mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
2079 lines
303 KiB
HTML
2079 lines
303 KiB
HTML
|
|
|
|
<!DOCTYPE html>
|
|
<html class="writer-html5" lang="en" data-content_root="../../../">
|
|
<head>
|
|
<meta charset="utf-8" />
|
|
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
|
|
<title>tensorrt_llm.models.modeling_utils — tensorrt_llm documentation</title>
|
|
<link rel="stylesheet" type="text/css" href="../../../_static/pygments.css?v=80d5e7a1" />
|
|
<link rel="stylesheet" type="text/css" href="../../../_static/css/theme.css?v=e59714d7" />
|
|
<link rel="stylesheet" type="text/css" href="../../../_static/copybutton.css?v=76b2166b" />
|
|
|
|
|
|
<script src="../../../_static/jquery.js?v=5d32c60e"></script>
|
|
<script src="../../../_static/_sphinx_javascript_frameworks_compat.js?v=2cd50e6c"></script>
|
|
<script src="../../../_static/documentation_options.js?v=5929fcd5"></script>
|
|
<script src="../../../_static/doctools.js?v=9a2dae69"></script>
|
|
<script src="../../../_static/sphinx_highlight.js?v=dc90522c"></script>
|
|
<script src="../../../_static/clipboard.min.js?v=a7894cd8"></script>
|
|
<script src="../../../_static/copybutton.js?v=65e89d2a"></script>
|
|
<script src="../../../_static/js/theme.js"></script>
|
|
<link rel="index" title="Index" href="../../../genindex.html" />
|
|
<link rel="search" title="Search" href="../../../search.html" />
|
|
</head>
|
|
|
|
<body class="wy-body-for-nav">
|
|
<div class="wy-grid-for-nav">
|
|
<nav data-toggle="wy-nav-shift" class="wy-nav-side">
|
|
<div class="wy-side-scroll">
|
|
<div class="wy-side-nav-search" >
|
|
|
|
|
|
|
|
<a href="../../../index.html" class="icon icon-home">
|
|
tensorrt_llm
|
|
</a>
|
|
<div role="search">
|
|
<form id="rtd-search-form" class="wy-form" action="../../../search.html" method="get">
|
|
<input type="text" name="q" placeholder="Search docs" aria-label="Search docs" />
|
|
<input type="hidden" name="check_keywords" value="yes" />
|
|
<input type="hidden" name="area" value="default" />
|
|
</form>
|
|
</div>
|
|
</div><div class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="Navigation menu">
|
|
<p class="caption" role="heading"><span class="caption-text">Getting Started</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../overview.html">Overview</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../quick-start-guide.html">Quick Start Guide</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../key-features.html">Key Features</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../release-notes.html">Release Notes</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Installation</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../installation/linux.html">Installing on Linux</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../installation/build-from-source-linux.html">Building from Source Code on Linux</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../installation/windows.html">Installing on Windows</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../installation/build-from-source-windows.html">Building from Source Code on Windows</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../installation/grace-hopper.html">Installing on Grace Hopper</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">LLM API</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../llm-api/index.html">API Introduction</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../llm-api/reference.html">API Reference</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">LLM API Examples</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../llm-api-examples/index.html">LLM Examples Introduction</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../llm-api-examples/customization.html">Common Customizations</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../llm-api-examples/llm_api_examples.html">Examples</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Model Definition API</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.layers.html">Layers</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.functional.html">Functionals</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.models.html">Models</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.plugin.html">Plugin</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.quantization.html">Quantization</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.runtime.html">Runtime</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">C++ API</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../_cpp_gen/executor.html">Executor</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../_cpp_gen/runtime.html">Runtime</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Command-Line Reference</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../commands/trtllm-build.html">trtllm-build</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../commands/trtllm-serve.html">trtllm-serve</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Architecture</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/overview.html">TensorRT-LLM Architecture</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/core-concepts.html">Model Definition</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/core-concepts.html#compilation">Compilation</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/core-concepts.html#runtime">Runtime</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/core-concepts.html#multi-gpu-and-multi-node-support">Multi-GPU and Multi-Node Support</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/checkpoint.html">TensorRT-LLM Checkpoint</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/workflow.html">TensorRT-LLM Build Workflow</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/add-model.html">Adding a Model</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Advanced</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/gpt-attention.html">Multi-Head, Multi-Query, and Group-Query Attention</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/gpt-runtime.html">C++ GPT Runtime</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/executor.html">Executor API</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/graph-rewriting.html">Graph Rewriting Module</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/inference-request.html">Inference Request</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/inference-request.html#responses">Responses</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/lora.html">Run gpt-2b + LoRA using GptManager / cpp runtime</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/expert-parallelism.html">Expert Parallelism in TensorRT-LLM</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/kv-cache-reuse.html">KV cache reuse</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/speculative-decoding.html">Speculative Sampling</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Performance</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../performance/perf-overview.html">Overview</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../performance/perf-benchmarking.html">Benchmarking</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../performance/perf-best-practices.html">Best Practices</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../performance/perf-analysis.html">Performance Analysis</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Reference</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../reference/troubleshooting.html">Troubleshooting</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../reference/support-matrix.html">Support Matrix</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../reference/precision.html">Numerical Precision</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../reference/memory.html">Memory Usage of TensorRT-LLM</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Blogs</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/H100vsA100.html">H100 has 4.6x A100 Performance in TensorRT-LLM, achieving 10,000 tok/s at 100ms to first token</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/H200launch.html">H200 achieves nearly 12,000 tokens/sec on Llama2-13B with TensorRT-LLM</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/Falcon180B-H200.html">Falcon-180B on a single H200 GPU with INT4 AWQ, and 6.7x faster Llama-70B over A100</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/quantization-in-TRT-LLM.html">Speed up inference with SOTA quantization techniques in TRT-LLM</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/XQA-kernel.html">New XQA-kernel provides 2.4x more Llama-70B throughput within the same latency budget</a></li>
|
|
</ul>
|
|
|
|
</div>
|
|
</div>
|
|
</nav>
|
|
|
|
<section data-toggle="wy-nav-shift" class="wy-nav-content-wrap"><nav class="wy-nav-top" aria-label="Mobile navigation menu" >
|
|
<i data-toggle="wy-nav-top" class="fa fa-bars"></i>
|
|
<a href="../../../index.html">tensorrt_llm</a>
|
|
</nav>
|
|
|
|
<div class="wy-nav-content">
|
|
<div class="rst-content">
|
|
<div role="navigation" aria-label="Page navigation">
|
|
<ul class="wy-breadcrumbs">
|
|
<li><a href="../../../index.html" class="icon icon-home" aria-label="Home"></a></li>
|
|
<li class="breadcrumb-item"><a href="../../index.html">Module code</a></li>
|
|
<li class="breadcrumb-item active">tensorrt_llm.models.modeling_utils</li>
|
|
<li class="wy-breadcrumbs-aside">
|
|
</li>
|
|
</ul>
|
|
<hr/>
|
|
</div>
|
|
<div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
|
|
<div itemprop="articleBody">
|
|
|
|
<h1>Source code for tensorrt_llm.models.modeling_utils</h1><div class="highlight"><pre>
|
|
<span></span><span class="kn">import</span> <span class="nn">argparse</span>
|
|
<span class="kn">import</span> <span class="nn">copy</span>
|
|
<span class="kn">import</span> <span class="nn">dataclasses</span>
|
|
<span class="kn">import</span> <span class="nn">json</span>
|
|
<span class="kn">import</span> <span class="nn">os</span>
|
|
<span class="kn">import</span> <span class="nn">re</span>
|
|
<span class="kn">from</span> <span class="nn">enum</span> <span class="kn">import</span> <span class="n">IntFlag</span><span class="p">,</span> <span class="n">auto</span>
|
|
<span class="kn">from</span> <span class="nn">functools</span> <span class="kn">import</span> <span class="n">cached_property</span>
|
|
<span class="kn">from</span> <span class="nn">pathlib</span> <span class="kn">import</span> <span class="n">Path</span>
|
|
<span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="p">(</span><span class="n">TYPE_CHECKING</span><span class="p">,</span> <span class="n">Callable</span><span class="p">,</span> <span class="n">Dict</span><span class="p">,</span> <span class="n">Generator</span><span class="p">,</span> <span class="n">List</span><span class="p">,</span> <span class="n">Optional</span><span class="p">,</span>
|
|
<span class="n">Union</span><span class="p">)</span>
|
|
|
|
<span class="kn">import</span> <span class="nn">numpy</span> <span class="k">as</span> <span class="nn">np</span>
|
|
<span class="kn">import</span> <span class="nn">safetensors</span>
|
|
<span class="kn">import</span> <span class="nn">torch</span>
|
|
|
|
<span class="kn">from</span> <span class="nn">.._common</span> <span class="kn">import</span> <span class="n">default_net</span>
|
|
<span class="kn">from</span> <span class="nn">.._utils</span> <span class="kn">import</span> <span class="p">(</span><span class="n">QuantModeWrapper</span><span class="p">,</span> <span class="n">get_init_params</span><span class="p">,</span> <span class="n">numpy_to_torch</span><span class="p">,</span>
|
|
<span class="n">release_gc</span><span class="p">,</span> <span class="n">str_dtype_to_torch</span><span class="p">,</span> <span class="n">str_dtype_to_trt</span><span class="p">,</span>
|
|
<span class="n">trt_dtype_to_torch</span><span class="p">)</span>
|
|
<span class="kn">from</span> <span class="nn">..bindings</span> <span class="kn">import</span> <span class="n">KVCacheType</span>
|
|
<span class="kn">from</span> <span class="nn">..bindings.executor</span> <span class="kn">import</span> <span class="n">RuntimeDefaults</span>
|
|
<span class="kn">from</span> <span class="nn">..functional</span> <span class="kn">import</span> <span class="p">(</span><span class="n">PositionEmbeddingType</span><span class="p">,</span> <span class="n">Tensor</span><span class="p">,</span> <span class="n">allgather</span><span class="p">,</span> <span class="n">constant</span><span class="p">,</span>
|
|
<span class="n">cp_split_plugin</span><span class="p">,</span> <span class="n">gather_last_token_logits</span><span class="p">,</span>
|
|
<span class="n">index_select</span><span class="p">,</span> <span class="n">tanh</span><span class="p">,</span> <span class="n">view</span><span class="p">)</span>
|
|
<span class="kn">from</span> <span class="nn">..layers</span> <span class="kn">import</span> <span class="p">(</span><span class="n">MLP</span><span class="p">,</span> <span class="n">AttentionParams</span><span class="p">,</span> <span class="n">Embedding</span><span class="p">,</span> <span class="n">FusedGatedMLP</span><span class="p">,</span>
|
|
<span class="n">FusedRgLru</span><span class="p">,</span> <span class="n">GatedMLP</span><span class="p">,</span> <span class="n">KeyValueCacheParams</span><span class="p">,</span> <span class="n">LoraParams</span><span class="p">,</span>
|
|
<span class="n">PromptTuningEmbedding</span><span class="p">,</span> <span class="n">RgLru</span><span class="p">)</span>
|
|
<span class="kn">from</span> <span class="nn">..layers.attention</span> <span class="kn">import</span> <span class="n">Attention</span><span class="p">,</span> <span class="n">BertAttention</span>
|
|
<span class="kn">from</span> <span class="nn">..layers.linear</span> <span class="kn">import</span> <span class="n">ColumnLinear</span><span class="p">,</span> <span class="n">Linear</span><span class="p">,</span> <span class="n">RowLinear</span>
|
|
<span class="kn">from</span> <span class="nn">..layers.lora</span> <span class="kn">import</span> <span class="n">Lora</span>
|
|
<span class="kn">from</span> <span class="nn">..layers.moe</span> <span class="kn">import</span> <span class="n">MOE</span><span class="p">,</span> <span class="n">MoeOOTB</span>
|
|
<span class="kn">from</span> <span class="nn">..logger</span> <span class="kn">import</span> <span class="n">logger</span>
|
|
<span class="kn">from</span> <span class="nn">..mapping</span> <span class="kn">import</span> <span class="n">Mapping</span>
|
|
<span class="kn">from</span> <span class="nn">..module</span> <span class="kn">import</span> <span class="n">Module</span><span class="p">,</span> <span class="n">ModuleList</span>
|
|
<span class="kn">from</span> <span class="nn">..parameter</span> <span class="kn">import</span> <span class="n">Parameter</span>
|
|
<span class="kn">from</span> <span class="nn">..plugin</span> <span class="kn">import</span> <span class="n">init_all_reduce_helper</span>
|
|
<span class="kn">from</span> <span class="nn">..quantization</span> <span class="kn">import</span> <span class="n">QuantMode</span>
|
|
<span class="kn">from</span> <span class="nn">..quantization.layers</span> <span class="kn">import</span> <span class="p">(</span><span class="n">FP8Linear</span><span class="p">,</span> <span class="n">Fp8RowwiseFusedGatedMLP</span><span class="p">,</span>
|
|
<span class="n">Fp8RowwiseGatedMLP</span><span class="p">,</span>
|
|
<span class="n">WeightOnlyGroupwiseQuantLinear</span><span class="p">,</span>
|
|
<span class="n">WeightOnlyGroupwiseQuantRowLinear</span><span class="p">,</span>
|
|
<span class="n">WeightOnlyQuantLinear</span><span class="p">,</span>
|
|
<span class="n">WeightOnlyQuantRowLinear</span><span class="p">)</span>
|
|
<span class="kn">from</span> <span class="nn">..quantization.mode</span> <span class="kn">import</span> <span class="p">(</span><span class="n">KV_CACHE_QUANT_ALGO_LIST</span><span class="p">,</span> <span class="n">QUANT_ALGO_LIST</span><span class="p">,</span>
|
|
<span class="n">W8A8_SQ_PLUGIN_LIST</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="p">)</span>
|
|
<span class="kn">from</span> <span class="nn">..top_model_mixin</span> <span class="kn">import</span> <span class="n">TopModelMixin</span>
|
|
<span class="kn">from</span> <span class="nn">.convert_utils</span> <span class="kn">import</span> <span class="n">weight_only_quantize_dict</span>
|
|
<span class="kn">from</span> <span class="nn">.generation_mixin</span> <span class="kn">import</span> <span class="n">GenerationMixin</span>
|
|
|
|
|
|
<span class="nd">@dataclasses</span><span class="o">.</span><span class="n">dataclass</span><span class="p">(</span><span class="n">kw_only</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">frozen</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
|
<span class="k">class</span> <span class="nc">Gemma2ConfigGroup</span><span class="p">:</span>
|
|
<span class="n">query_pre_attn_scalar</span><span class="p">:</span> <span class="nb">int</span>
|
|
<span class="n">final_logit_softcapping</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">float</span><span class="p">]</span>
|
|
<span class="n">attn_logit_softcapping</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">float</span><span class="p">]</span>
|
|
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">keys</span><span class="p">(</span><span class="bp">cls</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="p">{</span><span class="n">f</span><span class="o">.</span><span class="n">name</span> <span class="k">for</span> <span class="n">f</span> <span class="ow">in</span> <span class="n">dataclasses</span><span class="o">.</span><span class="n">fields</span><span class="p">(</span><span class="bp">cls</span><span class="p">)}</span>
|
|
|
|
|
|
<span class="k">if</span> <span class="n">TYPE_CHECKING</span><span class="p">:</span>
|
|
<span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">Type</span><span class="p">,</span> <span class="n">TypeVar</span>
|
|
|
|
<span class="kn">from</span> <span class="nn">typing_extensions</span> <span class="kn">import</span> <span class="n">Self</span>
|
|
|
|
<span class="n">ConfigGroups</span> <span class="o">=</span> <span class="n">Union</span><span class="p">[</span><span class="n">Gemma2ConfigGroup</span><span class="p">]</span>
|
|
<span class="w"> </span><span class="sd">"""Groupings of config where, if one of said properties exists, we assume all of the properties exist (even if they are `None`)"""</span>
|
|
<span class="n">CG</span> <span class="o">=</span> <span class="n">TypeVar</span><span class="p">(</span><span class="s2">"CG"</span><span class="p">,</span> <span class="n">bound</span><span class="o">=</span><span class="n">ConfigGroups</span><span class="p">)</span>
|
|
|
|
<span class="n">RuntimeDefaultsIn</span> <span class="o">=</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Union</span><span class="p">[</span><span class="n">RuntimeDefaults</span><span class="p">,</span> <span class="nb">dict</span><span class="p">]]</span>
|
|
|
|
|
|
<div class="viewcode-block" id="SpeculativeDecodingMode">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.SpeculativeDecodingMode">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">SpeculativeDecodingMode</span><span class="p">(</span><span class="n">IntFlag</span><span class="p">):</span>
|
|
<span class="c1"># [WARNING] KEEP BELOW DEFINITION IN SYNC WITH cpp/tensorrt_llm/runtime/speculativeDecodingMode.h</span>
|
|
<span class="n">NONE</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
|
|
<span class="n">DRAFT_TOKENS_EXTERNAL</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
|
|
<span class="n">MEDUSA</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
|
|
<span class="n">LOOKAHEAD_DECODING</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
|
|
<span class="n">EXPLICIT_DRAFT_TOKENS</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
|
|
<span class="n">EAGLE</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
|
|
|
|
<div class="viewcode-block" id="SpeculativeDecodingMode.from_arguments">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.SpeculativeDecodingMode.from_arguments">[docs]</a>
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_arguments</span><span class="p">(</span><span class="n">args</span><span class="p">:</span> <span class="n">argparse</span><span class="o">.</span><span class="n">Namespace</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">NONE</span>
|
|
<span class="k">elif</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">==</span> <span class="s2">"draft_tokens_external"</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">DRAFT_TOKENS_EXTERNAL</span>
|
|
<span class="k">elif</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">==</span> <span class="s2">"medusa"</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">MEDUSA</span>
|
|
<span class="k">elif</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">==</span> <span class="s2">"lookahead_decoding"</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">LOOKAHEAD_DECODING</span>
|
|
<span class="k">elif</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">==</span> <span class="s2">"explicit_draft_tokens"</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">EXPLICIT_DRAFT_TOKENS</span>
|
|
<span class="k">elif</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">==</span> <span class="s2">"eagle"</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">EAGLE</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="kc">False</span><span class="p">,</span> <span class="s2">"Unknown speculative_decoding_mode "</span> <span class="o">+</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span></div>
|
|
</div>
|
|
|
|
|
|
|
|
<div class="viewcode-block" id="QuantConfig">
|
|
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.QuantConfig">[docs]</a>
|
|
<span class="nd">@dataclasses</span><span class="o">.</span><span class="n">dataclass</span>
|
|
<span class="k">class</span> <span class="nc">QuantConfig</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">'''Serializable quantization configuration class, part of the PretrainedConfig</span>
|
|
<span class="sd"> '''</span>
|
|
|
|
<span class="n">quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantAlgo</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="n">kv_cache_quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantAlgo</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="n">group_size</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="mi">128</span>
|
|
<span class="n">smoothquant_val</span><span class="p">:</span> <span class="nb">float</span> <span class="o">=</span> <span class="mf">0.5</span>
|
|
<span class="n">clamp_val</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">List</span><span class="p">[</span><span class="nb">float</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="n">use_meta_recipe</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
<span class="n">has_zero_point</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">bool</span><span class="p">]</span> <span class="o">=</span> <span class="kc">False</span>
|
|
<span class="n">pre_quant_scale</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">bool</span><span class="p">]</span> <span class="o">=</span> <span class="kc">False</span>
|
|
<span class="n">exclude_modules</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">List</span><span class="p">[</span><span class="nb">str</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">use_plugin_sq</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">in</span> <span class="n">W8A8_SQ_PLUGIN_LIST</span>
|
|
|
|
<span class="nd">@cached_property</span>
|
|
<span class="k">def</span> <span class="nf">quant_mode</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">QuantModeWrapper</span><span class="p">:</span>
|
|
<span class="n">quant_mode_list</span> <span class="o">=</span> <span class="p">[</span>
|
|
<span class="n">QuantMode</span><span class="o">.</span><span class="n">from_quant_algo</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="p">]</span>
|
|
<span class="k">return</span> <span class="n">QuantModeWrapper</span><span class="p">(</span><span class="n">quant_mode_list</span><span class="p">)</span>
|
|
|
|
<span class="nd">@cached_property</span>
|
|
<span class="k">def</span> <span class="nf">layer_quant_mode</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">QuantMode</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">QuantMode</span><span class="o">.</span><span class="n">from_quant_algo</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">requires_calibration</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">in</span> <span class="p">(</span><span class="nb">set</span><span class="p">(</span><span class="n">QUANT_ALGO_LIST</span><span class="p">)</span> <span class="o">-</span> <span class="p">{</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A16</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8_PER_CHANNEL_PER_TOKEN</span>
|
|
<span class="p">})</span> <span class="ow">or</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span> <span class="ow">in</span> <span class="n">KV_CACHE_QUANT_ALGO_LIST</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">requires_modelopt_quantization</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">in</span> <span class="p">[</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">NVFP4</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16_AWQ</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A8_SQ_PER_CHANNEL</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">MIXED_PRECISION</span>
|
|
<span class="p">]:</span>
|
|
<span class="k">return</span> <span class="kc">True</span>
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">True</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
|
|
<div class="viewcode-block" id="QuantConfig.get_quant_cfg">
|
|
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.QuantConfig.get_quant_cfg">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">get_quant_cfg</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">module_name</span><span class="o">=</span><span class="kc">None</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="bp">self</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="QuantConfig.get_modelopt_qformat">
|
|
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.QuantConfig.get_modelopt_qformat">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">get_modelopt_qformat</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">algo_to_modelopt_map</span> <span class="o">=</span> <span class="p">{</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A16</span><span class="p">:</span> <span class="s2">"int8_wo"</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16</span><span class="p">:</span> <span class="s2">"int4_wo"</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">NVFP4</span><span class="p">:</span> <span class="s2">"nvfp4"</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span> <span class="s2">"fp8"</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16_AWQ</span><span class="p">:</span> <span class="s2">"int4_awq"</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span><span class="p">:</span> <span class="s2">"w4a8_awq"</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A8_SQ_PER_CHANNEL</span><span class="p">:</span> <span class="s2">"int8_sq"</span><span class="p">,</span>
|
|
<span class="p">}</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="o">!=</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">MIXED_PRECISION</span><span class="p">,</span> <span class="sa">f</span><span class="s2">"We don't support mixed precision in QuantConfig"</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">in</span> <span class="n">algo_to_modelopt_map</span><span class="p">,</span> <span class="sa">f</span><span class="s2">"We don't use Modelopt for quantization algorithm </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span><span class="si">}</span><span class="s2">, you probably shall not call this"</span>
|
|
<span class="k">return</span> <span class="n">algo_to_modelopt_map</span><span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span><span class="p">]</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="s1">'full_prec'</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="QuantConfig.get_modelopt_kv_cache_dtype">
|
|
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.QuantConfig.get_modelopt_kv_cache_dtype">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">get_modelopt_kv_cache_dtype</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">algo_to_modelopt_map</span> <span class="o">=</span> <span class="p">{</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span> <span class="s1">'fp8'</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">INT8</span><span class="p">:</span> <span class="s1">'int8'</span><span class="p">,</span>
|
|
<span class="p">}</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span> <span class="ow">in</span> <span class="n">algo_to_modelopt_map</span><span class="p">,</span> <span class="sa">f</span><span class="s2">"We don't use Modelopt for quantization algorithm </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="si">}</span><span class="s2">, you probably shall not call this"</span>
|
|
<span class="k">return</span> <span class="n">algo_to_modelopt_map</span><span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="p">]</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">None</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="QuantConfig.from_dict">
|
|
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.QuantConfig.from_dict">[docs]</a>
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_dict</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="nb">dict</span><span class="p">):</span>
|
|
<span class="n">obj</span> <span class="o">=</span> <span class="bp">cls</span><span class="p">(</span><span class="o">**</span><span class="n">config</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">obj</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="QuantConfig.to_dict">
|
|
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.QuantConfig.to_dict">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">to_dict</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="n">dataclasses</span><span class="o">.</span><span class="n">asdict</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span></div>
|
|
</div>
|
|
|
|
|
|
|
|
<span class="nd">@dataclasses</span><span class="o">.</span><span class="n">dataclass</span>
|
|
<span class="k">class</span> <span class="nc">LayerQuantConfig</span><span class="p">(</span><span class="n">QuantConfig</span><span class="p">):</span>
|
|
<span class="n">quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="n">kv_cache_quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="n">quantized_layers</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">QuantConfig</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="n">exclude_modules</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">List</span><span class="p">[</span><span class="nb">str</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span>
|
|
<span class="o">*</span><span class="p">,</span>
|
|
<span class="n">quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">kv_cache_quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">quantized_layers</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">QuantConfig</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">exclude_modules</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">List</span><span class="p">[</span><span class="nb">str</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="o">=</span> <span class="n">quant_algo</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">quantized_layers</span> <span class="o">=</span> <span class="n">quantized_layers</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span> <span class="o">=</span> <span class="n">kv_cache_quant_algo</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">exclude_modules</span> <span class="o">=</span> <span class="n">exclude_modules</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">auto_quant_mode</span> <span class="o">=</span> <span class="p">{}</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer_config</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantized_layers</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">auto_quant_mode</span><span class="o">.</span><span class="n">update</span><span class="p">({</span>
|
|
<span class="n">name</span><span class="p">:</span>
|
|
<span class="n">QuantMode</span><span class="o">.</span><span class="n">from_quant_algo</span><span class="p">(</span>
|
|
<span class="n">layer_config</span><span class="o">.</span><span class="n">quant_algo</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="p">})</span>
|
|
<span class="k">for</span> <span class="n">key</span> <span class="ow">in</span> <span class="n">kwargs</span><span class="p">:</span>
|
|
<span class="n">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Warning: Unrecognized parameter '</span><span class="si">{</span><span class="n">key</span><span class="si">}</span><span class="s2">' with value '</span><span class="si">{</span><span class="n">kwargs</span><span class="p">[</span><span class="n">key</span><span class="p">]</span><span class="si">}</span><span class="s2">'"</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="nd">@cached_property</span>
|
|
<span class="k">def</span> <span class="nf">quant_mode</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">quant_mode_list</span> <span class="o">=</span> <span class="nb">list</span><span class="p">(</span><span class="nb">set</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">auto_quant_mode</span><span class="o">.</span><span class="n">values</span><span class="p">()))</span>
|
|
<span class="k">return</span> <span class="n">QuantModeWrapper</span><span class="p">(</span><span class="n">quant_mode_list</span><span class="p">)</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">layer_quant_mode</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">QuantMode</span><span class="p">]:</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">auto_quant_mode</span>
|
|
|
|
<span class="nd">@cached_property</span>
|
|
<span class="k">def</span> <span class="nf">auto_quant_list</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">quant_list</span> <span class="o">=</span> <span class="p">[]</span>
|
|
<span class="k">for</span> <span class="n">_</span><span class="p">,</span> <span class="n">layer_config</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantized_layers</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="n">quant_list</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">layer_config</span><span class="o">.</span><span class="n">quant_algo</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="nb">list</span><span class="p">(</span><span class="nb">set</span><span class="p">(</span><span class="n">quant_list</span><span class="p">))</span>
|
|
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_dict</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="nb">dict</span><span class="p">):</span>
|
|
<span class="n">quantized_layers</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">'quantized_layers'</span><span class="p">,</span> <span class="p">{})</span>
|
|
|
|
<span class="n">quantized_layers_dict</span> <span class="o">=</span> <span class="p">{</span>
|
|
<span class="n">layer_name</span><span class="p">:</span> <span class="n">QuantConfig</span><span class="p">(</span><span class="o">**</span><span class="n">layer_config</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">layer_name</span><span class="p">,</span> <span class="n">layer_config</span> <span class="ow">in</span> <span class="n">quantized_layers</span><span class="o">.</span><span class="n">items</span><span class="p">()</span>
|
|
<span class="p">}</span>
|
|
|
|
<span class="n">obj</span> <span class="o">=</span> <span class="bp">cls</span><span class="p">(</span><span class="n">quantized_layers</span><span class="o">=</span><span class="n">quantized_layers_dict</span><span class="p">,</span> <span class="o">**</span><span class="n">config</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">obj</span>
|
|
|
|
<span class="k">def</span> <span class="nf">get_quant_cfg</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">module_name</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">module_name</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantized_layers</span><span class="o">.</span><span class="n">keys</span><span class="p">():</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantized_layers</span><span class="p">[</span><span class="n">module_name</span><span class="p">]</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">QuantConfig</span><span class="p">()</span>
|
|
|
|
<span class="k">def</span> <span class="nf">get_modelopt_qformat</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">algo_to_modelopt_map</span> <span class="o">=</span> <span class="p">{</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">NVFP4</span><span class="p">:</span> <span class="s2">"nvfp4"</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span> <span class="s2">"fp8"</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16_AWQ</span><span class="p">:</span> <span class="s2">"int4_awq"</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span><span class="p">:</span> <span class="s2">"w4a8_awq"</span><span class="p">,</span>
|
|
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A8_SQ_PER_CHANNEL</span><span class="p">:</span> <span class="s2">"int8_sq"</span><span class="p">,</span>
|
|
<span class="p">}</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">MIXED_PRECISION</span><span class="p">,</span> <span class="sa">f</span><span class="s2">"We only support mixed precision quantization in LayerQuantConfig"</span>
|
|
<span class="n">autoq_format</span> <span class="o">=</span> <span class="s1">','</span><span class="o">.</span><span class="n">join</span><span class="p">(</span>
|
|
<span class="p">[</span><span class="n">algo_to_modelopt_map</span><span class="p">[</span><span class="n">item</span><span class="p">]</span> <span class="k">for</span> <span class="n">item</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">auto_quant_list</span><span class="p">])</span>
|
|
<span class="k">return</span> <span class="n">autoq_format</span>
|
|
|
|
<span class="k">def</span> <span class="nf">to_dict</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">output</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="vm">__dict__</span><span class="p">)</span>
|
|
<span class="n">output</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">'auto_quant_mode'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="n">output</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">'quant_mode'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="n">output</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">'exclude_modules'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">per_layer_config</span> <span class="ow">in</span> <span class="n">output</span><span class="p">[</span><span class="s1">'quantized_layers'</span><span class="p">]</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="n">per_layer_config</span> <span class="o">=</span> <span class="n">per_layer_config</span><span class="o">.</span><span class="n">to_dict</span><span class="p">()</span>
|
|
<span class="n">per_layer_config</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">'exclude_modules'</span><span class="p">)</span>
|
|
<span class="n">output</span><span class="p">[</span><span class="s1">'quantized_layers'</span><span class="p">][</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">per_layer_config</span>
|
|
<span class="k">return</span> <span class="n">output</span>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">PretrainedConfig</span><span class="p">:</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span>
|
|
<span class="o">*</span><span class="p">,</span>
|
|
<span class="n">architecture</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
|
|
<span class="n">hidden_size</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
|
|
<span class="n">num_hidden_layers</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
|
|
<span class="n">num_attention_heads</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
|
|
<span class="n">vocab_size</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">hidden_act</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">'gelu'</span><span class="p">,</span>
|
|
<span class="n">logits_dtype</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">'float32'</span><span class="p">,</span>
|
|
<span class="n">norm_epsilon</span><span class="p">:</span> <span class="nb">float</span> <span class="o">=</span> <span class="mf">1e-5</span><span class="p">,</span>
|
|
<span class="n">position_embedding_type</span><span class="p">:</span> <span class="n">Union</span><span class="p">[</span>
|
|
<span class="n">PositionEmbeddingType</span><span class="p">,</span>
|
|
<span class="nb">str</span><span class="p">]</span> <span class="o">=</span> <span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">learned_absolute</span><span class="p">,</span>
|
|
<span class="n">max_position_embeddings</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_dim</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">num_key_value_heads</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">intermediate_size</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">mapping</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Union</span><span class="p">[</span><span class="n">Mapping</span><span class="p">,</span> <span class="nb">dict</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">quantization</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Union</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">,</span> <span class="nb">dict</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">use_parallel_embedding</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">embedding_sharding_dim</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">head_size</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">qk_layernorm</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">runtime_defaults</span><span class="p">:</span> <span class="s2">"RuntimeDefaultsIn"</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">architecture</span> <span class="o">=</span> <span class="n">architecture</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">dtype</span> <span class="o">=</span> <span class="n">dtype</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">vocab_size</span> <span class="o">=</span> <span class="n">vocab_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">hidden_size</span> <span class="o">=</span> <span class="n">hidden_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_hidden_layers</span> <span class="o">=</span> <span class="n">num_hidden_layers</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">=</span> <span class="n">num_attention_heads</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">hidden_act</span> <span class="o">=</span> <span class="n">hidden_act</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">logits_dtype</span> <span class="o">=</span> <span class="n">logits_dtype</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">norm_epsilon</span> <span class="o">=</span> <span class="n">norm_epsilon</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">runtime_defaults</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">create_runtime_defaults</span><span class="p">(</span><span class="n">runtime_defaults</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">position_embedding_type</span><span class="p">,</span> <span class="nb">str</span><span class="p">):</span>
|
|
<span class="n">position_embedding_type</span> <span class="o">=</span> <span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">from_string</span><span class="p">(</span>
|
|
<span class="n">position_embedding_type</span><span class="p">)</span>
|
|
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">position_embedding_type</span><span class="p">,</span> <span class="n">PositionEmbeddingType</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span> <span class="o">=</span> <span class="n">position_embedding_type</span>
|
|
|
|
<span class="k">if</span> <span class="n">num_key_value_heads</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">num_key_value_heads</span> <span class="o">=</span> <span class="n">num_attention_heads</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_key_value_heads</span> <span class="o">=</span> <span class="n">num_key_value_heads</span>
|
|
|
|
<span class="k">if</span> <span class="n">intermediate_size</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">intermediate_size</span> <span class="o">=</span> <span class="n">hidden_size</span> <span class="o">*</span> <span class="mi">4</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">intermediate_size</span> <span class="o">=</span> <span class="n">intermediate_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span> <span class="o">=</span> <span class="n">max_position_embeddings</span>
|
|
|
|
<span class="k">if</span> <span class="n">mapping</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">mapping</span> <span class="o">=</span> <span class="n">Mapping</span><span class="p">()</span>
|
|
<span class="k">elif</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">mapping</span><span class="p">,</span> <span class="nb">dict</span><span class="p">):</span>
|
|
<span class="n">mapping</span> <span class="o">=</span> <span class="n">Mapping</span><span class="o">.</span><span class="n">from_dict</span><span class="p">(</span><span class="n">mapping</span><span class="p">)</span>
|
|
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">mapping</span><span class="p">,</span> <span class="n">Mapping</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">mapping</span> <span class="o">=</span> <span class="n">mapping</span>
|
|
|
|
<span class="k">if</span> <span class="n">quantization</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">quantization</span> <span class="o">=</span> <span class="n">QuantConfig</span><span class="p">()</span>
|
|
<span class="k">elif</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">quantization</span><span class="p">,</span> <span class="nb">dict</span><span class="p">):</span>
|
|
<span class="n">quantization</span> <span class="o">=</span> <span class="n">QuantConfig</span><span class="o">.</span><span class="n">from_dict</span><span class="p">(</span><span class="n">quantization</span><span class="p">)</span>
|
|
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">quantization</span><span class="p">,</span> <span class="n">QuantConfig</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">quantization</span> <span class="o">=</span> <span class="n">quantization</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">use_parallel_embedding</span> <span class="o">=</span> <span class="n">use_parallel_embedding</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">embedding_sharding_dim</span> <span class="o">=</span> <span class="n">embedding_sharding_dim</span>
|
|
|
|
<span class="k">if</span> <span class="n">head_size</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">head_size</span> <span class="o">=</span> <span class="n">hidden_size</span> <span class="o">//</span> <span class="n">num_attention_heads</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">head_size</span> <span class="o">=</span> <span class="n">head_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">qk_layernorm</span> <span class="o">=</span> <span class="n">qk_layernorm</span>
|
|
|
|
<span class="k">if</span> <span class="n">rotary_embedding_dim</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">rotary_embedding_percentage</span> <span class="o">=</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s1">'rotary_pct'</span><span class="p">,</span> <span class="mf">1.0</span><span class="p">)</span>
|
|
<span class="n">rotary_embedding_dim</span> <span class="o">=</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">get</span><span class="p">(</span>
|
|
<span class="s1">'rotary_dim'</span><span class="p">,</span> <span class="nb">int</span><span class="p">(</span><span class="n">head_size</span> <span class="o">*</span> <span class="n">rotary_embedding_percentage</span><span class="p">))</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span> <span class="o">=</span> <span class="n">rotary_embedding_dim</span>
|
|
|
|
<span class="k">for</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span> <span class="ow">in</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="k">try</span><span class="p">:</span>
|
|
<span class="nb">setattr</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span><span class="p">)</span>
|
|
<span class="n">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Implicitly setting </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="vm">__class__</span><span class="o">.</span><span class="vm">__name__</span><span class="si">}</span><span class="s2">.</span><span class="si">{</span><span class="n">key</span><span class="si">}</span><span class="s2"> = </span><span class="si">{</span><span class="n">value</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
<span class="k">except</span> <span class="ne">AttributeError</span> <span class="k">as</span> <span class="n">err</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="n">err</span>
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.create_runtime_defaults">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.create_runtime_defaults">[docs]</a>
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">create_runtime_defaults</span><span class="p">(</span>
|
|
<span class="n">defaults</span><span class="p">:</span> <span class="s2">"RuntimeDefaultsIn"</span> <span class="o">=</span> <span class="kc">None</span><span class="p">)</span> <span class="o">-></span> <span class="n">Optional</span><span class="p">[</span><span class="n">RuntimeDefaults</span><span class="p">]:</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">defaults</span><span class="p">,</span> <span class="nb">dict</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="n">RuntimeDefaults</span><span class="p">(</span><span class="o">**</span><span class="n">defaults</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">defaults</span></div>
|
|
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">kv_dtype</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="c1"># TODO: need to align the kv dtype</span>
|
|
<span class="c1"># now assume the kv cache is for all layers</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="o">.</span><span class="n">has_int8_kv_cache</span><span class="p">():</span>
|
|
<span class="k">return</span> <span class="s1">'int8'</span>
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="o">.</span><span class="n">has_fp8_kv_cache</span><span class="p">():</span>
|
|
<span class="k">return</span> <span class="s1">'fp8'</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">dtype</span>
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.set_if_not_exist">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.set_if_not_exist">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">set_if_not_exist</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="nb">hasattr</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">key</span><span class="p">):</span>
|
|
<span class="nb">setattr</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span><span class="p">)</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.from_dict">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.from_dict">[docs]</a>
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_dict</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="nb">dict</span><span class="p">):</span>
|
|
<span class="c1"># Maybe we need AutoConfig for this</span>
|
|
<span class="kn">from</span> <span class="nn">.</span> <span class="kn">import</span> <span class="n">MODEL_MAP</span>
|
|
<span class="n">model_cls</span> <span class="o">=</span> <span class="n">MODEL_MAP</span><span class="p">[</span><span class="n">config</span><span class="p">[</span><span class="s1">'architecture'</span><span class="p">]]</span>
|
|
<span class="n">config_cls</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">model_cls</span><span class="p">,</span> <span class="s1">'config_class'</span><span class="p">,</span> <span class="bp">cls</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">config_cls</span><span class="p">(</span><span class="o">**</span><span class="n">config</span><span class="p">)</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.to_dict">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.to_dict">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">to_dict</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">output</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="vm">__dict__</span><span class="p">)</span>
|
|
|
|
<span class="n">output</span><span class="p">[</span><span class="s1">'position_embedding_type'</span><span class="p">]</span> <span class="o">=</span> <span class="nb">str</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span><span class="p">)</span>
|
|
<span class="n">output</span><span class="p">[</span><span class="s1">'mapping'</span><span class="p">]</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">to_dict</span><span class="p">()</span>
|
|
<span class="n">output</span><span class="p">[</span><span class="s1">'mapping'</span><span class="p">]</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">'rank'</span><span class="p">)</span>
|
|
<span class="n">output</span><span class="p">[</span><span class="s1">'quantization'</span><span class="p">]</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">to_dict</span><span class="p">()</span>
|
|
|
|
<span class="k">return</span> <span class="n">output</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.from_json_file">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.from_json_file">[docs]</a>
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_json_file</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">config_file</span><span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
|
|
<span class="k">with</span> <span class="nb">open</span><span class="p">(</span><span class="n">config_file</span><span class="p">)</span> <span class="k">as</span> <span class="n">f</span><span class="p">:</span>
|
|
<span class="n">config</span> <span class="o">=</span> <span class="n">json</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="n">f</span><span class="p">)</span>
|
|
<span class="n">obj</span> <span class="o">=</span> <span class="bp">cls</span><span class="o">.</span><span class="n">from_dict</span><span class="p">(</span><span class="n">config</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">obj</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">MIXED_PRECISION</span><span class="p">:</span>
|
|
<span class="k">try</span><span class="p">:</span>
|
|
<span class="n">layer_config_path</span> <span class="o">=</span> <span class="nb">str</span><span class="p">(</span><span class="n">config_file</span><span class="p">)</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span>
|
|
<span class="s1">'config.json'</span><span class="p">,</span> <span class="s1">'quant_cfg.json'</span><span class="p">)</span>
|
|
<span class="n">obj</span><span class="o">.</span><span class="n">to_layer_quant_config</span><span class="p">(</span><span class="n">layer_config_path</span><span class="p">)</span>
|
|
<span class="k">except</span> <span class="ne">Exception</span> <span class="k">as</span> <span class="n">e</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Encounter error '</span><span class="si">{</span><span class="n">e</span><span class="si">}</span><span class="s2">' for read quantization config '</span><span class="si">{</span><span class="n">layer_config_path</span><span class="si">}</span><span class="s2">'"</span>
|
|
<span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">obj</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.from_checkpoint">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.from_checkpoint">[docs]</a>
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_checkpoint</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">ckpt_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="bp">cls</span><span class="o">.</span><span class="n">from_json_file</span><span class="p">(</span><span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">ckpt_dir</span><span class="p">,</span> <span class="s1">'config.json'</span><span class="p">))</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.to_json_file">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.to_json_file">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">to_json_file</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config_file</span><span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
|
|
<span class="k">with</span> <span class="nb">open</span><span class="p">(</span><span class="n">config_file</span><span class="p">,</span> <span class="s1">'w'</span><span class="p">)</span> <span class="k">as</span> <span class="n">f</span><span class="p">:</span>
|
|
<span class="n">json</span><span class="o">.</span><span class="n">dump</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">to_dict</span><span class="p">(),</span> <span class="n">f</span><span class="p">,</span> <span class="n">indent</span><span class="o">=</span><span class="mi">4</span><span class="p">)</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.to_layer_quant_config">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.to_layer_quant_config">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">to_layer_quant_config</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config_file</span><span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
|
|
<span class="k">with</span> <span class="nb">open</span><span class="p">(</span><span class="n">config_file</span><span class="p">)</span> <span class="k">as</span> <span class="n">f</span><span class="p">:</span>
|
|
<span class="n">config</span> <span class="o">=</span> <span class="n">json</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="n">f</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">quantization</span> <span class="o">=</span> <span class="n">LayerQuantConfig</span><span class="o">.</span><span class="n">from_dict</span><span class="p">(</span><span class="n">config</span><span class="p">)</span></div>
|
|
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">quant_mode</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">quant_mode</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">quant_algo</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">quant_algo</span>
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.get_quant_cfg">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.get_quant_cfg">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">get_quant_cfg</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">module_name</span><span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">get_quant_cfg</span><span class="p">(</span><span class="n">module_name</span><span class="p">)</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.set_rank">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.set_rank">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">set_rank</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">rank</span><span class="p">:</span> <span class="nb">int</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">rank</span> <span class="o">=</span> <span class="n">rank</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.get_config_group">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.get_config_group">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">get_config_group</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">group_cls</span><span class="p">:</span> <span class="s2">"Type[CG]"</span><span class="p">)</span> <span class="o">-></span> <span class="s2">"CG"</span><span class="p">:</span>
|
|
<span class="n">cfg</span> <span class="o">=</span> <span class="p">{</span><span class="n">k</span><span class="p">:</span> <span class="n">v</span> <span class="k">for</span> <span class="n">k</span><span class="p">,</span> <span class="n">v</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">to_dict</span><span class="p">()</span><span class="o">.</span><span class="n">items</span><span class="p">()</span> <span class="k">if</span> <span class="n">k</span> <span class="ow">in</span> <span class="n">group_cls</span><span class="o">.</span><span class="n">keys</span><span class="p">()}</span>
|
|
<span class="k">return</span> <span class="n">group_cls</span><span class="p">(</span><span class="o">**</span><span class="n">cfg</span><span class="p">)</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.has_config_group">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.has_config_group">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">has_config_group</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">group_cls</span><span class="p">:</span> <span class="s2">"Type[CG]"</span><span class="p">)</span> <span class="o">-></span> <span class="s2">"bool"</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="nb">all</span><span class="p">(</span><span class="nb">hasattr</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">key</span><span class="p">)</span> <span class="k">for</span> <span class="n">key</span> <span class="ow">in</span> <span class="n">group_cls</span><span class="o">.</span><span class="n">keys</span><span class="p">())</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedConfig.for_each_rank">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.for_each_rank">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">for_each_rank</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="s2">"Generator[Self, None, None]"</span><span class="p">:</span>
|
|
<span class="k">for</span> <span class="n">rank</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">world_size</span><span class="p">):</span>
|
|
<span class="n">config_copy</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span>
|
|
<span class="n">config_copy</span><span class="o">.</span><span class="n">set_rank</span><span class="p">(</span><span class="n">rank</span><span class="p">)</span>
|
|
<span class="k">yield</span> <span class="n">config_copy</span></div>
|
|
</div>
|
|
|
|
|
|
|
|
<span class="k">class</span> <span class="nc">DecoderLayerList</span><span class="p">(</span><span class="n">ModuleList</span><span class="p">):</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="bp">cls</span><span class="p">,</span> <span class="n">config</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_hidden_layers</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">num_hidden_layers</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">pp_layers</span><span class="p">(</span><span class="n">config</span><span class="o">.</span><span class="n">num_hidden_layers</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">quant_mode</span>
|
|
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">([</span><span class="bp">cls</span><span class="p">(</span><span class="n">config</span><span class="p">,</span> <span class="n">idx</span><span class="p">)</span> <span class="k">for</span> <span class="n">idx</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span><span class="p">])</span>
|
|
|
|
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span>
|
|
<span class="n">hidden_states</span><span class="p">,</span>
|
|
<span class="n">use_cache</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">attention_mask</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">kv_cache_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">attention_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">mrope_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">position_ids</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">lora_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">vision_token_mask</span><span class="o">=</span><span class="kc">None</span><span class="p">):</span>
|
|
<span class="n">kv_cache_params</span><span class="o">.</span><span class="n">fill_none_tensor_list</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span><span class="p">))</span>
|
|
|
|
<span class="k">if</span> <span class="n">use_cache</span><span class="p">:</span>
|
|
<span class="n">presents</span> <span class="o">=</span> <span class="p">[]</span>
|
|
|
|
<span class="k">for</span> <span class="n">layer_idx</span><span class="p">,</span> <span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">past</span><span class="p">)</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span>
|
|
<span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">kv_cache_params</span><span class="o">.</span><span class="n">past_key_value</span><span class="p">)):</span>
|
|
|
|
<span class="n">lora_layer_params</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="n">lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">lora_params</span><span class="o">.</span><span class="n">lora_ranks</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">lora_layer_params</span> <span class="o">=</span> <span class="n">lora_params</span><span class="o">.</span><span class="n">get_layer_params</span><span class="p">(</span><span class="n">layer_idx</span><span class="p">)</span>
|
|
|
|
<span class="n">kwargs</span> <span class="o">=</span> <span class="p">{}</span>
|
|
<span class="k">if</span> <span class="n">position_ids</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'position_ids'</span><span class="p">]</span> <span class="o">=</span> <span class="n">position_ids</span>
|
|
<span class="k">if</span> <span class="n">vision_token_mask</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'vision_token_mask'</span><span class="p">]</span> <span class="o">=</span> <span class="n">vision_token_mask</span>
|
|
<span class="k">if</span> <span class="n">lora_layer_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'lora_layer_params'</span><span class="p">]</span> <span class="o">=</span> <span class="n">lora_layer_params</span>
|
|
<span class="k">if</span> <span class="n">spec_decoding_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'spec_decoding_params'</span><span class="p">]</span> <span class="o">=</span> <span class="n">spec_decoding_params</span>
|
|
<span class="k">if</span> <span class="n">mrope_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'mrope_params'</span><span class="p">]</span> <span class="o">=</span> <span class="n">mrope_params</span>
|
|
<span class="k">if</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">reduce_fusion</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="n">layer_idx</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o"><</span> <span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span><span class="p">[</span><span class="o">-</span><span class="mi">1</span><span class="p">]:</span>
|
|
<span class="k">if</span> <span class="n">default_net</span><span class="p">(</span>
|
|
<span class="p">)</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">user_buffer</span> <span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="o">.</span><span class="n">has_fp8_qdq</span><span class="p">(</span>
|
|
<span class="p">):</span>
|
|
<span class="n">qkv_activation_scaling_factor</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="p">[</span><span class="n">layer_idx</span> <span class="o">+</span> <span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">attention</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span>
|
|
<span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="o">.</span><span class="n">copy</span><span class="p">())</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">qkv_activation_scaling_factor</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'next_layer_input_layernorm_args'</span><span class="p">]</span> <span class="o">=</span> <span class="p">(</span>
|
|
<span class="bp">self</span><span class="p">[</span><span class="n">layer_idx</span> <span class="o">+</span> <span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">input_layernorm</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="p">[</span><span class="n">layer_idx</span> <span class="o">+</span> <span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">input_layernorm</span><span class="o">.</span><span class="n">eps</span><span class="p">,</span>
|
|
<span class="n">qkv_activation_scaling_factor</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'next_layer_input_layernorm_args'</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">layer</span><span class="p">(</span>
|
|
<span class="n">hidden_states</span><span class="p">,</span>
|
|
<span class="n">use_cache</span><span class="o">=</span><span class="n">use_cache</span><span class="p">,</span>
|
|
<span class="n">attention_mask</span><span class="o">=</span><span class="n">attention_mask</span><span class="p">,</span>
|
|
<span class="n">kv_cache_params</span><span class="o">=</span><span class="n">KeyValueCacheParams</span><span class="p">(</span>
|
|
<span class="n">past_key_value</span><span class="o">=</span><span class="p">[</span><span class="n">past</span><span class="p">],</span>
|
|
<span class="n">host_past_key_value_lengths</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_past_key_value_lengths</span><span class="p">,</span>
|
|
<span class="n">host_max_attention_window_sizes</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_max_attention_window_sizes</span><span class="p">,</span>
|
|
<span class="n">host_sink_token_length</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_sink_token_length</span><span class="p">,</span>
|
|
<span class="n">kv_cache_block_offsets</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">kv_cache_block_offsets</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_block_offsets</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_block_offsets</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_pool_pointers</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_pool_pointers</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_pool_mapping</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_pool_mapping</span><span class="p">,</span>
|
|
<span class="n">cache_indirection</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span><span class="n">cache_indirection</span><span class="p">),</span>
|
|
<span class="n">attention_params</span><span class="o">=</span><span class="n">attention_params</span><span class="p">,</span>
|
|
<span class="o">**</span><span class="n">kwargs</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">use_cache</span><span class="p">:</span>
|
|
<span class="n">presents</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">[</span><span class="mi">1</span><span class="p">])</span>
|
|
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">hidden_states</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
|
|
|
|
<span class="k">if</span> <span class="n">use_cache</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">hidden_states</span><span class="p">,</span> <span class="n">presents</span>
|
|
<span class="k">return</span> <span class="n">hidden_states</span>
|
|
|
|
|
|
<span class="k">class</span> <span class="nc">PostInitCaller</span><span class="p">(</span><span class="nb">type</span><span class="p">):</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__call__</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="o">*</span><span class="n">args</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
|
|
<span class="n">obj</span> <span class="o">=</span> <span class="nb">type</span><span class="o">.</span><span class="fm">__call__</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="o">*</span><span class="n">args</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">)</span>
|
|
<span class="n">obj</span><span class="o">.</span><span class="n">__post_init__</span><span class="p">()</span>
|
|
<span class="k">return</span> <span class="n">obj</span>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedModel">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">PretrainedModel</span><span class="p">(</span><span class="n">Module</span><span class="p">,</span>
|
|
<span class="n">GenerationMixin</span><span class="p">,</span>
|
|
<span class="n">TopModelMixin</span><span class="p">,</span>
|
|
<span class="n">metaclass</span><span class="o">=</span><span class="n">PostInitCaller</span><span class="p">):</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="n">PretrainedConfig</span><span class="p">):</span>
|
|
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
|
|
<span class="n">init_all_reduce_helper</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">config</span> <span class="o">=</span> <span class="n">config</span>
|
|
|
|
<span class="k">def</span> <span class="nf">__post_init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="kn">from</span> <span class="nn">..quantization.quantize</span> <span class="kn">import</span> <span class="n">quantize</span>
|
|
<span class="n">quantize</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="p">)</span>
|
|
|
|
<span class="c1"># Currently, use_parallel_embedding must be enabled before weight loading;</span>
|
|
<span class="c1"># otherwise, the model will be inconsistent with the weights loaded from checkpoint.</span>
|
|
<span class="n">optimize_model</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="p">,</span> <span class="n">use_parallel_embedding</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">use_parallel_embedding</span><span class="p">)</span>
|
|
|
|
<div class="viewcode-block" id="PretrainedModel.release">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.release">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">release</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">release_gc</span><span class="p">()</span></div>
|
|
|
|
|
|
<span class="k">def</span> <span class="fm">__del__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">release</span><span class="p">()</span>
|
|
|
|
<div class="viewcode-block" id="PretrainedModel.check_config">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.check_config">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">check_config</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config</span><span class="p">):</span>
|
|
<span class="k">raise</span> <span class="ne">NotImplementedError</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"</span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="vm">__class__</span><span class="si">}</span><span class="s2"> is an abstract class. Only classes inheriting this class can be called."</span>
|
|
<span class="p">)</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedModel.from_config">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.from_config">[docs]</a>
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_config</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="n">PretrainedConfig</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="bp">cls</span><span class="p">(</span><span class="n">config</span><span class="p">)</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedModel.from_checkpoint">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.from_checkpoint">[docs]</a>
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_checkpoint</span><span class="p">(</span>
|
|
<span class="bp">cls</span><span class="p">,</span>
|
|
<span class="n">ckpt_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
|
|
<span class="n">rank</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">PretrainedConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="o">*</span><span class="p">,</span>
|
|
<span class="n">preprocess_weights_hook</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Callable</span><span class="p">[[</span><span class="n">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">Tensor</span><span class="p">]],</span>
|
|
<span class="n">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">Tensor</span><span class="p">]]]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">config</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">config</span> <span class="o">=</span> <span class="n">PretrainedConfig</span><span class="o">.</span><span class="n">from_json_file</span><span class="p">(</span>
|
|
<span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">ckpt_dir</span><span class="p">,</span> <span class="s1">'config.json'</span><span class="p">))</span>
|
|
|
|
<span class="k">if</span> <span class="n">rank</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">config</span><span class="o">.</span><span class="n">set_rank</span><span class="p">(</span><span class="n">rank</span><span class="p">)</span>
|
|
|
|
<span class="n">rank</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">rank</span>
|
|
<span class="k">if</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">auto_parallel</span><span class="p">:</span>
|
|
<span class="n">rank</span> <span class="o">=</span> <span class="mi">0</span>
|
|
<span class="k">elif</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">></span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="c1"># tp_cp_pp rank -> tp_pp rank: because different cp ranks share the same ckpt</span>
|
|
<span class="n">tp_size</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_size</span>
|
|
<span class="n">cp_size</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span>
|
|
<span class="n">rank</span> <span class="o">=</span> <span class="n">rank</span> <span class="o">%</span> <span class="n">tp_size</span> <span class="o">+</span> <span class="n">rank</span> <span class="o">//</span> <span class="p">(</span><span class="n">tp_size</span> <span class="o">*</span> <span class="n">cp_size</span><span class="p">)</span> <span class="o">*</span> <span class="n">tp_size</span>
|
|
<span class="n">weights_path</span> <span class="o">=</span> <span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">ckpt_dir</span><span class="p">,</span> <span class="sa">f</span><span class="s1">'rank</span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s1">.safetensors'</span><span class="p">)</span>
|
|
|
|
<span class="k">assert</span> <span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">isfile</span><span class="p">(</span><span class="n">weights_path</span><span class="p">)</span>
|
|
<span class="n">weights</span> <span class="o">=</span> <span class="n">safetensors</span><span class="o">.</span><span class="n">torch</span><span class="o">.</span><span class="n">load_file</span><span class="p">(</span><span class="n">weights_path</span><span class="p">)</span>
|
|
<span class="n">is_checkpoint_pruned</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">config</span><span class="p">,</span> <span class="s1">'is_pruned'</span><span class="p">,</span> <span class="kc">False</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">preprocess_weights_hook</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">weights</span> <span class="o">=</span> <span class="n">preprocess_weights_hook</span><span class="p">(</span><span class="n">weights</span><span class="p">)</span>
|
|
|
|
<span class="n">weights</span> <span class="o">=</span> <span class="n">preprocess_weights</span><span class="p">(</span><span class="n">weights</span><span class="p">,</span>
|
|
<span class="n">config</span><span class="p">,</span>
|
|
<span class="n">from_pruned</span><span class="o">=</span><span class="n">is_checkpoint_pruned</span><span class="p">)</span>
|
|
<span class="n">model</span> <span class="o">=</span> <span class="bp">cls</span><span class="p">(</span><span class="n">config</span><span class="p">)</span>
|
|
<span class="n">model</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="n">weights</span><span class="p">,</span> <span class="n">from_pruned</span><span class="o">=</span><span class="n">is_checkpoint_pruned</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">model</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedModel.load">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.load">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">load</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">weights</span><span class="p">,</span> <span class="n">from_pruned</span><span class="o">=</span><span class="kc">False</span><span class="p">):</span>
|
|
<span class="n">required_names</span> <span class="o">=</span> <span class="nb">set</span><span class="p">()</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">named_parameters</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">param</span><span class="o">.</span><span class="n">is_inited</span><span class="p">():</span>
|
|
<span class="k">continue</span>
|
|
<span class="k">if</span> <span class="n">name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">weights</span><span class="p">:</span>
|
|
<span class="c1"># Exemption for embedding sharing</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'lm_head.weight'</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">any</span><span class="p">(</span>
|
|
<span class="n">k</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'vocab_embedding.weight'</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">k</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">keys</span><span class="p">()):</span>
|
|
<span class="k">continue</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'lm_head.per_channel_scale'</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">any</span><span class="p">(</span>
|
|
<span class="n">k</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'vocab_embedding.per_channel_scale'</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">k</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">keys</span><span class="p">()):</span>
|
|
<span class="k">continue</span>
|
|
<span class="n">required_names</span><span class="o">.</span><span class="n">add</span><span class="p">(</span><span class="n">name</span><span class="p">)</span>
|
|
|
|
<span class="n">provided_names</span> <span class="o">=</span> <span class="nb">set</span><span class="p">(</span><span class="n">weights</span><span class="o">.</span><span class="n">keys</span><span class="p">())</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">required_names</span><span class="o">.</span><span class="n">issubset</span><span class="p">(</span><span class="n">provided_names</span><span class="p">):</span>
|
|
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Required but not provided tensors:</span><span class="si">{</span><span class="n">required_names</span><span class="o">.</span><span class="n">difference</span><span class="p">(</span><span class="n">provided_names</span><span class="p">)</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">provided_names</span><span class="o">.</span><span class="n">issubset</span><span class="p">(</span><span class="n">required_names</span><span class="p">):</span>
|
|
<span class="n">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Provided but not required tensors: </span><span class="si">{</span><span class="n">provided_names</span><span class="o">.</span><span class="n">difference</span><span class="p">(</span><span class="n">required_names</span><span class="p">)</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">named_parameters</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">name</span> <span class="ow">in</span> <span class="n">provided_names</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">from_pruned</span><span class="p">:</span>
|
|
<span class="k">try</span><span class="p">:</span>
|
|
<span class="n">param</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span>
|
|
<span class="k">except</span> <span class="ne">Exception</span> <span class="k">as</span> <span class="n">e</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Encounter error '</span><span class="si">{</span><span class="n">e</span><span class="si">}</span><span class="s2">' for parameter '</span><span class="si">{</span><span class="n">name</span><span class="si">}</span><span class="s2">'"</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">param</span><span class="o">.</span><span class="n">set_value_or_dummy</span><span class="p">(</span><span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">])</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedModel.save_checkpoint">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.save_checkpoint">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">save_checkpoint</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">output_dir</span><span class="p">,</span> <span class="n">save_config</span><span class="o">=</span><span class="kc">True</span><span class="p">):</span>
|
|
<span class="c1"># multiple ranks could share same config.json, so adding a save_config parameter to let user avoiding writing config.json in all ranks</span>
|
|
<span class="n">rank</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">rank</span>
|
|
<span class="n">weights</span> <span class="o">=</span> <span class="p">{</span>
|
|
<span class="n">name</span><span class="p">:</span> <span class="n">numpy_to_torch</span><span class="p">(</span><span class="n">param</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">named_parameters</span><span class="p">()</span>
|
|
<span class="p">}</span>
|
|
<span class="c1"># If there are some tensors share memory, this will lead to error when we call "save_file". So, for repeated tensors, we</span>
|
|
<span class="c1"># clone the tensors to prevent this issue.</span>
|
|
<span class="n">data_ptrs</span> <span class="o">=</span> <span class="nb">set</span><span class="p">()</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">param</span><span class="o">.</span><span class="n">data_ptr</span><span class="p">()</span> <span class="ow">in</span> <span class="n">data_ptrs</span><span class="p">:</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span><span class="o">.</span><span class="n">clone</span><span class="p">()</span>
|
|
<span class="n">data_ptrs</span><span class="o">.</span><span class="n">add</span><span class="p">(</span><span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span><span class="o">.</span><span class="n">data_ptr</span><span class="p">())</span>
|
|
<span class="n">safetensors</span><span class="o">.</span><span class="n">torch</span><span class="o">.</span><span class="n">save_file</span><span class="p">(</span>
|
|
<span class="n">weights</span><span class="p">,</span> <span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">output_dir</span><span class="p">,</span> <span class="sa">f</span><span class="s1">'rank</span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s1">.safetensors'</span><span class="p">))</span>
|
|
<span class="k">if</span> <span class="n">save_config</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">to_json_file</span><span class="p">(</span><span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">output_dir</span><span class="p">,</span> <span class="s1">'config.json'</span><span class="p">))</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedModel.prepare_inputs">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.prepare_inputs">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">prepare_inputs</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="p">,</span>
|
|
<span class="n">max_batch_size</span><span class="p">,</span>
|
|
<span class="n">max_input_len</span><span class="p">,</span>
|
|
<span class="n">max_seq_len</span><span class="p">,</span>
|
|
<span class="n">max_num_tokens</span><span class="p">,</span>
|
|
<span class="n">use_cache</span><span class="p">,</span>
|
|
<span class="n">max_beam_width</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">opt_num_tokens</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">prompt_embedding_table_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">position_encoding_2d</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">max_draft_len</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">speculative_decoding_draft_tokens_external</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_is_generation_length_variable</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">gather_context_logits</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">gather_generation_logits</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">lora_target_modules</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="nb">str</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">opt_batch_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">num_hidden_layers</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">mrope_rotary_cos_sin_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">'''@brief: Prepare inputs Tensors for the model, the given sizes are used to determine the</span>
|
|
<span class="sd"> ranges of the dimensions of when using TRT dynamic shapes.</span>
|
|
|
|
<span class="sd"> @return: a list contains values which can be fed into the self.forward()</span>
|
|
<span class="sd"> '''</span>
|
|
|
|
<span class="c1"># Prepare inputs</span>
|
|
<span class="n">remove_input_padding</span> <span class="o">=</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">remove_input_padding</span>
|
|
<span class="n">use_gpt_attention_plugin</span> <span class="o">=</span> <span class="n">default_net</span><span class="p">(</span>
|
|
<span class="p">)</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">gpt_attention_plugin</span>
|
|
<span class="n">use_gemm_plugin</span> <span class="o">=</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">gemm_plugin</span>
|
|
<span class="n">paged_kv_cache</span> <span class="o">=</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">paged_kv_cache</span>
|
|
<span class="n">tokens_per_block</span> <span class="o">=</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">tokens_per_block</span>
|
|
<span class="n">use_lora_plugin</span> <span class="o">=</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">lora_plugin</span>
|
|
<span class="n">multiple_profiles</span> <span class="o">=</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">multiple_profiles</span>
|
|
<span class="n">streamingllm</span> <span class="o">=</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">streamingllm</span>
|
|
<span class="n">pp_reduce_scatter</span> <span class="o">=</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">pp_reduce_scatter</span>
|
|
|
|
<span class="n">kv_cache_type</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">use_cache</span><span class="p">:</span>
|
|
<span class="n">kv_cache_type</span> <span class="o">=</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">DISABLED</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="n">paged_kv_cache</span><span class="p">:</span>
|
|
<span class="n">kv_cache_type</span> <span class="o">=</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">PAGED</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">kv_cache_type</span> <span class="o">=</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">CONTINUOUS</span>
|
|
|
|
<span class="n">model_inputs</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">prepare_basic_inputs</span><span class="p">(</span>
|
|
<span class="n">max_batch_size</span><span class="o">=</span><span class="n">max_batch_size</span><span class="p">,</span>
|
|
<span class="n">max_beam_width</span><span class="o">=</span><span class="n">max_beam_width</span><span class="p">,</span>
|
|
<span class="n">max_input_len</span><span class="o">=</span><span class="n">max_input_len</span><span class="p">,</span>
|
|
<span class="n">max_seq_len</span><span class="o">=</span><span class="n">max_seq_len</span><span class="p">,</span>
|
|
<span class="n">hidden_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">num_kv_heads</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">num_key_value_heads</span><span class="p">,</span>
|
|
<span class="n">head_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">head_size</span><span class="p">,</span>
|
|
<span class="n">num_layers</span><span class="o">=</span><span class="n">num_hidden_layers</span>
|
|
<span class="k">if</span> <span class="n">num_hidden_layers</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="k">else</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">num_hidden_layers</span><span class="p">,</span>
|
|
<span class="n">kv_dtype</span><span class="o">=</span><span class="n">str_dtype_to_trt</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">kv_dtype</span><span class="p">),</span>
|
|
<span class="n">remove_input_padding</span><span class="o">=</span><span class="n">remove_input_padding</span><span class="p">,</span>
|
|
<span class="n">use_gpt_attention_plugin</span><span class="o">=</span><span class="n">use_gpt_attention_plugin</span><span class="p">,</span>
|
|
<span class="n">use_gemm_plugin</span><span class="o">=</span><span class="n">use_gemm_plugin</span><span class="p">,</span>
|
|
<span class="n">kv_cache_type</span><span class="o">=</span><span class="n">kv_cache_type</span><span class="p">,</span>
|
|
<span class="n">tokens_per_block</span><span class="o">=</span><span class="n">tokens_per_block</span><span class="p">,</span>
|
|
<span class="n">num_heads</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">,</span>
|
|
<span class="n">max_num_tokens</span><span class="o">=</span><span class="n">max_num_tokens</span><span class="p">,</span>
|
|
<span class="n">opt_num_tokens</span><span class="o">=</span><span class="n">opt_num_tokens</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">str_dtype_to_trt</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">dtype</span><span class="p">),</span>
|
|
<span class="n">prompt_embedding_table_size</span><span class="o">=</span><span class="n">prompt_embedding_table_size</span><span class="p">,</span>
|
|
<span class="n">position_encoding_2d</span><span class="o">=</span><span class="n">position_encoding_2d</span><span class="p">,</span>
|
|
<span class="n">mapping</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="p">,</span>
|
|
<span class="n">gather_context_logits</span><span class="o">=</span><span class="n">gather_context_logits</span><span class="p">,</span>
|
|
<span class="n">gather_generation_logits</span><span class="o">=</span><span class="n">gather_generation_logits</span><span class="p">,</span>
|
|
<span class="n">use_lora_plugin</span><span class="o">=</span><span class="n">use_lora_plugin</span><span class="p">,</span>
|
|
<span class="n">max_draft_len</span><span class="o">=</span><span class="n">max_draft_len</span><span class="p">,</span>
|
|
<span class="n">speculative_decoding_draft_tokens_external</span><span class="o">=</span>
|
|
<span class="n">speculative_decoding_draft_tokens_external</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_is_generation_length_variable</span><span class="o">=</span>
|
|
<span class="n">spec_decoding_is_generation_length_variable</span><span class="p">,</span>
|
|
<span class="n">lora_target_modules</span><span class="o">=</span><span class="n">lora_target_modules</span><span class="p">,</span>
|
|
<span class="n">multiple_profiles</span><span class="o">=</span><span class="n">multiple_profiles</span><span class="p">,</span>
|
|
<span class="n">streamingllm</span><span class="o">=</span><span class="n">streamingllm</span><span class="p">,</span>
|
|
<span class="n">opt_batch_size</span><span class="o">=</span><span class="n">opt_batch_size</span><span class="p">,</span>
|
|
<span class="n">pp_reduce_scatter</span><span class="o">=</span><span class="n">pp_reduce_scatter</span><span class="p">,</span>
|
|
<span class="n">mrope_rotary_cos_sin_size</span><span class="o">=</span><span class="n">mrope_rotary_cos_sin_size</span><span class="p">)</span>
|
|
|
|
<span class="n">result</span> <span class="o">=</span> <span class="p">{</span>
|
|
<span class="s1">'input_ids'</span><span class="p">:</span>
|
|
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">'input_ids'</span><span class="p">],</span>
|
|
<span class="s1">'position_ids'</span><span class="p">:</span>
|
|
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">'position_ids'</span><span class="p">],</span>
|
|
<span class="s1">'use_cache'</span><span class="p">:</span>
|
|
<span class="n">kv_cache_type</span> <span class="o">!=</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">DISABLED</span><span class="p">,</span>
|
|
<span class="s1">'last_token_ids'</span><span class="p">:</span>
|
|
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">'last_token_ids'</span><span class="p">],</span>
|
|
<span class="s1">'attention_mask'</span><span class="p">:</span>
|
|
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">'attention_mask'</span><span class="p">],</span>
|
|
<span class="s1">'kv_cache_params'</span><span class="p">:</span>
|
|
<span class="n">KeyValueCacheParams</span><span class="p">(</span>
|
|
<span class="n">past_key_value</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'past_key_value'</span><span class="p">],</span>
|
|
<span class="n">host_past_key_value_lengths</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span>
|
|
<span class="s1">'host_past_key_value_lengths'</span><span class="p">],</span>
|
|
<span class="n">host_max_attention_window_sizes</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span>
|
|
<span class="s1">'host_max_attention_window_sizes'</span><span class="p">],</span>
|
|
<span class="n">host_sink_token_length</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'host_sink_token_length'</span><span class="p">],</span>
|
|
<span class="n">kv_cache_block_offsets</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'kv_cache_block_offsets'</span><span class="p">],</span>
|
|
<span class="n">host_kv_cache_block_offsets</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span>
|
|
<span class="s1">'host_kv_cache_block_offsets'</span><span class="p">],</span>
|
|
<span class="n">host_kv_cache_pool_pointers</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span>
|
|
<span class="s1">'host_kv_cache_pool_pointers'</span><span class="p">],</span>
|
|
<span class="n">host_kv_cache_pool_mapping</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span>
|
|
<span class="s1">'host_kv_cache_pool_mapping'</span><span class="p">],</span>
|
|
<span class="n">cache_indirection</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'cache_indirection'</span><span class="p">],</span>
|
|
<span class="p">),</span>
|
|
<span class="s1">'attention_params'</span><span class="p">:</span>
|
|
<span class="n">AttentionParams</span><span class="p">(</span>
|
|
<span class="n">sequence_length</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'sequence_length'</span><span class="p">],</span>
|
|
<span class="n">context_lengths</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'context_lengths'</span><span class="p">],</span>
|
|
<span class="n">host_context_lengths</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'host_context_lengths'</span><span class="p">],</span>
|
|
<span class="n">max_context_length</span><span class="o">=</span><span class="n">max_input_len</span><span class="p">,</span>
|
|
<span class="n">host_request_types</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'host_request_types'</span><span class="p">],</span>
|
|
<span class="n">host_runtime_perf_knobs</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'host_runtime_perf_knobs'</span><span class="p">],</span>
|
|
<span class="n">host_context_progress</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'host_context_progress'</span><span class="p">],</span>
|
|
<span class="p">)</span>
|
|
<span class="p">}</span>
|
|
|
|
<span class="k">if</span> <span class="n">prompt_embedding_table_size</span> <span class="o">></span> <span class="mi">0</span><span class="p">:</span>
|
|
<span class="n">result</span><span class="p">[</span><span class="s1">'prompt_embedding_table'</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span>
|
|
<span class="s1">'prompt_embedding_table'</span><span class="p">]</span>
|
|
<span class="n">result</span><span class="p">[</span><span class="s1">'prompt_tasks'</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">'tasks'</span><span class="p">]</span>
|
|
<span class="n">result</span><span class="p">[</span><span class="s1">'prompt_vocab_size'</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">'prompt_vocab_size'</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">'hidden_states_input'</span><span class="p">]</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">result</span><span class="p">[</span><span class="s1">'hidden_states'</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">'hidden_states_input'</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="n">use_lora_plugin</span><span class="p">:</span>
|
|
<span class="n">result</span><span class="p">[</span><span class="s1">'lora_params'</span><span class="p">]</span> <span class="o">=</span> <span class="n">LoraParams</span><span class="p">(</span>
|
|
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">'lora_ranks'</span><span class="p">],</span>
|
|
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">'lora_weights_pointers'</span><span class="p">],</span>
|
|
<span class="n">host_context_lengths</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'host_context_lengths'</span><span class="p">],</span>
|
|
<span class="n">host_request_types</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">'host_request_types'</span><span class="p">])</span>
|
|
<span class="k">if</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">'spec_decoding_params'</span><span class="p">]</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">result</span><span class="p">[</span><span class="s1">'spec_decoding_params'</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span>
|
|
<span class="s1">'spec_decoding_params'</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">'mrope_params'</span><span class="p">]</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">result</span><span class="p">[</span><span class="s1">'mrope_params'</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">'mrope_params'</span><span class="p">]</span>
|
|
|
|
<span class="k">return</span> <span class="n">result</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="PretrainedModel.quantize">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.quantize">[docs]</a>
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">quantize</span><span class="p">(</span>
|
|
<span class="bp">cls</span><span class="p">,</span>
|
|
<span class="n">hf_model_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
|
|
<span class="n">output_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">'auto'</span><span class="p">,</span>
|
|
<span class="n">mapping</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Mapping</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">quant_config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="o">*</span><span class="p">,</span>
|
|
<span class="n">device</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">'cuda'</span><span class="p">,</span>
|
|
<span class="n">calib_dataset</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">'cnn_dailymail'</span><span class="p">,</span>
|
|
<span class="n">calib_batches</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">512</span><span class="p">,</span>
|
|
<span class="n">calib_batch_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">calib_max_seq_length</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">512</span><span class="p">,</span>
|
|
<span class="n">random_seed</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1234</span><span class="p">,</span>
|
|
<span class="n">tokenizer_max_seq_length</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">2048</span><span class="p">,</span>
|
|
<span class="o">**</span><span class="n">kwargs</span><span class="p">,</span>
|
|
<span class="p">):</span>
|
|
<span class="n">config_cls</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="s1">'config_class'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">config_cls</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">NotImplementedError</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"</span><span class="si">{</span><span class="bp">cls</span><span class="o">.</span><span class="vm">__name__</span><span class="si">}</span><span class="s2"> has not implemented corresponding config class, which is needed for correct config parsing."</span>
|
|
<span class="p">)</span>
|
|
<span class="n">config</span><span class="p">:</span> <span class="n">PretrainedConfig</span> <span class="o">=</span> <span class="n">config_cls</span><span class="o">.</span><span class="n">from_hugging_face</span><span class="p">(</span>
|
|
<span class="n">hf_model_dir</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">,</span>
|
|
<span class="n">mapping</span><span class="o">=</span><span class="n">mapping</span><span class="p">,</span>
|
|
<span class="n">quant_config</span><span class="o">=</span><span class="n">quant_config</span><span class="p">,</span>
|
|
<span class="o">**</span><span class="n">kwargs</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">moe_ep_size</span> <span class="o">></span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">NotImplementedError</span><span class="p">(</span>
|
|
<span class="s2">"Quantization for expert parallelism is not supported"</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">requires_modelopt_quantization</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"The quant_config (</span><span class="si">{</span><span class="n">quant_config</span><span class="si">}</span><span class="s2">) should not call modelopt quantization"</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="kn">from</span> <span class="nn">..quantization</span> <span class="kn">import</span> <span class="n">quantize_and_export</span>
|
|
<span class="n">quantize_and_export</span><span class="p">(</span>
|
|
<span class="n">model_dir</span><span class="o">=</span><span class="nb">str</span><span class="p">(</span><span class="n">hf_model_dir</span><span class="p">),</span>
|
|
<span class="n">device</span><span class="o">=</span><span class="n">device</span><span class="p">,</span>
|
|
<span class="n">calib_dataset</span><span class="o">=</span><span class="n">calib_dataset</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span>
|
|
<span class="n">qformat</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">get_modelopt_qformat</span><span class="p">(),</span>
|
|
<span class="n">kv_cache_dtype</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">get_modelopt_kv_cache_dtype</span><span class="p">(),</span>
|
|
<span class="n">calib_size</span><span class="o">=</span><span class="n">calib_batches</span><span class="p">,</span>
|
|
<span class="n">batch_size</span><span class="o">=</span><span class="n">calib_batch_size</span><span class="p">,</span>
|
|
<span class="n">calib_max_seq_length</span><span class="o">=</span><span class="n">calib_max_seq_length</span><span class="p">,</span>
|
|
<span class="n">awq_block_size</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">group_size</span><span class="p">,</span>
|
|
<span class="n">output_dir</span><span class="o">=</span><span class="n">output_dir</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">pp_size</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">pp_size</span><span class="p">,</span>
|
|
<span class="n">cp_size</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span><span class="p">,</span>
|
|
<span class="n">seed</span><span class="o">=</span><span class="n">random_seed</span><span class="p">,</span>
|
|
<span class="n">tokenizer_max_seq_length</span><span class="o">=</span><span class="n">tokenizer_max_seq_length</span><span class="p">,</span>
|
|
<span class="p">)</span></div>
|
|
</div>
|
|
|
|
|
|
|
|
<span class="k">class</span> <span class="nc">DecoderModelForCausalLM</span><span class="p">(</span><span class="n">PretrainedModel</span><span class="p">):</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="n">PretrainedConfig</span><span class="p">,</span> <span class="n">transformer</span><span class="p">,</span> <span class="n">lm_head</span><span class="p">):</span>
|
|
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">(</span><span class="n">config</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">transformer</span> <span class="o">=</span> <span class="n">transformer</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">lm_head</span> <span class="o">=</span> <span class="n">lm_head</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">mup_width_multiplier</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">config</span><span class="p">,</span> <span class="s1">'mup_width_multiplier'</span><span class="p">,</span>
|
|
<span class="kc">None</span><span class="p">)</span>
|
|
<span class="c1"># Create constant attention parameters to be reused by all layers.</span>
|
|
<span class="n">Attention</span><span class="o">.</span><span class="n">create_attention_const_params</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">position_embedding_type</span>
|
|
|
|
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span>
|
|
<span class="n">input_ids</span><span class="p">:</span> <span class="n">Tensor</span><span class="p">,</span>
|
|
<span class="n">position_ids</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">use_cache</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">last_token_ids</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">attention_mask</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">kv_cache_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">attention_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">mrope_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">hidden_states</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">prompt_embedding_table</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Tensor</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">prompt_tasks</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Tensor</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">prompt_vocab_size</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Tensor</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">lora_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_params</span><span class="o">=</span><span class="kc">None</span><span class="p">):</span>
|
|
|
|
<span class="c1"># fill attention params.</span>
|
|
<span class="n">attention_params</span> <span class="o">=</span> <span class="n">Attention</span><span class="o">.</span><span class="n">fill_attention_params</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="p">,</span> <span class="n">attention_params</span><span class="p">)</span>
|
|
|
|
<span class="c1"># split the sequence for context parallelism</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">></span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="nb">len</span><span class="p">(</span><span class="n">input_ids</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="o">==</span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="c1"># input shape is [-1]</span>
|
|
<span class="n">input_ids</span><span class="p">,</span> <span class="n">cp_join_index</span> <span class="o">=</span> <span class="n">cp_split_plugin</span><span class="p">(</span>
|
|
<span class="n">input_ids</span><span class="p">,</span>
|
|
<span class="n">attention_params</span><span class="o">.</span><span class="n">host_request_types</span><span class="p">,</span>
|
|
<span class="n">attention_params</span><span class="o">.</span><span class="n">host_context_lengths</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_rank</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="kc">False</span><span class="p">,</span> <span class="s2">"Context parallelism with non-remove-padding is not supported yet."</span>
|
|
|
|
<span class="n">kwargs</span> <span class="o">=</span> <span class="p">{</span>
|
|
<span class="s1">'input_ids'</span><span class="p">:</span> <span class="n">input_ids</span><span class="p">,</span>
|
|
<span class="s1">'position_ids'</span><span class="p">:</span> <span class="n">position_ids</span><span class="p">,</span>
|
|
<span class="s1">'use_cache'</span><span class="p">:</span> <span class="n">use_cache</span><span class="p">,</span>
|
|
<span class="s1">'attention_mask'</span><span class="p">:</span> <span class="n">attention_mask</span><span class="p">,</span>
|
|
<span class="s1">'kv_cache_params'</span><span class="p">:</span> <span class="n">kv_cache_params</span><span class="p">,</span>
|
|
<span class="s1">'attention_params'</span><span class="p">:</span> <span class="n">attention_params</span><span class="p">,</span>
|
|
<span class="p">}</span>
|
|
<span class="k">if</span> <span class="n">lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'lora_params'</span><span class="p">]</span> <span class="o">=</span> <span class="n">lora_params</span>
|
|
<span class="k">if</span> <span class="n">hidden_states</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'hidden_states'</span><span class="p">]</span> <span class="o">=</span> <span class="n">hidden_states</span>
|
|
<span class="k">if</span> <span class="n">prompt_embedding_table</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'prompt_embedding_table'</span><span class="p">]</span> <span class="o">=</span> <span class="n">prompt_embedding_table</span>
|
|
<span class="k">if</span> <span class="n">prompt_tasks</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'prompt_tasks'</span><span class="p">]</span> <span class="o">=</span> <span class="n">prompt_tasks</span>
|
|
<span class="k">if</span> <span class="n">prompt_vocab_size</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'prompt_vocab_size'</span><span class="p">]</span> <span class="o">=</span> <span class="n">prompt_vocab_size</span>
|
|
|
|
<span class="k">if</span> <span class="n">spec_decoding_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'spec_decoding_params'</span><span class="p">]</span> <span class="o">=</span> <span class="n">spec_decoding_params</span>
|
|
<span class="k">if</span> <span class="n">mrope_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kwargs</span><span class="p">[</span><span class="s1">'mrope_params'</span><span class="p">]</span> <span class="o">=</span> <span class="n">mrope_params</span>
|
|
|
|
<span class="n">hidden_states</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">transformer</span><span class="o">.</span><span class="n">forward</span><span class="p">(</span><span class="o">**</span><span class="n">kwargs</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">use_cache</span><span class="p">:</span>
|
|
<span class="n">hidden_states</span><span class="p">,</span> <span class="n">presents</span> <span class="o">=</span> <span class="n">hidden_states</span>
|
|
|
|
<span class="c1"># All gather and rebuild sequence after transformer layer for context parallelism</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">></span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="nb">len</span><span class="p">(</span><span class="n">hidden_states</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="o">==</span> <span class="mi">2</span><span class="p">:</span>
|
|
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">allgather</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_group</span><span class="p">,</span>
|
|
<span class="n">gather_dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
|
|
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">view</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span>
|
|
<span class="p">[</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="n">hidden_states</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="o">-</span><span class="mi">1</span><span class="p">]])</span>
|
|
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">index_select</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="n">cp_join_index</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="kc">False</span><span class="p">,</span> <span class="s2">"Context parallelism with non-remove-padding is not supported yet."</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">is_last_pp_rank</span><span class="p">():</span>
|
|
<span class="n">all_hidden_states</span> <span class="o">=</span> <span class="n">hidden_states</span>
|
|
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">gather_last_token_logits</span><span class="p">(</span>
|
|
<span class="n">hidden_states</span><span class="p">,</span> <span class="n">last_token_ids</span><span class="p">,</span>
|
|
<span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">remove_input_padding</span><span class="p">)</span>
|
|
|
|
<span class="c1"># [batch_size, hidden_size] -> [batch_size, vocab_size]</span>
|
|
<span class="n">lm_logits</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">lm_head</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="nb">hasattr</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="p">,</span> <span class="s1">'output_multiplier_scale'</span><span class="p">):</span>
|
|
<span class="n">lm_logits</span> <span class="o">*=</span> <span class="nb">getattr</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="p">,</span> <span class="s1">'output_multiplier_scale'</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">mup_width_multiplier</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">lm_logits</span> <span class="o">=</span> <span class="n">lm_logits</span> <span class="o">/</span> <span class="bp">self</span><span class="o">.</span><span class="n">mup_width_multiplier</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">has_config_group</span><span class="p">(</span><span class="n">Gemma2ConfigGroup</span><span class="p">):</span>
|
|
<span class="n">softcap</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">get_config_group</span><span class="p">(</span>
|
|
<span class="n">Gemma2ConfigGroup</span><span class="p">)</span><span class="o">.</span><span class="n">final_logit_softcapping</span>
|
|
<span class="k">if</span> <span class="n">softcap</span><span class="p">:</span>
|
|
<span class="n">lm_logits</span> <span class="o">=</span> <span class="n">lm_logits</span> <span class="o">*</span> <span class="nb">float</span><span class="p">(</span><span class="mi">1</span> <span class="o">/</span> <span class="n">softcap</span><span class="p">)</span>
|
|
<span class="n">lm_logits</span> <span class="o">=</span> <span class="n">tanh</span><span class="p">(</span><span class="n">lm_logits</span><span class="p">)</span> <span class="o">*</span> <span class="nb">float</span><span class="p">(</span><span class="n">softcap</span><span class="p">)</span>
|
|
<span class="n">lm_logits</span><span class="o">.</span><span class="n">mark_output</span><span class="p">(</span><span class="s1">'logits'</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">logits_dtype</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">hidden_states</span><span class="o">.</span><span class="n">mark_output</span><span class="p">(</span><span class="s1">'hidden_states_output'</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">use_cache</span> <span class="ow">and</span> <span class="ow">not</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">paged_kv_cache</span><span class="p">:</span>
|
|
<span class="k">for</span> <span class="n">i</span><span class="p">,</span> <span class="n">present</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">pp_layers</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">num_hidden_layers</span><span class="p">),</span> <span class="n">presents</span><span class="p">):</span>
|
|
<span class="n">present</span><span class="o">.</span><span class="n">mark_output</span><span class="p">(</span><span class="sa">f</span><span class="s1">'present_key_value_</span><span class="si">{</span><span class="n">i</span><span class="si">}</span><span class="s1">'</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">kv_dtype</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">is_last_pp_rank</span><span class="p">():</span>
|
|
<span class="k">return</span> <span class="p">(</span><span class="n">lm_logits</span><span class="p">,</span> <span class="n">presents</span><span class="p">,</span> <span class="n">hidden_states</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span> <span class="n">presents</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">is_last_pp_rank</span><span class="p">():</span>
|
|
<span class="k">return</span> <span class="n">lm_logits</span><span class="p">,</span> <span class="n">hidden_states</span><span class="p">,</span> <span class="n">all_hidden_states</span>
|
|
<span class="k">return</span> <span class="n">hidden_states</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">fuse_gate_mlp</span><span class="p">(</span>
|
|
<span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">,</span>
|
|
<span class="n">gemm_swiglu_plugin_dtype</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">str</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">low_latency_gemm_swiglu_plugin_dtype</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">str</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="p">)</span> <span class="o">-></span> <span class="n">PretrainedModel</span><span class="p">:</span>
|
|
<span class="kn">from</span> <span class="nn">..quantization.quantize</span> <span class="kn">import</span> <span class="n">fp8_quantize</span>
|
|
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">mlp</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">mlp</span><span class="p">,</span> <span class="n">GatedMLP</span><span class="p">):</span>
|
|
<span class="n">init_params</span> <span class="o">=</span> <span class="n">get_init_params</span><span class="p">(</span><span class="n">mlp</span><span class="p">)</span>
|
|
|
|
<span class="n">hidden_act</span> <span class="o">=</span> <span class="n">init_params</span><span class="p">[</span><span class="s2">"hidden_act"</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="n">hidden_act</span> <span class="ow">not</span> <span class="ow">in</span> <span class="p">[</span><span class="s2">"silu"</span><span class="p">,</span> <span class="s2">"gelu"</span><span class="p">]:</span>
|
|
<span class="n">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"fuse_gate_mlp cannot be done for </span><span class="si">{</span><span class="n">name</span><span class="si">}</span><span class="s2"> due to unsupported activation </span><span class="si">{</span><span class="n">hidden_act</span><span class="si">}</span><span class="s2">. Skipping."</span>
|
|
<span class="p">)</span>
|
|
<span class="k">continue</span>
|
|
|
|
<span class="n">init_params</span><span class="p">[</span><span class="s2">"inner_layernorm"</span><span class="p">]</span> <span class="o">=</span> <span class="n">mlp</span><span class="o">.</span><span class="n">inner_layernorm</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="n">fused_layer</span> <span class="o">=</span> <span class="n">FusedGatedMLP</span><span class="p">(</span><span class="o">**</span><span class="n">init_params</span><span class="p">)</span>
|
|
|
|
<span class="n">fc_name</span> <span class="o">=</span> <span class="n">name</span> <span class="o">+</span> <span class="s1">'.fc'</span>
|
|
<span class="n">layer_quant_cfg</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">get_quant_cfg</span><span class="p">(</span><span class="n">fc_name</span><span class="p">)</span>
|
|
<span class="n">layer_quant_algo</span> <span class="o">=</span> <span class="n">layer_quant_cfg</span><span class="o">.</span><span class="n">quant_algo</span>
|
|
<span class="k">if</span> <span class="n">layer_quant_algo</span> <span class="o">!=</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span> <span class="ow">and</span> <span class="n">layer_quant_algo</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">continue</span>
|
|
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">exclude_modules</span><span class="p">,</span> <span class="nb">list</span><span class="p">)</span> \
|
|
<span class="ow">and</span> <span class="n">fc_name</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">exclude_modules</span><span class="p">:</span>
|
|
<span class="n">layer_quant_algo</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="k">if</span> <span class="n">layer_quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span>
|
|
<span class="n">fused_layer</span> <span class="o">=</span> <span class="n">fp8_quantize</span><span class="p">(</span><span class="n">fused_layer</span><span class="p">,</span> <span class="n">layer_quant_cfg</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">mlp</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span> <span class="nb">str</span><span class="p">):</span>
|
|
<span class="n">dtype</span> <span class="o">=</span> <span class="n">str_dtype_to_torch</span><span class="p">(</span><span class="n">mlp</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">dtype</span> <span class="o">=</span> <span class="n">trt_dtype_to_torch</span><span class="p">(</span><span class="n">mlp</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
|
|
<span class="n">gate_weight</span> <span class="o">=</span> <span class="n">numpy_to_torch</span><span class="p">(</span><span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
|
|
<span class="n">fc_weight</span> <span class="o">=</span> <span class="n">numpy_to_torch</span><span class="p">(</span><span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
|
|
<span class="k">assert</span> <span class="n">gate_weight</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">fc_weight</span><span class="o">.</span><span class="n">dtype</span>
|
|
<span class="n">need_qdq</span> <span class="o">=</span> <span class="n">gate_weight</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">torch</span><span class="o">.</span><span class="n">float8_e4m3fn</span>
|
|
|
|
<span class="n">gate_weight</span> <span class="o">=</span> <span class="n">gate_weight</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dtype</span><span class="p">)</span>
|
|
<span class="n">fc_weight</span> <span class="o">=</span> <span class="n">fc_weight</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dtype</span><span class="p">)</span>
|
|
<span class="c1"># dequantize if needed</span>
|
|
<span class="k">if</span> <span class="n">need_qdq</span><span class="p">:</span>
|
|
<span class="n">gate_weight</span> <span class="o">=</span> <span class="n">gate_weight</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dtype</span><span class="p">)</span> <span class="o">*</span> <span class="n">numpy_to_torch</span><span class="p">(</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
|
|
<span class="n">fc_weight</span> <span class="o">=</span> <span class="n">fc_weight</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dtype</span><span class="p">)</span> <span class="o">*</span> <span class="n">numpy_to_torch</span><span class="p">(</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
|
|
|
|
<span class="c1"># concat</span>
|
|
<span class="n">fused_weight</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">([</span><span class="n">gate_weight</span><span class="p">,</span> <span class="n">fc_weight</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
|
|
|
|
<span class="n">fused_weight_scaling_factor</span> <span class="o">=</span> <span class="n">numpy_to_torch</span><span class="p">(</span>
|
|
<span class="nb">max</span><span class="p">(</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="p">))</span>
|
|
<span class="c1"># quantize if needed</span>
|
|
<span class="k">if</span> <span class="n">need_qdq</span><span class="p">:</span>
|
|
<span class="n">fused_weight</span> <span class="o">=</span> <span class="p">(</span><span class="n">fused_weight</span> <span class="o">/</span>
|
|
<span class="n">fused_weight_scaling_factor</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span>
|
|
<span class="n">torch</span><span class="o">.</span><span class="n">float8_e4m3fn</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">gemm_swiglu_plugin_dtype</span> <span class="o">==</span> <span class="s1">'fp8'</span> <span class="ow">or</span> <span class="n">low_latency_gemm_swiglu_plugin_dtype</span> <span class="o">==</span> <span class="s1">'fp8'</span><span class="p">:</span>
|
|
<span class="c1"># gemm_swiglu_plugin needs (k, n) weights</span>
|
|
<span class="c1"># but weights should still be k-major for fp8</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weight</span> <span class="o">=</span> <span class="n">Parameter</span><span class="p">(</span>
|
|
<span class="n">shape</span><span class="o">=</span><span class="p">(</span><span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">in_features</span><span class="p">,</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">out_features</span><span class="p">),</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="s1">'fp8'</span><span class="p">)</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">fused_weight</span><span class="o">.</span><span class="n">view</span><span class="p">(</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">in_features</span><span class="p">,</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">out_features</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">fused_weight</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">fused_weight_scaling_factor</span>
|
|
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="nb">max</span><span class="p">(</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="k">elif</span> <span class="n">layer_quant_algo</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
|
|
<span class="p">[</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">mlp</span><span class="o">.</span><span class="n">bias</span><span class="p">:</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
|
|
<span class="p">[</span><span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span> <span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span><span class="sa">f</span><span class="s1">'Unsupported quant algo: </span><span class="si">{</span><span class="n">layer_quant_algo</span><span class="si">}</span><span class="s1">'</span><span class="p">)</span>
|
|
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">proj</span> <span class="o">=</span> <span class="n">mlp</span><span class="o">.</span><span class="n">proj</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">inner_layernorm</span> <span class="o">=</span> <span class="n">mlp</span><span class="o">.</span><span class="n">inner_layernorm</span>
|
|
|
|
<span class="n">_</span><span class="p">,</span> <span class="n">mlp_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">'.'</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
|
|
<span class="nb">setattr</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">mlp_name</span><span class="p">,</span> <span class="n">fused_layer</span><span class="p">)</span>
|
|
|
|
<span class="k">elif</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">mlp</span><span class="p">,</span> <span class="n">Fp8RowwiseGatedMLP</span><span class="p">):</span>
|
|
<span class="n">init_params</span> <span class="o">=</span> <span class="n">get_init_params</span><span class="p">(</span><span class="n">mlp</span><span class="p">)</span>
|
|
|
|
<span class="n">hidden_act</span> <span class="o">=</span> <span class="n">init_params</span><span class="p">[</span><span class="s2">"hidden_act"</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="n">hidden_act</span> <span class="ow">not</span> <span class="ow">in</span> <span class="p">[</span><span class="s2">"silu"</span><span class="p">,</span> <span class="s2">"gelu"</span><span class="p">]:</span>
|
|
<span class="n">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"fuse_gate_mlp cannot be done for </span><span class="si">{</span><span class="n">name</span><span class="si">}</span><span class="s2"> due to unsupported activation </span><span class="si">{</span><span class="n">hidden_act</span><span class="si">}</span><span class="s2">. Skipping."</span>
|
|
<span class="p">)</span>
|
|
<span class="k">continue</span>
|
|
|
|
<span class="k">if</span> <span class="n">mlp</span><span class="o">.</span><span class="n">clamp_val</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">init_params</span><span class="p">[</span><span class="s2">"clamp_val"</span><span class="p">]</span> <span class="o">=</span> <span class="n">mlp</span><span class="o">.</span><span class="n">clamp_val</span><span class="o">.</span><span class="n">raw_value</span><span class="o">.</span><span class="n">tolist</span><span class="p">()</span>
|
|
<span class="n">fused_layer</span> <span class="o">=</span> <span class="n">Fp8RowwiseFusedGatedMLP</span><span class="p">(</span><span class="o">**</span><span class="n">init_params</span><span class="p">)</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
|
|
<span class="p">[</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">per_channel_scale</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
|
|
<span class="p">[</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">per_channel_scale</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">per_channel_scale</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">mlp</span><span class="o">.</span><span class="n">bias</span><span class="p">:</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
|
|
<span class="p">[</span><span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span> <span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">],</span> <span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
|
|
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">proj</span> <span class="o">=</span> <span class="n">mlp</span><span class="o">.</span><span class="n">proj</span>
|
|
<span class="n">_</span><span class="p">,</span> <span class="n">mlp_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">'.'</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
|
|
<span class="nb">setattr</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">mlp_name</span><span class="p">,</span> <span class="n">fused_layer</span><span class="p">)</span>
|
|
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">unfuse_qkv_gemm</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-></span> <span class="n">PretrainedModel</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">'''Split all the models' Attention layer's QKV GEMM into 3 GEMMs layer.q layer.k, layer.v and return the changed model</span>
|
|
<span class="sd"> '''</span>
|
|
<span class="kn">from</span> <span class="nn">..quantization.quantize</span> <span class="kn">import</span> <span class="n">quantize</span>
|
|
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">Attention</span><span class="p">)</span> <span class="ow">and</span> <span class="ow">not</span> <span class="n">layer</span><span class="o">.</span><span class="n">cross_attention</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">==</span> <span class="mi">1</span><span class="p">,</span> <span class="s2">"please disable manual tp when enable auto parallel"</span>
|
|
<span class="k">if</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">continue</span>
|
|
<span class="n">qkv_params</span> <span class="o">=</span> <span class="n">get_init_params</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="p">,</span> <span class="n">ColumnLinear</span><span class="p">)</span>
|
|
<span class="n">qkv_params</span><span class="p">[</span><span class="s2">"bias"</span><span class="p">]</span> <span class="o">=</span> <span class="n">qkv_params</span><span class="p">[</span><span class="s2">"bias"</span><span class="p">]</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="n">qkv_params</span><span class="p">[</span><span class="s2">"strict_dtype"</span><span class="p">]</span> <span class="o">=</span> <span class="n">qkv_params</span><span class="o">.</span><span class="n">get</span><span class="p">(</span>
|
|
<span class="s2">"strict_dtype"</span><span class="p">)</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="n">q</span> <span class="o">=</span> <span class="n">ColumnLinear</span><span class="p">(</span>
|
|
<span class="o">**</span><span class="p">{</span>
|
|
<span class="o">**</span><span class="n">qkv_params</span><span class="p">,</span>
|
|
<span class="s2">"out_features"</span><span class="p">:</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="p">})</span>
|
|
<span class="n">k</span> <span class="o">=</span> <span class="n">ColumnLinear</span><span class="p">(</span>
|
|
<span class="o">**</span><span class="p">{</span>
|
|
<span class="o">**</span><span class="n">qkv_params</span><span class="p">,</span>
|
|
<span class="s2">"out_features"</span><span class="p">:</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="p">})</span>
|
|
<span class="n">v</span> <span class="o">=</span> <span class="n">ColumnLinear</span><span class="p">(</span>
|
|
<span class="o">**</span><span class="p">{</span>
|
|
<span class="o">**</span><span class="n">qkv_params</span><span class="p">,</span>
|
|
<span class="s2">"out_features"</span><span class="p">:</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="p">})</span>
|
|
<span class="n">layer_quant_cfg</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">get_quant_cfg</span><span class="p">(</span><span class="n">name</span> <span class="o">+</span> <span class="s1">'.qkv'</span><span class="p">)</span>
|
|
<span class="n">q</span> <span class="o">=</span> <span class="n">quantize</span><span class="p">(</span><span class="n">q</span><span class="p">,</span> <span class="n">layer_quant_cfg</span><span class="p">)</span>
|
|
<span class="n">k</span> <span class="o">=</span> <span class="n">quantize</span><span class="p">(</span><span class="n">k</span><span class="p">,</span> <span class="n">layer_quant_cfg</span><span class="p">)</span>
|
|
<span class="n">v</span> <span class="o">=</span> <span class="n">quantize</span><span class="p">(</span><span class="n">v</span><span class="p">,</span> <span class="n">layer_quant_cfg</span><span class="p">)</span>
|
|
<span class="n">out_features</span> <span class="o">=</span> <span class="n">q</span><span class="o">.</span><span class="n">out_features</span> <span class="o">+</span> <span class="n">k</span><span class="o">.</span><span class="n">out_features</span> <span class="o">+</span> <span class="n">v</span><span class="o">.</span><span class="n">out_features</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="p">,</span> <span class="p">(</span>
|
|
<span class="n">WeightOnlyQuantLinear</span><span class="p">,</span>
|
|
<span class="n">WeightOnlyQuantRowLinear</span><span class="p">,</span>
|
|
<span class="n">WeightOnlyGroupwiseQuantLinear</span><span class="p">,</span>
|
|
<span class="n">WeightOnlyGroupwiseQuantRowLinear</span><span class="p">,</span>
|
|
<span class="p">)):</span>
|
|
<span class="n">out_dim</span> <span class="o">=</span> <span class="mi">1</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">out_dim</span> <span class="o">=</span> <span class="mi">0</span>
|
|
<span class="k">if</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">is_inited</span><span class="p">():</span>
|
|
<span class="n">qkv_weight</span> <span class="o">=</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span>
|
|
<span class="n">weights</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">split</span><span class="p">(</span><span class="n">qkv_weight</span><span class="p">,</span> <span class="p">[</span>
|
|
<span class="n">qkv_weight</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="n">out_dim</span><span class="p">]</span> <span class="o">*</span> <span class="n">q</span><span class="o">.</span><span class="n">out_features</span> <span class="o">//</span> <span class="n">out_features</span><span class="p">,</span>
|
|
<span class="n">qkv_weight</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="n">out_dim</span><span class="p">]</span> <span class="o">*</span>
|
|
<span class="p">(</span><span class="n">q</span><span class="o">.</span><span class="n">out_features</span> <span class="o">+</span> <span class="n">k</span><span class="o">.</span><span class="n">out_features</span><span class="p">)</span> <span class="o">//</span> <span class="n">out_features</span><span class="p">,</span>
|
|
<span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=</span><span class="n">out_dim</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">gemm</span><span class="p">,</span> <span class="n">weight</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">([</span><span class="n">q</span><span class="p">,</span> <span class="n">k</span><span class="p">,</span> <span class="n">v</span><span class="p">],</span> <span class="n">weights</span><span class="p">):</span>
|
|
<span class="n">gemm</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">weight</span>
|
|
<span class="k">if</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">bias</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">is_inited</span><span class="p">():</span>
|
|
<span class="n">qkv_bias</span> <span class="o">=</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span>
|
|
<span class="n">biases</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">split</span><span class="p">(</span><span class="n">qkv_bias</span><span class="p">,</span> <span class="p">[</span>
|
|
<span class="n">qkv_bias</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="n">out_dim</span><span class="p">]</span> <span class="o">*</span> <span class="n">q</span><span class="o">.</span><span class="n">out_features</span> <span class="o">//</span> <span class="n">out_features</span><span class="p">,</span>
|
|
<span class="n">qkv_bias</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="n">out_dim</span><span class="p">]</span> <span class="o">*</span>
|
|
<span class="p">(</span><span class="n">q</span><span class="o">.</span><span class="n">out_features</span> <span class="o">+</span> <span class="n">k</span><span class="o">.</span><span class="n">out_features</span><span class="p">)</span> <span class="o">//</span> <span class="n">out_features</span><span class="p">,</span>
|
|
<span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=</span><span class="n">out_dim</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">gemm</span><span class="p">,</span> <span class="n">bias</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">([</span><span class="n">q</span><span class="p">,</span> <span class="n">k</span><span class="p">,</span> <span class="n">v</span><span class="p">],</span> <span class="n">biases</span><span class="p">):</span>
|
|
<span class="n">gemm</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">bias</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">parameter</span> <span class="ow">in</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">_parameters</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="p">[</span><span class="s2">"weight"</span><span class="p">,</span> <span class="s2">"bias"</span><span class="p">]:</span>
|
|
<span class="k">for</span> <span class="n">gemm</span> <span class="ow">in</span> <span class="p">[</span><span class="n">q</span><span class="p">,</span> <span class="n">k</span><span class="p">,</span> <span class="n">v</span><span class="p">]:</span>
|
|
<span class="nb">setattr</span><span class="p">(</span><span class="n">gemm</span><span class="p">,</span> <span class="n">name</span><span class="p">,</span> <span class="n">parameter</span><span class="p">)</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">q</span> <span class="o">=</span> <span class="n">q</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">k</span> <span class="o">=</span> <span class="n">k</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">v</span> <span class="o">=</span> <span class="n">v</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">qkv</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">fuse_rg_lru</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-></span> <span class="n">PretrainedModel</span><span class="p">:</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">rg_lru</span><span class="p">,</span> <span class="n">parent</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">rg_lru</span><span class="p">,</span> <span class="n">RgLru</span><span class="p">):</span>
|
|
<span class="n">fused_layer</span> <span class="o">=</span> <span class="n">FusedRgLru</span><span class="p">(</span><span class="o">**</span><span class="n">get_init_params</span><span class="p">(</span><span class="n">rg_lru</span><span class="p">))</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
|
|
<span class="p">[</span>
|
|
<span class="n">rg_lru</span><span class="o">.</span><span class="n">input_gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="n">rg_lru</span><span class="o">.</span><span class="n">recurrent_gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=-</span><span class="mi">1</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
|
|
<span class="p">[</span>
|
|
<span class="n">rg_lru</span><span class="o">.</span><span class="n">input_gate</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="n">rg_lru</span><span class="o">.</span><span class="n">recurrent_gate</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
|
|
<span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=-</span><span class="mi">1</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="n">fused_layer</span><span class="o">.</span><span class="n">recurrent_param</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">rg_lru</span><span class="o">.</span><span class="n">recurrent_param</span><span class="o">.</span><span class="n">raw_value</span>
|
|
<span class="n">rg_lru_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">'.'</span><span class="p">,</span> <span class="mi">1</span><span class="p">)[</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span>
|
|
<span class="nb">setattr</span><span class="p">(</span><span class="n">parent</span><span class="p">,</span> <span class="n">rg_lru_name</span><span class="p">,</span> <span class="n">fused_layer</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">set_prompt_tuning</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-></span> <span class="n">PretrainedModel</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">'''Replace the given models embedding layer with a PromptTuningEmbedding layer in-place, return the changed model</span>
|
|
<span class="sd"> Pre-conditions: vocab_embedding exists</span>
|
|
<span class="sd"> Post-conditions: isinstance(vocab_embedding, PromptTuningEmbedding)</span>
|
|
|
|
<span class="sd"> '''</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">embedding</span><span class="p">,</span> <span class="n">parent</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
|
|
<span class="n">layer_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">'.'</span><span class="p">,</span> <span class="mi">1</span><span class="p">)[</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="n">layer_name</span> <span class="o">==</span> <span class="s2">"vocab_embedding"</span> <span class="ow">and</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">embedding</span><span class="p">,</span> <span class="n">Embedding</span><span class="p">):</span>
|
|
<span class="n">ptuning_embedding</span> <span class="o">=</span> <span class="n">PromptTuningEmbedding</span><span class="p">(</span>
|
|
<span class="o">**</span><span class="n">get_init_params</span><span class="p">(</span><span class="n">embedding</span><span class="p">))</span>
|
|
<span class="n">ptuning_embedding</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">embedding</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span>
|
|
<span class="n">parent</span><span class="o">.</span><span class="n">vocab_embedding</span> <span class="o">=</span> <span class="n">ptuning_embedding</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">add_lora</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">,</span>
|
|
<span class="n">max_lora_rank</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">])</span> <span class="o">-></span> <span class="n">PretrainedModel</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">''' Add lora layers to the Attention/BertAttention/Linear/RowLinear/FusedGatedMLP layers to the given model, return the changed model</span>
|
|
<span class="sd"> '''</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules</span><span class="p">():</span>
|
|
<span class="n">max_rank</span> <span class="o">=</span> <span class="n">max_lora_rank</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="p">(</span><span class="n">Attention</span><span class="p">,</span> <span class="n">BertAttention</span><span class="p">)):</span>
|
|
<span class="k">if</span> <span class="n">max_rank</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">max_rank</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">)</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">qkv_lora</span> <span class="o">=</span> <span class="n">Lora</span><span class="p">(</span>
|
|
<span class="n">in_hidden_size</span><span class="o">=</span><span class="n">layer</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">out_hidden_sizes</span><span class="o">=</span><span class="p">[</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span>
|
|
<span class="p">],</span>
|
|
<span class="n">max_low_rank</span><span class="o">=</span><span class="n">max_rank</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="p">(</span><span class="n">Linear</span><span class="p">,</span> <span class="n">RowLinear</span><span class="p">)):</span>
|
|
<span class="k">if</span> <span class="n">max_rank</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">max_rank</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">in_features</span><span class="p">,</span> <span class="n">layer</span><span class="o">.</span><span class="n">out_features</span><span class="p">)</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">lora</span> <span class="o">=</span> <span class="n">Lora</span><span class="p">(</span>
|
|
<span class="n">in_hidden_size</span><span class="o">=</span><span class="n">layer</span><span class="o">.</span><span class="n">in_features</span><span class="p">,</span>
|
|
<span class="n">out_hidden_sizes</span><span class="o">=</span><span class="p">[</span><span class="n">layer</span><span class="o">.</span><span class="n">out_features</span><span class="p">],</span>
|
|
<span class="n">max_low_rank</span><span class="o">=</span><span class="n">max_rank</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="p">(</span><span class="n">MLP</span><span class="p">,</span> <span class="n">FusedGatedMLP</span><span class="p">)):</span>
|
|
<span class="k">if</span> <span class="n">max_rank</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">max_rank</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span><span class="p">)</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">lora</span> <span class="o">=</span> <span class="n">Lora</span><span class="p">(</span>
|
|
<span class="n">in_hidden_size</span><span class="o">=</span><span class="n">layer</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">out_hidden_sizes</span><span class="o">=</span><span class="p">[</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span>
|
|
<span class="p">],</span>
|
|
<span class="n">max_low_rank</span><span class="o">=</span><span class="n">max_rank</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">MOE</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">max_rank</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">max_rank</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span><span class="p">)</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">max_low_rank</span> <span class="o">=</span> <span class="n">max_rank</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">to_ootb_moe</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-></span> <span class="n">PretrainedModel</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">''' Use OOTB MoE instead of MoE plugin, return the changed model</span>
|
|
<span class="sd"> '''</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer</span><span class="p">,</span> <span class="n">parent</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">MOE</span><span class="p">):</span>
|
|
<span class="n">layer_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">'.'</span><span class="p">,</span> <span class="mi">1</span><span class="p">)[</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span>
|
|
<span class="n">ootb_layer</span> <span class="o">=</span> <span class="n">layer</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">MoeOOTB</span><span class="p">,</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="p">)</span>
|
|
<span class="nb">setattr</span><span class="p">(</span><span class="n">parent</span><span class="p">,</span> <span class="n">layer_name</span><span class="p">,</span> <span class="n">ootb_layer</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">parallelize_embedding</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-></span> <span class="n">PretrainedModel</span><span class="p">:</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">embedding</span><span class="p">,</span> <span class="n">parent</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
|
|
<span class="n">layer_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">'.'</span><span class="p">,</span> <span class="mi">1</span><span class="p">)[</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">embedding</span><span class="p">,</span> <span class="n">Embedding</span><span class="p">)</span> <span class="ow">and</span> <span class="n">embedding</span><span class="o">.</span><span class="n">tp_group</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">init_params</span> <span class="o">=</span> <span class="n">get_init_params</span><span class="p">(</span><span class="n">embedding</span><span class="p">)</span>
|
|
<span class="n">init_params</span><span class="p">[</span><span class="s2">"tp_group"</span><span class="p">]</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_group</span>
|
|
<span class="n">init_params</span><span class="p">[</span><span class="s2">"tp_size"</span><span class="p">]</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_size</span>
|
|
<span class="n">init_params</span><span class="p">[</span><span class="s2">"tp_rank"</span><span class="p">]</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_rank</span>
|
|
<span class="n">init_params</span><span class="p">[</span><span class="s2">"sharding_dim"</span><span class="p">]</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">embedding_sharding_dim</span>
|
|
<span class="n">new_embedding</span> <span class="o">=</span> <span class="n">embedding</span><span class="o">.</span><span class="vm">__class__</span><span class="p">(</span><span class="o">**</span><span class="n">init_params</span><span class="p">)</span>
|
|
<span class="nb">setattr</span><span class="p">(</span><span class="n">parent</span><span class="p">,</span> <span class="n">layer_name</span><span class="p">,</span> <span class="n">new_embedding</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">share_embedding</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-></span> <span class="n">PretrainedModel</span><span class="p">:</span>
|
|
<span class="n">lm_head</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="n">vocab_embedding</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules</span><span class="p">():</span>
|
|
<span class="n">layer_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">'.'</span><span class="p">,</span> <span class="mi">1</span><span class="p">)[</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="n">layer_name</span> <span class="o">==</span> <span class="s2">"lm_head"</span><span class="p">:</span>
|
|
<span class="n">lm_head</span> <span class="o">=</span> <span class="n">layer</span>
|
|
<span class="k">if</span> <span class="n">layer_name</span> <span class="o">==</span> <span class="s2">"vocab_embedding"</span><span class="p">:</span>
|
|
<span class="n">vocab_embedding</span> <span class="o">=</span> <span class="n">layer</span>
|
|
<span class="k">if</span> <span class="n">lm_head</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">vocab_embedding</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">break</span>
|
|
|
|
<span class="c1"># Cannot find either lm_head or vocab_embedding, e.g., pipeline parallel</span>
|
|
<span class="k">if</span> <span class="n">lm_head</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">or</span> <span class="n">vocab_embedding</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
<span class="c1"># lm_head and vocab_embedding have different shapes, e.g., tensor parallel without embedding parallel</span>
|
|
<span class="k">if</span> <span class="n">lm_head</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">shape</span> <span class="o">!=</span> <span class="n">vocab_embedding</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">shape</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
<span class="c1"># Don't assume weight can be shared if vocab_embedding is not initialized, e.g., dummy weights</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">vocab_embedding</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">is_inited</span><span class="p">():</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
<span class="k">if</span> <span class="n">lm_head</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">is_inited</span><span class="p">():</span>
|
|
<span class="n">lm_head_weight</span> <span class="o">=</span> <span class="n">numpy_to_torch</span><span class="p">(</span><span class="n">lm_head</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
|
|
<span class="n">vocab_embed_weight</span> <span class="o">=</span> <span class="n">numpy_to_torch</span><span class="p">(</span><span class="n">vocab_embedding</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
|
|
<span class="c1"># The lm_head and vocab_embedding have different weights</span>
|
|
<span class="k">if</span> <span class="p">(</span><span class="n">lm_head_weight</span> <span class="o">-</span> <span class="n">vocab_embed_weight</span><span class="p">)</span><span class="o">.</span><span class="n">abs</span><span class="p">()</span><span class="o">.</span><span class="n">max</span><span class="p">()</span><span class="o">.</span><span class="n">item</span><span class="p">()</span> <span class="o">></span> <span class="mf">1e-6</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
<span class="n">lm_head</span><span class="o">.</span><span class="n">weight</span> <span class="o">=</span> <span class="n">vocab_embedding</span><span class="o">.</span><span class="n">weight</span>
|
|
<span class="k">if</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">lm_head</span><span class="p">,</span> <span class="s1">'per_channel_scale'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">getattr</span><span class="p">(</span>
|
|
<span class="n">vocab_embedding</span><span class="p">,</span> <span class="s1">'per_channel_scale'</span><span class="p">,</span> <span class="kc">None</span><span class="p">):</span>
|
|
<span class="n">lm_head</span><span class="o">.</span><span class="n">per_channel_scale</span> <span class="o">=</span> <span class="n">vocab_embedding</span><span class="o">.</span><span class="n">per_token_scale</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">set_fp8_context_fhma</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-></span> <span class="n">PretrainedModel</span><span class="p">:</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">Attention</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">hasattr</span><span class="p">(</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">dense</span><span class="p">,</span> <span class="s1">'activation_scaling_factor'</span><span class="p">):</span>
|
|
<span class="n">scale</span> <span class="o">=</span> <span class="p">[</span><span class="mf">1.0</span><span class="p">]</span> <span class="o">/</span> <span class="n">layer</span><span class="o">.</span><span class="n">dense</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">attention_output_orig_quant_scale</span> <span class="o">=</span> <span class="n">Parameter</span><span class="p">(</span>
|
|
<span class="n">value</span><span class="o">=</span><span class="n">scale</span><span class="o">.</span><span class="n">astype</span><span class="p">(</span><span class="n">np</span><span class="o">.</span><span class="n">float32</span><span class="p">),</span> <span class="n">dtype</span><span class="o">=</span><span class="s1">'float32'</span><span class="p">)</span>
|
|
<span class="k">elif</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">Attention</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">hasattr</span><span class="p">(</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">dense</span><span class="p">,</span> <span class="s1">'activation_scaling_factor_2'</span><span class="p">):</span>
|
|
<span class="n">scale</span> <span class="o">=</span> <span class="n">layer</span><span class="o">.</span><span class="n">dense</span><span class="o">.</span><span class="n">activation_scaling_factor_2</span><span class="o">.</span><span class="n">raw_value</span>
|
|
<span class="n">layer</span><span class="o">.</span><span class="n">attention_output_orig_quant_scale</span> <span class="o">=</span> <span class="n">Parameter</span><span class="p">(</span>
|
|
<span class="n">value</span><span class="o">=</span><span class="n">scale</span><span class="o">.</span><span class="n">astype</span><span class="p">(</span><span class="n">np</span><span class="o">.</span><span class="n">float32</span><span class="p">),</span> <span class="n">dtype</span><span class="o">=</span><span class="s1">'float32'</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">optimize_model</span><span class="p">(</span>
|
|
<span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">,</span>
|
|
<span class="n">use_parallel_embedding</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">share_embedding_table</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">use_ootb_moe</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">use_fused_mlp</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">gemm_swiglu_plugin_dtype</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">str</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">low_latency_gemm_swiglu_plugin_dtype</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">str</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">use_fused_rg_lru</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">use_unfused_qkv_gemm</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">use_prompt_tuning</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">use_lora</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">max_lora_rank</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">use_fp8_context_fmha</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">use_optimize_cross_qkv</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="p">)</span> <span class="o">-></span> <span class="n">PretrainedModel</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">"""</span>
|
|
<span class="sd"> Run optimization passes on model.</span>
|
|
<span class="sd"> There are dependencies between some passes,</span>
|
|
<span class="sd"> so we always run passes in the order of arguments to guarantee the execution order.</span>
|
|
<span class="sd"> """</span>
|
|
<span class="c1"># before weight loading</span>
|
|
<span class="k">if</span> <span class="n">use_parallel_embedding</span><span class="p">:</span>
|
|
<span class="n">model</span> <span class="o">=</span> <span class="n">parallelize_embedding</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">share_embedding_table</span><span class="p">:</span>
|
|
<span class="c1"># if share_embedding_table is enabled, only one copy of the embedding table is store in converted ckpt</span>
|
|
<span class="c1"># this pass is required to make lm_head.weight and vocab_embedding.weight point to the same tensor</span>
|
|
<span class="c1"># however even if share_embedding_table is not enabled, trt would still only keep one copy of the table if the weights are identical</span>
|
|
<span class="n">model</span> <span class="o">=</span> <span class="n">share_embedding</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
|
|
|
|
<span class="c1"># After weight loading</span>
|
|
<span class="k">if</span> <span class="n">use_ootb_moe</span><span class="p">:</span>
|
|
<span class="n">model</span> <span class="o">=</span> <span class="n">to_ootb_moe</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">use_fused_mlp</span><span class="p">:</span>
|
|
<span class="n">model</span> <span class="o">=</span> <span class="n">fuse_gate_mlp</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">gemm_swiglu_plugin_dtype</span><span class="p">,</span>
|
|
<span class="n">low_latency_gemm_swiglu_plugin_dtype</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">use_fused_rg_lru</span><span class="p">:</span>
|
|
<span class="n">model</span> <span class="o">=</span> <span class="n">fuse_rg_lru</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">use_unfused_qkv_gemm</span><span class="p">:</span>
|
|
<span class="n">model</span> <span class="o">=</span> <span class="n">unfuse_qkv_gemm</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">use_prompt_tuning</span><span class="p">:</span>
|
|
<span class="n">model</span> <span class="o">=</span> <span class="n">set_prompt_tuning</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">use_lora</span><span class="p">:</span>
|
|
<span class="n">model</span> <span class="o">=</span> <span class="n">add_lora</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">max_lora_rank</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">use_fp8_context_fmha</span><span class="p">:</span>
|
|
<span class="n">model</span> <span class="o">=</span> <span class="n">set_fp8_context_fhma</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">use_lora</span> <span class="ow">and</span> <span class="n">use_optimize_cross_qkv</span> <span class="ow">is</span> <span class="kc">True</span><span class="p">:</span>
|
|
<span class="c1"># This optimization is not supported when we use lora</span>
|
|
<span class="n">model</span> <span class="o">=</span> <span class="n">optimize_cross_qkv</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
|
|
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">optimize_cross_qkv</span><span class="p">(</span><span class="n">model</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">"""</span>
|
|
<span class="sd"> For cross attention layer, we can skip computing the query of encoder_output.</span>
|
|
<span class="sd"> So, add a new attribute 'kv' in the cross_attention layer. This might lead to</span>
|
|
<span class="sd"> additional memory cost on model size, but save the memory usage on runtime.</span>
|
|
|
|
<span class="sd"> Currently, this function only detect the ColumnLinear and FP8Linear. It does not supports</span>
|
|
<span class="sd"> other quantization now.</span>
|
|
<span class="sd"> """</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">attn</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">attn</span><span class="p">,</span> <span class="n">Attention</span><span class="p">)</span> <span class="ow">and</span> <span class="n">attn</span><span class="o">.</span><span class="n">cross_attention</span> <span class="ow">and</span> \
|
|
<span class="p">(</span><span class="nb">type</span><span class="p">(</span><span class="n">attn</span><span class="o">.</span><span class="n">qkv</span><span class="p">)</span> <span class="o">==</span> <span class="n">ColumnLinear</span> <span class="ow">or</span> <span class="nb">type</span><span class="p">(</span><span class="n">attn</span><span class="o">.</span><span class="n">qkv</span><span class="p">)</span> <span class="o">==</span> <span class="n">FP8Linear</span><span class="p">):</span>
|
|
<span class="n">old_qkv</span> <span class="o">=</span> <span class="n">attn</span><span class="o">.</span><span class="n">qkv</span>
|
|
<span class="n">linear_class</span> <span class="o">=</span> <span class="nb">type</span><span class="p">(</span><span class="n">old_qkv</span><span class="p">)</span>
|
|
<span class="n">new_kv</span> <span class="o">=</span> <span class="n">linear_class</span><span class="p">(</span>
|
|
<span class="n">in_features</span><span class="o">=</span><span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">out_features</span><span class="o">=</span><span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="n">old_qkv</span><span class="o">.</span><span class="n">bias</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">old_qkv</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span>
|
|
<span class="n">tp_group</span><span class="o">=</span><span class="n">old_qkv</span><span class="o">.</span><span class="n">tp_group</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="n">old_qkv</span><span class="o">.</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">gather_output</span><span class="o">=</span><span class="n">old_qkv</span><span class="o">.</span><span class="n">gather_output</span><span class="p">,</span>
|
|
<span class="n">prefer_managed_weight</span><span class="o">=</span><span class="n">old_qkv</span><span class="o">.</span><span class="n">prefer_managed_weight</span><span class="p">,</span>
|
|
<span class="n">is_qkv</span><span class="o">=</span><span class="n">old_qkv</span><span class="o">.</span><span class="n">is_qkv</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="n">old_qkv_weight_value</span> <span class="o">=</span> <span class="n">old_qkv</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span>
|
|
<span class="k">if</span> <span class="p">(</span><span class="n">old_qkv_weight_value</span><span class="o">.</span><span class="n">shape</span> <span class="o">==</span> <span class="n">np</span><span class="o">.</span><span class="n">asarray</span><span class="p">([</span>
|
|
<span class="p">(</span><span class="n">attn</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">+</span> <span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">)</span> <span class="o">*</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span> <span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span>
|
|
<span class="p">]))</span><span class="o">.</span><span class="n">all</span><span class="p">():</span>
|
|
|
|
<span class="n">q_weight</span><span class="p">,</span> <span class="n">kv_weight</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">array_split</span><span class="p">(</span>
|
|
<span class="n">old_qkv_weight_value</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">+</span>
|
|
<span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">,</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span> <span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">),</span>
|
|
<span class="p">[</span><span class="n">attn</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
|
|
<span class="n">new_kv</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">kv_weight</span><span class="o">.</span><span class="n">reshape</span><span class="p">([</span>
|
|
<span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span>
|
|
<span class="p">])</span>
|
|
<span class="k">elif</span> <span class="p">(</span><span class="n">old_qkv_weight_value</span><span class="o">.</span><span class="n">shape</span> <span class="o">==</span> <span class="n">np</span><span class="o">.</span><span class="n">asarray</span><span class="p">([</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="p">(</span><span class="n">attn</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">+</span> <span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">)</span> <span class="o">*</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span>
|
|
<span class="p">]))</span><span class="o">.</span><span class="n">all</span><span class="p">():</span>
|
|
<span class="n">q_weight</span><span class="p">,</span> <span class="n">kv_weight</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">array_split</span><span class="p">(</span>
|
|
<span class="n">old_qkv_weight_value</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">+</span>
|
|
<span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">,</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">),</span> <span class="p">[</span><span class="n">attn</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">new_kv</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">kv_weight</span><span class="o">.</span><span class="n">reshape</span><span class="p">([</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span>
|
|
<span class="p">])</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="kc">False</span>
|
|
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">attn</span><span class="o">.</span><span class="n">qkv</span><span class="p">,</span> <span class="n">FP8Linear</span><span class="p">):</span>
|
|
<span class="n">new_kv</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">old_qkv</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span>
|
|
<span class="n">new_kv</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">old_qkv</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span>
|
|
|
|
<span class="k">if</span> <span class="n">old_qkv</span><span class="o">.</span><span class="n">bias</span><span class="p">:</span>
|
|
<span class="n">q_bias</span><span class="p">,</span> <span class="n">kv_bias</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">array_split</span><span class="p">(</span><span class="n">old_qkv</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">+</span> <span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">,</span>
|
|
<span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">),</span> <span class="p">[</span><span class="n">attn</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
|
|
<span class="n">new_kv</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">kv_bias</span><span class="o">.</span><span class="n">reshape</span><span class="p">([</span>
|
|
<span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span>
|
|
<span class="p">])</span>
|
|
<span class="nb">setattr</span><span class="p">(</span><span class="n">attn</span><span class="p">,</span> <span class="s2">"kv"</span><span class="p">,</span> <span class="n">new_kv</span><span class="p">)</span>
|
|
|
|
<span class="k">return</span> <span class="n">model</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">preprocess_perlayer_weights</span><span class="p">(</span><span class="n">weights</span><span class="p">,</span>
|
|
<span class="n">model_config</span><span class="p">,</span>
|
|
<span class="n">quant_algo</span><span class="p">,</span>
|
|
<span class="n">from_pruned</span><span class="o">=</span><span class="kc">False</span><span class="p">):</span>
|
|
<span class="n">exclude_modules</span> <span class="o">=</span> <span class="n">model_config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">exclude_modules</span>
|
|
<span class="c1"># INT4_AWQ</span>
|
|
<span class="k">if</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span> <span class="ow">or</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16_AWQ</span><span class="p">:</span>
|
|
<span class="n">preprocessor</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">trtllm</span><span class="o">.</span><span class="n">preprocess_weights_for_mixed_gemm</span>
|
|
<span class="k">if</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span><span class="p">:</span>
|
|
<span class="n">activation_type</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">float8_e4m3fn</span>
|
|
<span class="k">elif</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16_AWQ</span><span class="p">:</span>
|
|
<span class="n">activation_type</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">float16</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">from_pruned</span> <span class="ow">and</span> <span class="n">param</span><span class="o">.</span><span class="n">numel</span><span class="p">()</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
|
|
<span class="k">continue</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'weight'</span><span class="p">)</span> <span class="ow">and</span> <span class="n">param</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">torch</span><span class="o">.</span><span class="n">int8</span><span class="p">:</span>
|
|
<span class="n">dtype</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">float16</span>
|
|
<span class="k">if</span> <span class="n">model_config</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="s2">"bfloat16"</span><span class="p">:</span>
|
|
<span class="n">dtype</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">bfloat16</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">preprocessor</span><span class="p">(</span><span class="n">param</span><span class="o">.</span><span class="n">T</span><span class="o">.</span><span class="n">contiguous</span><span class="p">(),</span>
|
|
<span class="n">torch</span><span class="o">.</span><span class="n">quint4x2</span><span class="p">,</span>
|
|
<span class="n">activation_type</span><span class="p">)</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">dtype</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'weights_scaling_factor'</span>
|
|
<span class="p">)</span> <span class="ow">and</span> <span class="n">param</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">></span> <span class="n">param</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</span><span class="p">]:</span>
|
|
<span class="c1"># TODO: refine on supporting ModelOpt HF-AWQ</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span><span class="o">.</span><span class="n">T</span><span class="o">.</span><span class="n">contiguous</span><span class="p">()</span><span class="o">.</span><span class="n">to</span><span class="p">(</span>
|
|
<span class="n">str_dtype_to_torch</span><span class="p">(</span><span class="n">model_config</span><span class="o">.</span><span class="n">dtype</span><span class="p">))</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'prequant_scaling_factor'</span><span class="p">):</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="o">-</span><span class="mi">1</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">model_config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_rank</span> <span class="o">></span> <span class="mi">0</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'attention.dense.bias'</span><span class="p">)</span> <span class="ow">or</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span>
|
|
<span class="s1">'mlp.proj.bias'</span><span class="p">):</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros_like</span><span class="p">(</span><span class="n">param</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span><span class="p">:</span>
|
|
<span class="k">for</span> <span class="n">name</span> <span class="ow">in</span> <span class="nb">list</span><span class="p">(</span><span class="n">weights</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'weights_scaling_factor'</span><span class="p">):</span>
|
|
<span class="n">activation_scaling_factor</span> <span class="o">=</span> <span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span>
|
|
<span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">'weights_scaling_factor'</span><span class="p">,</span>
|
|
<span class="s1">'activation_scaling_factor'</span><span class="p">))</span>
|
|
<span class="n">weights_scaling_factor_2</span> <span class="o">=</span> <span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span>
|
|
<span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">'weights_scaling_factor'</span><span class="p">,</span>
|
|
<span class="s1">'weights_scaling_factor_2'</span><span class="p">))</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">/=</span> <span class="n">weights_scaling_factor_2</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">float16</span><span class="p">)</span><span class="o">.</span><span class="n">view</span><span class="p">(</span>
|
|
<span class="n">str_dtype_to_torch</span><span class="p">(</span><span class="n">model_config</span><span class="o">.</span><span class="n">dtype</span><span class="p">))</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span>
|
|
<span class="s1">'weights_scaling_factor'</span><span class="p">,</span>
|
|
<span class="s1">'prequant_scaling_factor'</span><span class="p">)]</span> <span class="o">/=</span> <span class="n">activation_scaling_factor</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span>
|
|
<span class="s1">'weights_scaling_factor'</span><span class="p">,</span> <span class="s1">'alpha'</span>
|
|
<span class="p">)]</span> <span class="o">=</span> <span class="n">activation_scaling_factor</span> <span class="o">*</span> <span class="n">weights_scaling_factor_2</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">'weights_scaling_factor'</span><span class="p">,</span>
|
|
<span class="s1">'activation_scaling_factor'</span>
|
|
<span class="p">)]</span> <span class="o">=</span> <span class="n">activation_scaling_factor</span>
|
|
|
|
<span class="c1"># FP8</span>
|
|
<span class="k">elif</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'weight'</span><span class="p">)</span> <span class="ow">and</span> <span class="n">param</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">torch</span><span class="o">.</span><span class="n">int8</span><span class="p">:</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">float8_e4m3fn</span><span class="p">)</span>
|
|
<span class="c1"># lm_head is not quantized to FP8</span>
|
|
<span class="k">if</span> <span class="s2">"lm_head.weight"</span> <span class="ow">in</span> <span class="n">weights</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="n">weights</span><span class="p">[</span><span class="s1">'lm_head.weight'</span><span class="p">]</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">str_dtype_to_torch</span><span class="p">(</span>
|
|
<span class="n">model_config</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">'lm_head.weights_scaling_factor'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">'lm_head.activation_scaling_factor'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="k">elif</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8_PER_CHANNEL_PER_TOKEN</span><span class="p">:</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'weight'</span><span class="p">)</span> <span class="ow">and</span> <span class="n">param</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">torch</span><span class="o">.</span><span class="n">int8</span><span class="p">:</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">float8_e4m3fn</span><span class="p">)</span>
|
|
<span class="c1"># lm_head is not quantized to FP8</span>
|
|
<span class="k">if</span> <span class="s2">"lm_head.weight"</span> <span class="ow">in</span> <span class="n">weights</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="n">weights</span><span class="p">[</span><span class="s1">'lm_head.weight'</span><span class="p">]</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">str_dtype_to_torch</span><span class="p">(</span>
|
|
<span class="n">model_config</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">'lm_head.weights_scaling_factor'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">'lm_head.activation_scaling_factor'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="c1"># FP4</span>
|
|
<span class="k">elif</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">NVFP4</span><span class="p">:</span>
|
|
<span class="c1"># Interleave block scale for NVFP4 plugin.</span>
|
|
<span class="k">for</span> <span class="n">name</span> <span class="ow">in</span> <span class="nb">list</span><span class="p">(</span><span class="n">weights</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'weights_scaling_factor'</span><span class="p">):</span>
|
|
<span class="n">ori_shape</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span><span class="o">.</span><span class="n">shape</span>
|
|
<span class="n">weights</span><span class="p">[</span>
|
|
<span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">tensorrt_llm</span><span class="o">.</span><span class="n">nvfp4_block_scale_interleave</span><span class="p">(</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span><span class="o">.</span><span class="n">view</span><span class="p">(</span>
|
|
<span class="n">torch</span><span class="o">.</span><span class="n">int8</span><span class="p">)</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span><span class="o">.</span><span class="n">contiguous</span><span class="p">())</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span>
|
|
<span class="n">ori_shape</span><span class="p">)</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">int32</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'activation_scaling_factor'</span><span class="p">):</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span> <span class="o">+</span> <span class="s1">'_2'</span><span class="p">]</span> <span class="o">=</span> <span class="mf">1.0</span> <span class="o">/</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span>
|
|
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="n">name</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'weights_scaling_factor_2'</span><span class="p">):</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="mf">1.0</span> <span class="o">/</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'weight'</span><span class="p">)</span> <span class="ow">and</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">torch</span><span class="o">.</span><span class="n">uint8</span><span class="p">:</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">int64</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">name</span> <span class="ow">in</span> <span class="nb">list</span><span class="p">(</span><span class="n">weights</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'weights_scaling_factor_2'</span><span class="p">):</span>
|
|
<span class="n">weight_global_sf</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span>
|
|
<span class="n">act_global_sf</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span>
|
|
<span class="s1">'weights_scaling_factor_2'</span><span class="p">,</span> <span class="s1">'activation_scaling_factor_2'</span><span class="p">)]</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span>
|
|
<span class="s1">'weights_scaling_factor_2'</span><span class="p">,</span>
|
|
<span class="s1">'alpha'</span><span class="p">)]</span> <span class="o">=</span> <span class="mf">1.0</span> <span class="o">/</span> <span class="p">(</span><span class="n">act_global_sf</span> <span class="o">*</span> <span class="n">weight_global_sf</span><span class="p">)</span>
|
|
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="n">name</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="k">elif</span> <span class="n">quant_algo</span> <span class="ow">in</span> <span class="p">[</span><span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A16</span><span class="p">]:</span>
|
|
<span class="n">weights</span> <span class="o">=</span> <span class="n">weight_only_quantize_dict</span><span class="p">(</span><span class="n">weights</span><span class="o">=</span><span class="n">weights</span><span class="p">,</span>
|
|
<span class="n">quant_algo</span><span class="o">=</span><span class="n">quant_algo</span><span class="p">,</span>
|
|
<span class="n">exclude_modules</span><span class="o">=</span><span class="n">exclude_modules</span><span class="p">,</span>
|
|
<span class="n">plugin</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">preprocess_weights</span><span class="p">(</span><span class="n">weights</span><span class="p">:</span> <span class="n">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">],</span>
|
|
<span class="n">model_config</span><span class="p">:</span> <span class="n">PretrainedConfig</span><span class="p">,</span>
|
|
<span class="n">from_pruned</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span> <span class="o">-></span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">"""This function in-place modifies weights and model_config, making them compatible with each other.</span>
|
|
|
|
<span class="sd"> Note: Typically, it should be called before model creation and weight loading. For example,</span>
|
|
<span class="sd"> preprocess_weights(weights, model_config)</span>
|
|
<span class="sd"> model = XXXForCausalLM(model_config)</span>
|
|
<span class="sd"> model.load(weights)</span>
|
|
<span class="sd"> """</span>
|
|
<span class="n">quant_config</span> <span class="o">=</span> <span class="n">model_config</span><span class="o">.</span><span class="n">quantization</span>
|
|
<span class="n">quant_algo</span> <span class="o">=</span> <span class="n">quant_config</span><span class="o">.</span><span class="n">quant_algo</span>
|
|
|
|
<span class="n">pattern_info</span> <span class="o">=</span> <span class="p">[</span><span class="s1">'fc'</span><span class="p">,</span> <span class="s1">'gate'</span><span class="p">,</span> <span class="s1">'proj'</span><span class="p">,</span> <span class="s1">'qkv'</span><span class="p">,</span> <span class="s1">'dense'</span><span class="p">]</span>
|
|
|
|
<span class="k">def</span> <span class="nf">add_kv_cache_rcp_scaling_factor</span><span class="p">(</span><span class="n">weights</span><span class="p">:</span> <span class="n">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">]):</span>
|
|
<span class="n">new_entries</span> <span class="o">=</span> <span class="p">[]</span>
|
|
<span class="c1"># The unified converter generate_tllm_weights() already generates these rcp weights, but legacy</span>
|
|
<span class="c1"># converters do not. Handle it here.</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">'.kv_cache_scaling_factor'</span><span class="p">):</span>
|
|
<span class="n">rcp_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">'kv_cache_scaling_factor'</span><span class="p">,</span>
|
|
<span class="s1">'kv_cache_rcp_scaling_factor'</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">rcp_name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">weights</span><span class="p">:</span>
|
|
<span class="n">new_entries</span><span class="o">.</span><span class="n">append</span><span class="p">((</span><span class="n">rcp_name</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">reciprocal</span><span class="p">(</span><span class="n">param</span><span class="p">)))</span>
|
|
<span class="n">weights</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">new_entries</span><span class="p">)</span>
|
|
|
|
<span class="n">add_kv_cache_rcp_scaling_factor</span><span class="p">(</span><span class="n">weights</span><span class="p">)</span>
|
|
|
|
<span class="n">per_layer_weights</span> <span class="o">=</span> <span class="p">{}</span>
|
|
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="n">in_mode</span> <span class="o">=</span> <span class="kc">False</span>
|
|
<span class="k">for</span> <span class="n">info</span> <span class="ow">in</span> <span class="n">pattern_info</span><span class="p">:</span>
|
|
<span class="n">pattern</span> <span class="o">=</span> <span class="sa">rf</span><span class="s1">'(.*?</span><span class="si">{</span><span class="n">info</span><span class="si">}</span><span class="s1">.*?)'</span>
|
|
<span class="n">pattern_match</span> <span class="o">=</span> <span class="n">re</span><span class="o">.</span><span class="n">match</span><span class="p">(</span><span class="n">pattern</span><span class="p">,</span> <span class="n">name</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">pattern_match</span><span class="p">:</span>
|
|
<span class="n">base_name</span> <span class="o">=</span> <span class="n">pattern_match</span><span class="o">.</span><span class="n">group</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">base_name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">per_layer_weights</span><span class="o">.</span><span class="n">keys</span><span class="p">():</span>
|
|
<span class="n">per_layer_weights</span><span class="p">[</span><span class="n">base_name</span><span class="p">]</span> <span class="o">=</span> <span class="p">{}</span>
|
|
<span class="n">per_layer_weights</span><span class="p">[</span><span class="n">base_name</span><span class="p">][</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span>
|
|
<span class="n">in_mode</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="k">break</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">in_mode</span><span class="p">:</span>
|
|
<span class="c1"># [lm_head.weight, ln_f.weight, vocab_embedding.weight]</span>
|
|
<span class="n">base_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">'.'</span><span class="p">,</span> <span class="mi">1</span><span class="p">)[</span><span class="mi">0</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="n">base_name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">per_layer_weights</span><span class="o">.</span><span class="n">keys</span><span class="p">():</span>
|
|
<span class="n">per_layer_weights</span><span class="p">[</span><span class="n">base_name</span><span class="p">]</span> <span class="o">=</span> <span class="p">{}</span>
|
|
<span class="n">per_layer_weights</span><span class="p">[</span><span class="n">base_name</span><span class="p">][</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span>
|
|
|
|
<span class="n">new_weights</span> <span class="o">=</span> <span class="p">{}</span>
|
|
<span class="k">for</span> <span class="n">base_name</span><span class="p">,</span> <span class="n">layer_weights</span> <span class="ow">in</span> <span class="n">per_layer_weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">quant_algo</span> <span class="o">!=</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">MIXED_PRECISION</span><span class="p">:</span>
|
|
<span class="n">layer_quant_algo</span> <span class="o">=</span> <span class="n">quant_algo</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="n">base_name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">quant_config</span><span class="o">.</span><span class="n">quantized_layers</span><span class="o">.</span><span class="n">keys</span><span class="p">():</span>
|
|
<span class="n">new_weights</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">layer_weights</span><span class="p">)</span>
|
|
<span class="k">continue</span>
|
|
<span class="n">layer_quant_algo</span> <span class="o">=</span> <span class="n">quant_config</span><span class="o">.</span><span class="n">quantized_layers</span><span class="p">[</span>
|
|
<span class="n">base_name</span><span class="p">]</span><span class="o">.</span><span class="n">quant_algo</span>
|
|
|
|
<span class="n">preprocess_perlayer_weights</span><span class="p">(</span><span class="n">layer_weights</span><span class="p">,</span> <span class="n">model_config</span><span class="p">,</span>
|
|
<span class="n">layer_quant_algo</span><span class="p">,</span> <span class="n">from_pruned</span><span class="p">)</span>
|
|
<span class="n">new_weights</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">layer_weights</span><span class="p">)</span>
|
|
|
|
<span class="n">weights</span> <span class="o">=</span> <span class="n">new_weights</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">model_config</span><span class="o">.</span><span class="n">architecture</span> <span class="o">==</span> <span class="s1">'GPTJForCausalLM'</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="n">model_config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_rank</span> <span class="o">></span> <span class="mi">0</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="s1">'attention.dense.bias'</span> <span class="ow">in</span> <span class="n">name</span> <span class="ow">or</span> <span class="s1">'mlp.proj.bias'</span> <span class="ow">in</span> <span class="n">name</span><span class="p">:</span>
|
|
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros_like</span><span class="p">(</span><span class="n">param</span><span class="p">)</span>
|
|
|
|
<span class="k">return</span> <span class="n">weights</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">get_kv_cache_type_from_legacy</span><span class="p">(</span><span class="n">use_cache</span><span class="p">:</span> <span class="nb">bool</span><span class="p">,</span>
|
|
<span class="n">paged_kv_cache</span><span class="p">:</span> <span class="nb">bool</span><span class="p">)</span> <span class="o">-></span> <span class="n">KVCacheType</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="n">use_cache</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="n">paged_kv_cache</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">PAGED</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">CONTINUOUS</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">DISABLED</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">save_config</span><span class="p">(</span><span class="n">config</span><span class="p">:</span> <span class="n">PretrainedConfig</span><span class="p">,</span> <span class="o">*</span><span class="p">,</span> <span class="n">output_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
|
|
<span class="n">log</span><span class="p">:</span> <span class="nb">bool</span><span class="p">)</span> <span class="o">-></span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">config_path</span> <span class="o">=</span> <span class="n">Path</span><span class="p">(</span><span class="n">output_dir</span><span class="p">)</span> <span class="o">/</span> <span class="s2">"config.json"</span>
|
|
<span class="k">if</span> <span class="n">log</span><span class="p">:</span>
|
|
<span class="n">logger</span><span class="o">.</span><span class="n">debug</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Saving TensorRT-LLM configuration to </span><span class="si">{</span><span class="n">config_path</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
|
|
<span class="n">config_path</span><span class="o">.</span><span class="n">parent</span><span class="o">.</span><span class="n">mkdir</span><span class="p">(</span><span class="n">exist_ok</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">parents</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
|
<span class="n">config_path</span><span class="o">.</span><span class="n">write_text</span><span class="p">(</span><span class="n">json</span><span class="o">.</span><span class="n">dumps</span><span class="p">(</span><span class="n">config</span><span class="o">.</span><span class="n">to_dict</span><span class="p">(),</span> <span class="n">indent</span><span class="o">=</span><span class="mi">4</span><span class="p">))</span>
|
|
|
|
|
|
<span class="k">def</span> <span class="nf">save_checkpoint</span><span class="p">(</span><span class="o">*</span><span class="p">,</span> <span class="n">output_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span> <span class="n">weights</span><span class="p">:</span> <span class="nb">dict</span><span class="p">,</span> <span class="n">rank</span><span class="p">:</span> <span class="nb">int</span><span class="p">)</span> <span class="o">-></span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">""" Checkpoint saver for weight loader."""</span>
|
|
<span class="n">safetensors</span><span class="o">.</span><span class="n">torch</span><span class="o">.</span><span class="n">save_file</span><span class="p">(</span>
|
|
<span class="n">weights</span><span class="p">,</span> <span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">output_dir</span><span class="p">,</span> <span class="sa">f</span><span class="s1">'rank</span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s1">.safetensors'</span><span class="p">))</span>
|
|
</pre></div>
|
|
|
|
</div>
|
|
</div>
|
|
<footer>
|
|
|
|
<hr/>
|
|
|
|
<div role="contentinfo">
|
|
<jinja2.runtime.BlockReference object at 0x7f5e6e4a5700>
|
|
|
|
<div class="footer">
|
|
<p>
|
|
Copyright © 2024 NVIDIA Corporation
|
|
</p>
|
|
<p>
|
|
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/privacy-policy/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Privacy Policy</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/privacy-center/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Manage My Privacy</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/preferences/start/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Do Not Sell or Share My Data</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/terms-of-service/" target="_blank"
|
|
rel="noopener" data-cms-ai="0">Terms of Service</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/accessibility/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Accessibility</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/company-policies/" target="_blank"
|
|
rel="noopener" data-cms-ai="0">Corporate Policies</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/product-security/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Product Security</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/contact/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Contact</a>
|
|
</p>
|
|
</div>
|
|
|
|
|
|
</div>
|
|
|
|
|
|
|
|
</footer>
|
|
</div>
|
|
</div>
|
|
</section>
|
|
</div>
|
|
<script>
|
|
jQuery(function () {
|
|
SphinxRtdTheme.Navigation.enable(true);
|
|
});
|
|
</script>
|
|
|
|
</body>
|
|
</html> |