb5-107" aria-hidden="true" tabindex="-1"></a>For combining FSDP with QLoRA, see our <span class="co">[</span><span class="ot">dedicated guide</span><span class="co">](fsdp_qlora.qmd)</span>.</span>
|
||||
<span id="cb5-108"><a href="#cb5-108" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-109"><a href="#cb5-109" aria-hidden="true" tabindex="-1"></a><span class="fu">## Performance Optimization {#sec-performance}</span></span>
|
||||
<span id="cb5-110"><a href="#cb5-110" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-111"><a href="#cb5-111" aria-hidden="true" tabindex="-1"></a><span class="fu">### Liger Kernel Integration {#sec-liger}</span></span>
|
||||
<span id="cb5-112"><a href="#cb5-112" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-113"><a href="#cb5-113" aria-hidden="true" tabindex="-1"></a>Please see <span class="co">[</span><span class="ot">docs</span><span class="co">](custom_integrations.qmd#liger)</span> for more info.</span>
|
||||
<span id="cb5-114"><a href="#cb5-114" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-115"><a href="#cb5-115" aria-hidden="true" tabindex="-1"></a><span class="fu">## Troubleshooting {#sec-troubleshooting}</span></span>
|
||||
<span id="cb5-116"><a href="#cb5-116" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-117"><a href="#cb5-117" aria-hidden="true" tabindex="-1"></a><span class="fu">### NCCL Issues {#sec-nccl}</span></span>
|
||||
<span id="cb5-118"><a href="#cb5-118" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-119"><a href="#cb5-119" aria-hidden="true" tabindex="-1"></a>For NCCL-related problems, see our <span class="co">[</span><span class="ot">NCCL troubleshooting guide</span><span class="co">](nccl.qmd)</span>.</span>
|
||||
<span id="cb5-120"><a href="#cb5-120" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-121"><a href="#cb5-121" aria-hidden="true" tabindex="-1"></a><span class="fu">### Common Problems {#sec-common-problems}</span></span>
|
||||
<span id="cb5-122"><a href="#cb5-122" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-123"><a href="#cb5-123" aria-hidden="true" tabindex="-1"></a>::: {.panel-tabset}</span>
|
||||
<span id="cb5-124"><a href="#cb5-124" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-125"><a href="#cb5-125" aria-hidden="true" tabindex="-1"></a><span class="fu">## Memory Issues</span></span>
|
||||
<span id="cb5-126"><a href="#cb5-126" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-127"><a href="#cb5-127" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Reduce <span class="in">`micro_batch_size`</span></span>
|
||||
<span id="cb5-128"><a href="#cb5-128" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Reduce <span class="in">`eval_batch_size`</span></span>
|
||||
<span id="cb5-129"><a href="#cb5-129" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Adjust <span class="in">`gradient_accumulation_steps`</span></span>
|
||||
<span id="cb5-130"><a href="#cb5-130" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Consider using a higher ZeRO stage</span>
|
||||
<span id="cb5-131"><a href="#cb5-131" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-132"><a href="#cb5-132" aria-hidden="true" tabindex="-1"></a><span class="fu">## Training Instability</span></span>
|
||||
<span id="cb5-133"><a href="#cb5-133" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-134"><a href="#cb5-134" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Start with DeepSpeed ZeRO-2</span>
|
||||
<span id="cb5-135"><a href="#cb5-135" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Monitor loss values</span>
|
||||
<span id="cb5-136"><a href="#cb5-136" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Check learning rates</span>
|
||||
<span id="cb5-137"><a href="#cb5-137" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-138"><a href="#cb5-138" aria-hidden="true" tabindex="-1"></a>:::</span>
|
||||
<span id="cb5-139"><a href="#cb5-139" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb5-140"><a href="#cb5-140" aria-hidden="true" tabindex="-1"></a>For more detailed troubleshooting, see our <span class="co">[</span><span class="ot">debugging guide</span><span class="co">](debugging.qmd)</span>.</span></code><button title="Copy to Clipboard" class="code-copy-button" data-in-quarto-modal=""><i class="bi"></i></button></pre></div>
|
||||
<div class="sourceCode" id="cb4" data-shortcodes="false"><pre class="sourceCode markdown code-with-copy"><code class="sourceCode markdown"><span id="cb4-1"><a href="#cb4-1" aria-hidden="true" tabindex="-1"></a><span class="co">---</span></span>
|
||||
<span id="cb4-2"><a href="#cb4-2" aria-hidden="true" tabindex="-1"></a><span class="an">title:</span><span class="co"> "Multi-GPU"</span></span>
|
||||
<span id="cb4-3"><a href="#cb4-3" aria-hidden="true" tabindex="-1"></a><span class="an">format:</span></span>
|
||||
<span id="cb4-4"><a href="#cb4-4" aria-hidden="true" tabindex="-1"></a><span class="co"> html:</span></span>
|
||||
<span id="cb4-5"><a href="#cb4-5" aria-hidden="true" tabindex="-1"></a><span class="co"> toc: true</span></span>
|
||||
<span id="cb4-6"><a href="#cb4-6" aria-hidden="true" tabindex="-1"></a><span class="co"> toc-depth: 3</span></span>
|
||||
<span id="cb4-7"><a href="#cb4-7" aria-hidden="true" tabindex="-1"></a><span class="co"> number-sections: true</span></span>
|
||||
<span id="cb4-8"><a href="#cb4-8" aria-hidden="true" tabindex="-1"></a><span class="co"> code-tools: true</span></span>
|
||||
<span id="cb4-9"><a href="#cb4-9" aria-hidden="true" tabindex="-1"></a><span class="an">execute:</span></span>
|
||||
<span id="cb4-10"><a href="#cb4-10" aria-hidden="true" tabindex="-1"></a><span class="co"> enabled: false</span></span>
|
||||
<span id="cb4-11"><a href="#cb4-11" aria-hidden="true" tabindex="-1"></a><span class="co">---</span></span>
|
||||
<span id="cb4-12"><a href="#cb4-12" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-13"><a href="#cb4-13" aria-hidden="true" tabindex="-1"></a>This guide covers advanced training configurations for multi-GPU setups using Axolotl.</span>
|
||||
<span id="cb4-14"><a href="#cb4-14" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-15"><a href="#cb4-15" aria-hidden="true" tabindex="-1"></a><span class="fu">## Overview {#sec-overview}</span></span>
|
||||
<span id="cb4-16"><a href="#cb4-16" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-17"><a href="#cb4-17" aria-hidden="true" tabindex="-1"></a>Axolotl supports several methods for multi-GPU training:</span>
|
||||
<span id="cb4-18"><a href="#cb4-18" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-19"><a href="#cb4-19" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>DeepSpeed (recommended)</span>
|
||||
<span id="cb4-20"><a href="#cb4-20" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>FSDP (Fully Sharded Data Parallel)</span>
|
||||
<span id="cb4-21"><a href="#cb4-21" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Sequence parallelism</span>
|
||||
<span id="cb4-22"><a href="#cb4-22" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>FSDP + QLoRA</span>
|
||||
<span id="cb4-23"><a href="#cb4-23" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-24"><a href="#cb4-24" aria-hidden="true" tabindex="-1"></a><span class="fu">## DeepSpeed {#sec-deepspeed}</span></span>
|
||||
<span id="cb4-25"><a href="#cb4-25" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-26"><a href="#cb4-26" aria-hidden="true" tabindex="-1"></a>DeepSpeed is the recommended approach for multi-GPU training due to its stability and performance. It provides various optimization levels through ZeRO stages.</span>
|
||||
<span id="cb4-27"><a href="#cb4-27" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-28"><a href="#cb4-28" aria-hidden="true" tabindex="-1"></a><span class="fu">### Configuration {#sec-deepspeed-config}</span></span>
|
||||
<span id="cb4-29"><a href="#cb4-29" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-30"><a href="#cb4-30" aria-hidden="true" tabindex="-1"></a>Add to your YAML config:</span>
|
||||
<span id="cb4-31"><a href="#cb4-31" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-32"><a href="#cb4-32" aria-hidden="true" tabindex="-1"></a><span class="in">```{.yaml}</span></span>
|
||||
<span id="cb4-33"><a href="#cb4-33" aria-hidden="true" tabindex="-1"></a><span class="fu">deepspeed</span><span class="kw">:</span><span class="at"> deepspeed_configs/zero1.json</span></span>
|
||||
<span id="cb4-34"><a href="#cb4-34" aria-hidden="true" tabindex="-1"></a><span class="in">```</span></span>
|
||||
<span id="cb4-35"><a href="#cb4-35" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-36"><a href="#cb4-36" aria-hidden="true" tabindex="-1"></a><span class="fu">### Usage {#sec-deepspeed-usage}</span></span>
|
||||
<span id="cb4-37"><a href="#cb4-37" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-38"><a href="#cb4-38" aria-hidden="true" tabindex="-1"></a><span class="in">```{.bash}</span></span>
|
||||
<span id="cb4-39"><a href="#cb4-39" aria-hidden="true" tabindex="-1"></a><span class="co"># Fetch deepspeed configs (if not already present)</span></span>
|
||||
<span id="cb4-40"><a href="#cb4-40" aria-hidden="true" tabindex="-1"></a><span class="ex">axolotl</span> fetch deepspeed_configs</span>
|
||||
<span id="cb4-41"><a href="#cb4-41" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-42"><a href="#cb4-42" aria-hidden="true" tabindex="-1"></a><span class="co"># Passing arg via config</span></span>
|
||||
<span id="cb4-43"><a href="#cb4-43" aria-hidden="true" tabindex="-1"></a><span class="ex">axolotl</span> train config.yml</span>
|
||||
<span id="cb4-44"><a href="#cb4-44" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-45"><a href="#cb4-45" aria-hidden="true" tabindex="-1"></a><span class="co"># Passing arg via cli</span></span>
|
||||
<span id="cb4-46"><a href="#cb4-46" aria-hidden="true" tabindex="-1"></a><span class="ex">axolotl</span> train config.yml <span class="at">--deepspeed</span> deepspeed_configs/zero1.json</span>
|
||||
<span id="cb4-47"><a href="#cb4-47" aria-hidden="true" tabindex="-1"></a><span class="in">```</span></span>
|
||||
<span id="cb4-48"><a href="#cb4-48" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-49"><a href="#cb4-49" aria-hidden="true" tabindex="-1"></a><span class="fu">### ZeRO Stages {#sec-zero-stages}</span></span>
|
||||
<span id="cb4-50"><a href="#cb4-50" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-51"><a href="#cb4-51" aria-hidden="true" tabindex="-1"></a>We provide default configurations for:</span>
|
||||
<span id="cb4-52"><a href="#cb4-52" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-53"><a href="#cb4-53" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 1 (<span class="in">`zero1.json`</span>)</span>
|
||||
<span id="cb4-54"><a href="#cb4-54" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 1 with torch compile (<span class="in">`zero1_torch_compile.json`</span>)</span>
|
||||
<span id="cb4-55"><a href="#cb4-55" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 2 (<span class="in">`zero2.json`</span>)</span>
|
||||
<span id="cb4-56"><a href="#cb4-56" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 3 (<span class="in">`zero3.json`</span>)</span>
|
||||
<span id="cb4-57"><a href="#cb4-57" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 3 with bf16 (<span class="in">`zero3_bf16.json`</span>)</span>
|
||||
<span id="cb4-58"><a href="#cb4-58" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 3 with bf16 and CPU offload params(<span class="in">`zero3_bf16_cpuoffload_params.json`</span>)</span>
|
||||
<span id="cb4-59"><a href="#cb4-59" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>ZeRO Stage 3 with bf16 and CPU offload params and optimizer (<span class="in">`zero3_bf16_cpuoffload_all.json`</span>)</span>
|
||||
<span id="cb4-60"><a href="#cb4-60" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-61"><a href="#cb4-61" aria-hidden="true" tabindex="-1"></a>::: {.callout-tip}</span>
|
||||
<span id="cb4-62"><a href="#cb4-62" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-63"><a href="#cb4-63" aria-hidden="true" tabindex="-1"></a>Choose the configuration that offloads the least amount to memory while still being able to fit on VRAM for best performance.</span>
|
||||
<span id="cb4-64"><a href="#cb4-64" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-65"><a href="#cb4-65" aria-hidden="true" tabindex="-1"></a>Start from Stage 1 -> Stage 2 -> Stage 3.</span>
|
||||
<span id="cb4-66"><a href="#cb4-66" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-67"><a href="#cb4-67" aria-hidden="true" tabindex="-1"></a>:::</span>
|
||||
<span id="cb4-68"><a href="#cb4-68" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-69"><a href="#cb4-69" aria-hidden="true" tabindex="-1"></a><span class="fu">## FSDP {#sec-fsdp}</span></span>
|
||||
<span id="cb4-70"><a href="#cb4-70" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-71"><a href="#cb4-71" aria-hidden="true" tabindex="-1"></a><span class="fu">### Basic FSDP Configuration {#sec-fsdp-config}</span></span>
|
||||
<span id="cb4-72"><a href="#cb4-72" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-73"><a href="#cb4-73" aria-hidden="true" tabindex="-1"></a><span class="in">```{.yaml}</span></span>
|
||||
<span id="cb4-74"><a href="#cb4-74" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp</span><span class="kw">:</span></span>
|
||||
<span id="cb4-75"><a href="#cb4-75" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="kw">-</span><span class="at"> </span>full_shard</span>
|
||||
<span id="cb4-76"><a href="#cb4-76" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="kw">-</span><span class="at"> </span>auto_wrap</span>
|
||||
<span id="cb4-77"><a href="#cb4-77" aria-hidden="true" tabindex="-1"></a><span class="fu">fsdp_config</span><span class="kw">:</span></span>
|
||||
<span id="cb4-78"><a href="#cb4-78" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">fsdp_offload_params</span><span class="kw">:</span><span class="at"> </span><span class="ch">true</span></span>
|
||||
<span id="cb4-79"><a href="#cb4-79" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">fsdp_state_dict_type</span><span class="kw">:</span><span class="at"> FULL_STATE_DICT</span></span>
|
||||
<span id="cb4-80"><a href="#cb4-80" aria-hidden="true" tabindex="-1"></a><span class="at"> </span><span class="fu">fsdp_transformer_layer_cls_to_wrap</span><span class="kw">:</span><span class="at"> LlamaDecoderLayer</span></span>
|
||||
<span id="cb4-81"><a href="#cb4-81" aria-hidden="true" tabindex="-1"></a><span class="in">```</span></span>
|
||||
<span id="cb4-82"><a href="#cb4-82" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-83"><a href="#cb4-83" aria-hidden="true" tabindex="-1"></a><span class="fu">## Sequence parallelism {#sec-sequence-parallelism}</span></span>
|
||||
<span id="cb4-84"><a href="#cb4-84" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-85"><a href="#cb4-85" aria-hidden="true" tabindex="-1"></a>We support sequence parallelism (SP) via the</span>
|
||||
<span id="cb4-86"><a href="#cb4-86" aria-hidden="true" tabindex="-1"></a><span class="co">[</span><span class="ot">ring-flash-attention</span><span class="co">](https://github.com/zhuzilin/ring-flash-attention)</span> project. This</span>
|
||||
<span id="cb4-87"><a href="#cb4-87" aria-hidden="true" tabindex="-1"></a>allows one to split up sequences across GPUs, which is useful in the event that a</span>
|
||||
<span id="cb4-88"><a href="#cb4-88" aria-hidden="true" tabindex="-1"></a>single sequence causes OOM errors during model training.</span>
|
||||
<span id="cb4-89"><a href="#cb4-89" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-90"><a href="#cb4-90" aria-hidden="true" tabindex="-1"></a>See our <span class="co">[</span><span class="ot">dedicated guide</span><span class="co">](sequence_parallelism.qmd)</span> for more information.</span>
|
||||
<span id="cb4-91"><a href="#cb4-91" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-92"><a href="#cb4-92" aria-hidden="true" tabindex="-1"></a><span class="fu">### FSDP + QLoRA {#sec-fsdp-qlora}</span></span>
|
||||
<span id="cb4-93"><a href="#cb4-93" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-94"><a href="#cb4-94" aria-hidden="true" tabindex="-1"></a>For combining FSDP with QLoRA, see our <span class="co">[</span><span class="ot">dedicated guide</span><span class="co">](fsdp_qlora.qmd)</span>.</span>
|
||||
<span id="cb4-95"><a href="#cb4-95" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-96"><a href="#cb4-96" aria-hidden="true" tabindex="-1"></a><span class="fu">## Performance Optimization {#sec-performance}</span></span>
|
||||
<span id="cb4-97"><a href="#cb4-97" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-98"><a href="#cb4-98" aria-hidden="true" tabindex="-1"></a><span class="fu">### Liger Kernel Integration {#sec-liger}</span></span>
|
||||
<span id="cb4-99"><a href="#cb4-99" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-100"><a href="#cb4-100" aria-hidden="true" tabindex="-1"></a>Please see <span class="co">[</span><span class="ot">docs</span><span class="co">](custom_integrations.qmd#liger)</span> for more info.</span>
|
||||
<span id="cb4-101"><a href="#cb4-101" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-102"><a href="#cb4-102" aria-hidden="true" tabindex="-1"></a><span class="fu">## Troubleshooting {#sec-troubleshooting}</span></span>
|
||||
<span id="cb4-103"><a href="#cb4-103" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-104"><a href="#cb4-104" aria-hidden="true" tabindex="-1"></a><span class="fu">### NCCL Issues {#sec-nccl}</span></span>
|
||||
<span id="cb4-105"><a href="#cb4-105" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-106"><a href="#cb4-106" aria-hidden="true" tabindex="-1"></a>For NCCL-related problems, see our <span class="co">[</span><span class="ot">NCCL troubleshooting guide</span><span class="co">](nccl.qmd)</span>.</span>
|
||||
<span id="cb4-107"><a href="#cb4-107" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-108"><a href="#cb4-108" aria-hidden="true" tabindex="-1"></a><span class="fu">### Common Problems {#sec-common-problems}</span></span>
|
||||
<span id="cb4-109"><a href="#cb4-109" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-110"><a href="#cb4-110" aria-hidden="true" tabindex="-1"></a>::: {.panel-tabset}</span>
|
||||
<span id="cb4-111"><a href="#cb4-111" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-112"><a href="#cb4-112" aria-hidden="true" tabindex="-1"></a><span class="fu">## Memory Issues</span></span>
|
||||
<span id="cb4-113"><a href="#cb4-113" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-114"><a href="#cb4-114" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Reduce <span class="in">`micro_batch_size`</span></span>
|
||||
<span id="cb4-115"><a href="#cb4-115" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Reduce <span class="in">`eval_batch_size`</span></span>
|
||||
<span id="cb4-116"><a href="#cb4-116" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Adjust <span class="in">`gradient_accumulation_steps`</span></span>
|
||||
<span id="cb4-117"><a href="#cb4-117" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Consider using a higher ZeRO stage</span>
|
||||
<span id="cb4-118"><a href="#cb4-118" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-119"><a href="#cb4-119" aria-hidden="true" tabindex="-1"></a><span class="fu">## Training Instability</span></span>
|
||||
<span id="cb4-120"><a href="#cb4-120" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-121"><a href="#cb4-121" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Start with DeepSpeed ZeRO-2</span>
|
||||
<span id="cb4-122"><a href="#cb4-122" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Monitor loss values</span>
|
||||
<span id="cb4-123"><a href="#cb4-123" aria-hidden="true" tabindex="-1"></a><span class="ss">- </span>Check learning rates</span>
|
||||
<span id="cb4-124"><a href="#cb4-124" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-125"><a href="#cb4-125" aria-hidden="true" tabindex="-1"></a>:::</span>
|
||||
<span id="cb4-126"><a href="#cb4-126" aria-hidden="true" tabindex="-1"></a></span>
|
||||
<span id="cb4-127"><a href="#cb4-127" aria-hidden="true" tabindex="-1"></a>For more detailed troubleshooting, see our <span class="co">[</span><span class="ot">debugging guide</span><span class="co">](debugging.qmd)</span>.</span></code><button title="Copy to Clipboard" class="code-copy-button" data-in-quarto-modal=""><i class="bi"></i></button></pre></div>
|
||||
</div></div></div></div></div>
|
||||
</div> <!-- /content -->
|
||||
|
||||
|
||||
Reference in New Issue
Block a user