kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-23 20:23:08 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity
1d1cd4259a
TensorRT-LLMs/_modules/tensorrt_llm/models/llama/model.html
石晓伟 1d1cd4259a
Update gh-pages (#1196)
2024-02-29 20:56:26 +08:00

617 lines
75 KiB
HTML
Raw Blame History

<!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.llama.model &mdash; 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=19f00094" />
<!--[if lt IE 9]>
<script src="../../../../_static/js/html5shiv.min.js"></script>
<![endif]-->
<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=888ff710"></script>
<script src="../../../../_static/sphinx_highlight.js?v=dc90522c"></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">Contents:</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../../../../architecture.html">TensorRT-LLM Architecture</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../gpt_runtime.html">C++ GPT Runtime</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../batch_manager.html">The Batch Manager in TensorRT-LLM</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../inference_request.html">Inference Request</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../gpt_attention.html">Multi-head, Multi-query and Group-query Attention</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../precision.html">Numerical Precision</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../build_from_source.html">Build from Source</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../performance.html">Performance of TensorRT-LLM</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../2023-05-19-how-to-debug.html">How to debug</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../2023-05-17-how-to-add-a-new-model.html">How to add a new model</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../graph-rewriting.html">Graph Rewriting Module</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../memory.html">Memory Usage of TensorRT-LLM</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../new_workflow.html">New Workflow</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../lora.html">Run gpt-2b + LoRA using GptManager / cpp runtime</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../perf_best_practices.html">Best Practices for Tuning the Performance of TensorRT-LLM</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../../performance_analysis.html">Performance Analysis of TensorRT-LLM</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Python 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/runtime.html">Runtime</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>
</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.llama.model</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.llama.model</h1><div class="highlight"><pre>
<span></span><span class="c1"># SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION &amp; AFFILIATES. All rights reserved.</span>
<span class="c1"># SPDX-License-Identifier: Apache-2.0</span>
<span class="c1">#</span>
<span class="c1"># Licensed under the Apache License, Version 2.0 (the &quot;License&quot;);</span>
<span class="c1"># you may not use this file except in compliance with the License.</span>
<span class="c1"># You may obtain a copy of the License at</span>
<span class="c1">#</span>
<span class="c1"># http://www.apache.org/licenses/LICENSE-2.0</span>
<span class="c1">#</span>
<span class="c1"># Unless required by applicable law or agreed to in writing, software</span>
<span class="c1"># distributed under the License is distributed on an &quot;AS IS&quot; BASIS,</span>
<span class="c1"># WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.</span>
<span class="c1"># See the License for the specific language governing permissions and</span>
<span class="c1"># limitations under the License.</span>
<span class="kn">import</span> <span class="nn">tempfile</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="n">Optional</span>
<span class="kn">from</span> <span class="nn">transformers</span> <span class="kn">import</span> <span class="n">AutoConfig</span>
<span class="kn">from</span> <span class="nn">tensorrt_llm.models.llama.weight</span> <span class="kn">import</span> <span class="p">(</span><span class="n">load_from_awq_llama</span><span class="p">,</span>
<span class="n">load_from_fp8_llama</span><span class="p">)</span>
<span class="kn">from</span> <span class="nn">...</span> <span class="kn">import</span> <span class="n">profiler</span>
<span class="kn">from</span> <span class="nn">..._utils</span> <span class="kn">import</span> <span class="n">pad_vocab_size</span>
<span class="kn">from</span> <span class="nn">...functional</span> <span class="kn">import</span> <span class="n">RotaryScalingType</span><span class="p">,</span> <span class="n">Tensor</span><span class="p">,</span> <span class="n">recv</span><span class="p">,</span> <span class="n">send</span>
<span class="kn">from</span> <span class="nn">...layers</span> <span class="kn">import</span> <span class="p">(</span><span class="n">MOE</span><span class="p">,</span> <span class="n">Attention</span><span class="p">,</span> <span class="n">AttentionMaskType</span><span class="p">,</span> <span class="n">ColumnLinear</span><span class="p">,</span>
<span class="n">Embedding</span><span class="p">,</span> <span class="n">GatedMLP</span><span class="p">,</span> <span class="n">MoeConfig</span><span class="p">,</span> <span class="n">PositionEmbeddingType</span><span class="p">,</span>
<span class="n">PromptTuningEmbedding</span><span class="p">,</span> <span class="n">RmsNorm</span><span class="p">)</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="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">...runtime.lora_manager</span> <span class="kn">import</span> <span class="n">LoraConfig</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">..modeling_utils</span> <span class="kn">import</span> <span class="p">(</span><span class="n">DecoderLayerList</span><span class="p">,</span> <span class="n">DecoderModelForCausalLM</span><span class="p">,</span>
<span class="n">PretrainedConfig</span><span class="p">)</span>
<span class="kn">from</span> <span class="nn">.weight</span> <span class="kn">import</span> <span class="n">load_from_hf_llama</span>
<span class="k">class</span> <span class="nc">LLaMADecoderLayer</span><span class="p">(</span><span class="n">Module</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">layer_idx</span><span class="p">:</span> <span class="nb">int</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="bp">self</span><span class="o">.</span><span class="n">layer_idx</span> <span class="o">=</span> <span class="n">layer_idx</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="bp">self</span><span class="o">.</span><span class="n">input_layernorm</span> <span class="o">=</span> <span class="n">RmsNorm</span><span class="p">(</span><span class="n">normalized_shape</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">eps</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">norm_epsilon</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="bp">self</span><span class="o">.</span><span class="n">attention</span> <span class="o">=</span> <span class="n">Attention</span><span class="p">(</span>
<span class="n">config</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="n">config</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">,</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">max_position_embeddings</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">max_position_embeddings</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">attention_mask_type</span><span class="o">=</span><span class="n">AttentionMaskType</span><span class="o">.</span><span class="n">causal</span><span class="p">,</span>
<span class="n">bias</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">attn_bias</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">rope_gpt_neox</span><span class="p">,</span>
<span class="n">rotary_embedding_base</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">rotary_base</span><span class="p">,</span>
<span class="n">rotary_embedding_scaling</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">rotary_scaling</span><span class="p">,</span>
<span class="n">tp_group</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="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">tp_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">tp_rank</span><span class="p">,</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="p">,</span>
<span class="n">enable_pos_shift</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">enable_pos_shift</span><span class="p">,</span>
<span class="n">dense_context_fmha</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">dense_context_fmha</span><span class="p">,</span>
<span class="n">max_lora_rank</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">max_lora_rank</span><span class="p">)</span>
<span class="n">mlp_hidden_size</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">hidden_size</span> <span class="o">*</span> <span class="mi">4</span> <span class="k">if</span> <span class="n">config</span><span class="o">.</span><span class="n">intermediate_size</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">config</span><span class="o">.</span><span class="n">intermediate_size</span>
<span class="n">ClsMLP</span> <span class="o">=</span> <span class="n">GatedMLP</span>
<span class="n">mlp_kwargs</span> <span class="o">=</span> <span class="p">{}</span>
<span class="k">if</span> <span class="n">config</span><span class="o">.</span><span class="n">moe_num_experts</span> <span class="o">&gt;</span> <span class="mi">1</span><span class="p">:</span>
<span class="n">ClsMLP</span> <span class="o">=</span> <span class="n">MOE</span>
<span class="n">mlp_kwargs</span> <span class="o">=</span> <span class="p">{</span>
<span class="s2">&quot;moe_config&quot;</span><span class="p">:</span>
<span class="n">MoeConfig</span><span class="p">(</span>
<span class="n">config</span><span class="o">.</span><span class="n">moe_num_experts</span><span class="p">,</span>
<span class="n">config</span><span class="o">.</span><span class="n">moe_top_k</span><span class="p">,</span>
<span class="n">config</span><span class="o">.</span><span class="n">moe_tp_mode</span><span class="p">,</span>
<span class="n">config</span><span class="o">.</span><span class="n">moe_normalization_mode</span><span class="p">,</span>
<span class="p">),</span>
<span class="s2">&quot;tp_rank&quot;</span><span class="p">:</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="p">,</span>
<span class="p">}</span>
<span class="bp">self</span><span class="o">.</span><span class="n">mlp</span> <span class="o">=</span> <span class="n">ClsMLP</span><span class="p">(</span><span class="n">hidden_size</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">ffn_hidden_size</span><span class="o">=</span><span class="n">mlp_hidden_size</span><span class="p">,</span>
<span class="n">hidden_act</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">hidden_act</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">bias</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">mlp_bias</span><span class="p">,</span>
<span class="n">tp_group</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="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">quant_mode</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">quant_mode</span><span class="p">,</span>
<span class="n">max_lora_rank</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">max_lora_rank</span><span class="p">,</span>
<span class="o">**</span><span class="n">mlp_kwargs</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">post_layernorm</span> <span class="o">=</span> <span class="n">RmsNorm</span><span class="p">(</span><span class="n">normalized_shape</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">eps</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">norm_epsilon</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="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">attention_mask</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">medusa_packed_mask</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="c1"># For Medusa support</span>
<span class="n">medusa_position_offsets</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">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">lora_layer_params</span><span class="o">=</span><span class="kc">None</span><span class="p">):</span>
<span class="n">residual</span> <span class="o">=</span> <span class="n">hidden_states</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">input_layernorm</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">)</span>
<span class="n">attention_output</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention</span><span class="p">(</span>
<span class="n">hidden_states</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">medusa_packed_mask</span><span class="o">=</span><span class="n">medusa_packed_mask</span><span class="p">,</span> <span class="c1"># For Medusa support</span>
<span class="n">medusa_position_offsets</span><span class="o">=</span><span class="n">medusa_position_offsets</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">kv_cache_params</span><span class="o">=</span><span class="n">kv_cache_params</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="n">lora_layer_params</span><span class="o">=</span><span class="n">lora_layer_params</span><span class="p">)</span>
<span class="k">if</span> <span class="n">use_cache</span><span class="p">:</span>
<span class="n">attention_output</span><span class="p">,</span> <span class="n">presents</span> <span class="o">=</span> <span class="n">attention_output</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">residual</span> <span class="o">+</span> <span class="n">attention_output</span>
<span class="n">residual</span> <span class="o">=</span> <span class="n">hidden_states</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">post_layernorm</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">)</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">mlp</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span>
<span class="n">lora_layer_params</span><span class="o">=</span><span class="n">lora_layer_params</span><span class="p">)</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">residual</span> <span class="o">+</span> <span class="n">hidden_states</span>
<span class="k">if</span> <span class="n">use_cache</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">return</span> <span class="n">hidden_states</span>
<div class="viewcode-block" id="LLaMAModel">
<a class="viewcode-back" href="../../../../python-api/tensorrt_llm.models.html#tensorrt_llm.models.LLaMAModel">[docs]</a>
<span class="k">class</span> <span class="nc">LLaMAModel</span><span class="p">(</span><span class="n">Module</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="o">-&gt;</span> <span class="kc">None</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">mapping</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span>
<span class="bp">self</span><span class="o">.</span><span class="n">use_prompt_tuning</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">use_prompt_tuning</span>
<span class="n">EmbeddingCls</span> <span class="o">=</span> <span class="n">PromptTuningEmbedding</span> <span class="k">if</span> <span class="n">config</span><span class="o">.</span><span class="n">use_prompt_tuning</span> <span class="k">else</span> <span class="n">Embedding</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">is_first_pp_rank</span><span class="p">():</span>
<span class="bp">self</span><span class="o">.</span><span class="n">vocab_embedding</span> <span class="o">=</span> <span class="n">EmbeddingCls</span><span class="p">(</span>
<span class="n">num_embeddings</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">vocab_size</span><span class="p">,</span>
<span class="n">embedding_dim</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">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">tp_size</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">tp_size</span>
<span class="k">if</span> <span class="n">config</span><span class="o">.</span><span class="n">use_parallel_embedding</span> <span class="k">else</span> <span class="mi">1</span><span class="p">,</span>
<span class="n">tp_group</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">tp_group</span>
<span class="k">if</span> <span class="n">config</span><span class="o">.</span><span class="n">use_parallel_embedding</span> <span class="k">else</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">sharding_dim</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">embedding_sharding_dim</span><span class="p">,</span>
<span class="n">tp_rank</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">tp_rank</span><span class="p">,</span>
<span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">layers</span> <span class="o">=</span> <span class="n">DecoderLayerList</span><span class="p">(</span><span class="n">LLaMADecoderLayer</span><span class="p">,</span> <span class="n">config</span><span class="p">)</span>
<span class="k">if</span> <span class="bp">self</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="bp">self</span><span class="o">.</span><span class="n">ln_f</span> <span class="o">=</span> <span class="n">RmsNorm</span><span class="p">(</span><span class="n">normalized_shape</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">eps</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">norm_epsilon</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>
<div class="viewcode-block" id="LLaMAModel.forward">
<a class="viewcode-back" href="../../../../python-api/tensorrt_llm.models.html#tensorrt_llm.models.LLaMAModel.forward">[docs]</a>
<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">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">attention_mask</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">medusa_position_offsets</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="c1"># For Medusa support</span>
<span class="n">medusa_packed_mask</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="c1"># For Medusa support</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">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">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">layers</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="n">ptuning_args</span> <span class="o">=</span> <span class="p">[</span>
<span class="n">prompt_embedding_table</span><span class="p">,</span> <span class="n">prompt_tasks</span><span class="p">,</span> <span class="n">prompt_vocab_size</span>
<span class="p">]</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_prompt_tuning</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">mapping</span><span class="o">.</span><span class="n">is_first_pp_rank</span><span class="p">():</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">vocab_embedding</span><span class="p">(</span><span class="n">input_ids</span><span class="p">,</span> <span class="o">*</span><span class="n">ptuning_args</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">recv</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">mapping</span><span class="o">.</span><span class="n">prev_pp_rank</span><span class="p">())</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">layers</span><span class="o">.</span><span class="n">forward</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">kv_cache_params</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="n">lora_params</span><span class="o">=</span><span class="n">lora_params</span><span class="p">,</span>
<span class="n">medusa_position_offsets</span><span class="o">=</span><span class="n">medusa_position_offsets</span><span class="p">,</span>
<span class="n">medusa_packed_mask</span><span class="o">=</span><span class="n">medusa_packed_mask</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="k">if</span> <span class="bp">self</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">hidden_states</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">ln_f</span><span class="p">(</span><span class="n">hidden_states</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">send</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">mapping</span><span class="o">.</span><span class="n">next_pp_rank</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="p">(</span><span class="n">hidden_states</span><span class="p">,</span> <span class="nb">tuple</span><span class="p">(</span><span class="n">presents</span><span class="p">))</span>
<span class="k">return</span> <span class="n">hidden_states</span></div>
</div>
<div class="viewcode-block" id="LLaMAForCausalLM">
<a class="viewcode-back" href="../../../../python-api/tensorrt_llm.models.html#tensorrt_llm.models.LLaMAForCausalLM">[docs]</a>
<span class="k">class</span> <span class="nc">LLaMAForCausalLM</span><span class="p">(</span><span class="n">DecoderModelForCausalLM</span><span class="p">,</span> <span class="n">TopModelMixin</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="bp">self</span><span class="o">.</span><span class="n">check_config</span><span class="p">(</span><span class="n">config</span><span class="p">)</span>
<span class="n">transformer</span> <span class="o">=</span> <span class="n">LLaMAModel</span><span class="p">(</span><span class="n">config</span><span class="p">)</span>
<span class="n">vocab_size_padded</span> <span class="o">=</span> <span class="n">pad_vocab_size</span><span class="p">(</span><span class="n">config</span><span class="o">.</span><span class="n">vocab_size</span><span class="p">,</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="k">if</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">lm_head</span> <span class="o">=</span> <span class="n">ColumnLinear</span><span class="p">(</span><span class="n">config</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="n">vocab_size_padded</span><span class="p">,</span>
<span class="n">bias</span><span class="o">=</span><span class="kc">False</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">tp_group</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="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">gather_output</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">lm_head</span> <span class="o">=</span> <span class="kc">None</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="bp">self</span><span class="o">.</span><span class="n">mapping</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</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="n">transformer</span><span class="p">,</span> <span class="n">lm_head</span><span class="p">)</span>
<div class="viewcode-block" id="LLaMAForCausalLM.check_config">
<a class="viewcode-back" href="../../../../python-api/tensorrt_llm.models.html#tensorrt_llm.models.LLaMAForCausalLM.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="n">config</span><span class="o">.</span><span class="n">set_if_not_exist</span><span class="p">(</span><span class="s1">&#39;mlp_bias&#39;</span><span class="p">,</span> <span class="kc">False</span><span class="p">)</span>
<span class="n">config</span><span class="o">.</span><span class="n">set_if_not_exist</span><span class="p">(</span><span class="s1">&#39;attn_bias&#39;</span><span class="p">,</span> <span class="kc">False</span><span class="p">)</span>
<span class="n">config</span><span class="o">.</span><span class="n">set_if_not_exist</span><span class="p">(</span><span class="s1">&#39;rotary_base&#39;</span><span class="p">,</span> <span class="mf">10000.0</span><span class="p">)</span>
<span class="n">config</span><span class="o">.</span><span class="n">set_if_not_exist</span><span class="p">(</span><span class="s1">&#39;rotary_scaling&#39;</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
<span class="n">config</span><span class="o">.</span><span class="n">set_if_not_exist</span><span class="p">(</span><span class="s1">&#39;enable_pos_shift&#39;</span><span class="p">,</span> <span class="kc">False</span><span class="p">)</span>
<span class="n">config</span><span class="o">.</span><span class="n">set_if_not_exist</span><span class="p">(</span><span class="s1">&#39;dense_context_fmha&#39;</span><span class="p">,</span> <span class="kc">False</span><span class="p">)</span>
<span class="n">config</span><span class="o">.</span><span class="n">set_if_not_exist</span><span class="p">(</span><span class="s1">&#39;moe_num_experts&#39;</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
<span class="n">config</span><span class="o">.</span><span class="n">set_if_not_exist</span><span class="p">(</span><span class="s1">&#39;moe_top_k&#39;</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
<span class="n">config</span><span class="o">.</span><span class="n">set_if_not_exist</span><span class="p">(</span><span class="s1">&#39;moe_tp_mode&#39;</span><span class="p">,</span>
<span class="n">MoeConfig</span><span class="o">.</span><span class="n">ParallelismMode</span><span class="o">.</span><span class="n">TENSOR_PARALLEL</span><span class="p">)</span>
<span class="n">config</span><span class="o">.</span><span class="n">set_if_not_exist</span><span class="p">(</span>
<span class="s1">&#39;moe_normalization_mode&#39;</span><span class="p">,</span>
<span class="n">MoeConfig</span><span class="o">.</span><span class="n">ExpertScaleNormalizationMode</span><span class="o">.</span><span class="n">RENORMALIZE</span><span class="p">)</span></div>
<div class="viewcode-block" id="LLaMAForCausalLM.from_hugging_face">
<a class="viewcode-back" href="../../../../python-api/tensorrt_llm.models.html#tensorrt_llm.models.LLaMAForCausalLM.from_hugging_face">[docs]</a>
<span class="nd">@classmethod</span>
<span class="k">def</span> <span class="nf">from_hugging_face</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="n">dtype</span><span class="o">=</span><span class="s1">&#39;float16&#39;</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_mode</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantMode</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="kn">import</span> <span class="nn">transformers</span>
<span class="kn">from</span> <span class="nn">transformers</span> <span class="kn">import</span> <span class="n">LlamaConfig</span>
<span class="kn">from</span> <span class="nn">...models.modeling_utils</span> <span class="kn">import</span> <span class="n">PretrainedConfig</span>
<span class="n">cfg</span> <span class="o">=</span> <span class="n">LlamaConfig</span><span class="o">.</span><span class="n">from_pretrained</span><span class="p">(</span><span class="n">hf_model_dir</span><span class="p">)</span>
<span class="n">num_kv_heads</span> <span class="o">=</span> <span class="n">cfg</span><span class="o">.</span><span class="n">num_key_value_heads</span> <span class="k">if</span> <span class="nb">hasattr</span><span class="p">(</span><span class="n">cfg</span><span class="p">,</span> <span class="s2">&quot;num_key_value_heads&quot;</span><span class="p">)</span> \
<span class="k">else</span> <span class="n">cfg</span><span class="o">.</span><span class="n">num_attention_heads</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">if</span> <span class="n">quant_mode</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">quant_mode</span> <span class="o">=</span> <span class="n">QuantMode</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
<span class="n">cfg</span><span class="o">.</span><span class="n">mapping</span> <span class="o">=</span> <span class="n">mapping</span>
<span class="n">cfg</span><span class="o">.</span><span class="n">dtype</span> <span class="o">=</span> <span class="n">dtype</span>
<span class="n">cfg</span><span class="o">.</span><span class="n">quant_mode</span> <span class="o">=</span> <span class="n">quant_mode</span>
<span class="n">moe_config</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="s2">&quot;moe_config&quot;</span><span class="p">,</span> <span class="n">MoeConfig</span><span class="p">())</span>
<span class="n">cfg</span><span class="o">.</span><span class="n">norm_epsilon</span> <span class="o">=</span> <span class="n">cfg</span><span class="o">.</span><span class="n">rms_norm_eps</span>
<span class="n">config</span> <span class="o">=</span> <span class="p">{</span>
<span class="s1">&#39;architecture&#39;</span><span class="p">:</span> <span class="n">cfg</span><span class="o">.</span><span class="n">architectures</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
<span class="s1">&#39;dtype&#39;</span><span class="p">:</span> <span class="n">cfg</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span>
<span class="s1">&#39;logits_dtype&#39;</span><span class="p">:</span> <span class="s1">&#39;float32&#39;</span><span class="p">,</span>
<span class="s1">&#39;num_hidden_layers&#39;</span><span class="p">:</span> <span class="n">cfg</span><span class="o">.</span><span class="n">num_hidden_layers</span><span class="p">,</span>
<span class="s1">&#39;num_attention_heads&#39;</span><span class="p">:</span> <span class="n">cfg</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">,</span>
<span class="s1">&#39;hidden_size&#39;</span><span class="p">:</span> <span class="n">cfg</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="s1">&#39;intermediate_size&#39;</span><span class="p">:</span> <span class="n">cfg</span><span class="o">.</span><span class="n">intermediate_size</span><span class="p">,</span>
<span class="s1">&#39;num_key_value_heads&#39;</span><span class="p">:</span> <span class="n">cfg</span><span class="o">.</span><span class="n">num_key_value_heads</span><span class="p">,</span>
<span class="s1">&#39;vocab_size&#39;</span><span class="p">:</span> <span class="n">cfg</span><span class="o">.</span><span class="n">vocab_size</span><span class="p">,</span>
<span class="s1">&#39;position_embedding_type&#39;</span><span class="p">:</span> <span class="s1">&#39;rope_gpt_neox&#39;</span><span class="p">,</span>
<span class="s1">&#39;max_position_embeddings&#39;</span><span class="p">:</span> <span class="n">cfg</span><span class="o">.</span><span class="n">max_position_embeddings</span><span class="p">,</span>
<span class="s1">&#39;hidden_act&#39;</span><span class="p">:</span> <span class="n">cfg</span><span class="o">.</span><span class="n">hidden_act</span><span class="p">,</span>
<span class="s1">&#39;rotary_base&#39;</span><span class="p">:</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">cfg</span><span class="p">,</span> <span class="s1">&#39;rotary_base&#39;</span><span class="p">,</span> <span class="mf">10000.0</span><span class="p">),</span>
<span class="s1">&#39;rotary_scaling&#39;</span><span class="p">:</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">cfg</span><span class="p">,</span> <span class="s1">&#39;rotary_scaling&#39;</span><span class="p">,</span> <span class="kc">None</span><span class="p">),</span>
<span class="s1">&#39;norm_epsilon&#39;</span><span class="p">:</span> <span class="n">cfg</span><span class="o">.</span><span class="n">rms_norm_eps</span><span class="p">,</span>
<span class="s1">&#39;quantization&#39;</span><span class="p">:</span> <span class="p">{</span>
<span class="s1">&#39;group_size&#39;</span><span class="p">:</span> <span class="mi">128</span><span class="p">,</span>
<span class="p">},</span>
<span class="s1">&#39;mapping&#39;</span><span class="p">:</span> <span class="p">{</span>
<span class="s1">&#39;world_size&#39;</span><span class="p">:</span> <span class="n">mapping</span><span class="o">.</span><span class="n">world_size</span><span class="p">,</span>
<span class="s1">&#39;tp_size&#39;</span><span class="p">:</span> <span class="n">mapping</span><span class="o">.</span><span class="n">world_size</span><span class="p">,</span>
<span class="p">},</span>
<span class="s1">&#39;use_parallel_embedding&#39;</span><span class="p">:</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">&quot;use_parallel_embedding&quot;</span><span class="p">,</span>
<span class="kc">False</span><span class="p">),</span>
<span class="s1">&#39;embedding_sharding_dim&#39;</span><span class="p">:</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">&quot;embedding_sharding_dim&quot;</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
<span class="s1">&#39;use_prompt_tuning&#39;</span><span class="p">:</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">&quot;use_prompt_tuning&quot;</span><span class="p">,</span> <span class="kc">False</span><span class="p">),</span>
<span class="s1">&#39;moe_num_experts&#39;</span><span class="p">:</span> <span class="n">moe_config</span><span class="o">.</span><span class="n">num_experts</span><span class="p">,</span>
<span class="s1">&#39;moe_top_k&#39;</span><span class="p">:</span> <span class="n">moe_config</span><span class="o">.</span><span class="n">top_k</span><span class="p">,</span>
<span class="s1">&#39;moe_tp_mode&#39;</span><span class="p">:</span> <span class="n">moe_config</span><span class="o">.</span><span class="n">tp_mode</span><span class="p">,</span>
<span class="s1">&#39;moe_normalization_mode&#39;</span><span class="p">:</span> <span class="n">moe_config</span><span class="o">.</span><span class="n">normalization_mode</span><span class="p">,</span>
<span class="s1">&#39;use_fused_mlp&#39;</span><span class="p">:</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">&quot;use_fused_mlp&quot;</span><span class="p">,</span> <span class="kc">False</span><span class="p">),</span>
<span class="s1">&#39;enable_pos_shift&#39;</span><span class="p">:</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">&quot;enable_pos_shift&quot;</span><span class="p">,</span> <span class="kc">False</span><span class="p">),</span>
<span class="s1">&#39;dense_context_fmha&#39;</span><span class="p">:</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">&quot;dense_context_fmha&quot;</span><span class="p">,</span> <span class="kc">False</span><span class="p">),</span>
<span class="p">}</span>
<span class="k">if</span> <span class="n">quant_mode</span><span class="o">.</span><span class="n">is_int4_weight_only_per_group</span><span class="p">():</span>
<span class="n">config</span><span class="p">[</span><span class="s1">&#39;quantization&#39;</span><span class="p">]</span><span class="o">.</span><span class="n">update</span><span class="p">({</span>
<span class="s1">&#39;quant_algo&#39;</span><span class="p">:</span> <span class="s1">&#39;W4A8_AWQ&#39;</span><span class="p">,</span>
<span class="s1">&#39;has_zero_point&#39;</span><span class="p">:</span> <span class="kc">False</span><span class="p">,</span>
<span class="s1">&#39;pre_quant_scale&#39;</span><span class="p">:</span> <span class="kc">True</span><span class="p">,</span>
<span class="s1">&#39;exclude_modules&#39;</span><span class="p">:</span> <span class="p">[],</span>
<span class="p">})</span>
<span class="k">elif</span> <span class="n">quant_mode</span><span class="o">.</span><span class="n">has_fp8_qdq</span><span class="p">()</span> <span class="ow">and</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="n">config</span><span class="p">[</span><span class="s1">&#39;quantization&#39;</span><span class="p">]</span><span class="o">.</span><span class="n">update</span><span class="p">({</span>
<span class="s1">&#39;quant_algo&#39;</span><span class="p">:</span> <span class="s1">&#39;FP8&#39;</span><span class="p">,</span>
<span class="s1">&#39;kv_cache_quant_algo&#39;</span><span class="p">:</span> <span class="s1">&#39;FP8&#39;</span>
<span class="p">})</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">if</span> <span class="n">quant_mode</span> <span class="o">!=</span> <span class="n">QuantMode</span><span class="p">(</span><span class="mi">0</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">&quot;Unsupported quantization mode: </span><span class="si">{</span><span class="n">quant_mode</span><span class="si">}</span><span class="s2">&quot;</span><span class="p">)</span>
<span class="n">tllm_llama</span> <span class="o">=</span> <span class="n">LLaMAForCausalLM</span><span class="p">(</span><span class="n">PretrainedConfig</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="n">q_weights</span> <span class="o">=</span> <span class="p">{}</span>
<span class="k">if</span> <span class="n">quant_mode</span><span class="o">.</span><span class="n">has_any_quant</span><span class="p">():</span>
<span class="n">q_weights</span> <span class="o">=</span> <span class="n">tllm_llama</span><span class="o">.</span><span class="n">_quantize</span><span class="p">(</span><span class="n">hf_model_dir</span><span class="p">,</span> <span class="n">dtype</span><span class="p">,</span> <span class="n">cfg</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">)</span>
<span class="c1"># For debug purpose, skip weights loading to be faster</span>
<span class="k">if</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">&quot;skip_loading_weights&quot;</span><span class="p">,</span> <span class="kc">False</span><span class="p">):</span>
<span class="k">return</span> <span class="n">tllm_llama</span>
<span class="c1"># TODO: support mixtral</span>
<span class="c1"># weights already loaded in _quantize for int4 weight only</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">quant_mode</span><span class="o">.</span><span class="n">is_int4_weight_only_per_group</span><span class="p">():</span>
<span class="n">hf_model</span> <span class="o">=</span> <span class="n">transformers</span><span class="o">.</span><span class="n">LlamaForCausalLM</span>
<span class="n">profiler</span><span class="o">.</span><span class="n">start</span><span class="p">(</span><span class="s2">&quot;Loading weights from HF&quot;</span><span class="p">)</span>
<span class="n">hf_llama</span> <span class="o">=</span> <span class="n">hf_model</span><span class="o">.</span><span class="n">from_pretrained</span><span class="p">(</span>
<span class="n">hf_model_dir</span><span class="p">,</span>
<span class="n">device_map</span><span class="o">=</span><span class="p">{</span>
<span class="s2">&quot;model&quot;</span><span class="p">:</span> <span class="s2">&quot;cpu&quot;</span><span class="p">,</span>
<span class="s2">&quot;lm_head&quot;</span><span class="p">:</span> <span class="s2">&quot;cpu&quot;</span><span class="p">,</span>
<span class="s2">&quot;embed_tokens&quot;</span><span class="p">:</span> <span class="s2">&quot;cpu&quot;</span><span class="p">,</span>
<span class="s2">&quot;layers&quot;</span><span class="p">:</span> <span class="s2">&quot;cpu&quot;</span><span class="p">,</span>
<span class="s2">&quot;norm&quot;</span><span class="p">:</span> <span class="s2">&quot;cpu&quot;</span><span class="p">,</span>
<span class="p">},</span> <span class="c1"># Load to CPU memory</span>
<span class="n">torch_dtype</span><span class="o">=</span><span class="s1">&#39;auto&#39;</span><span class="p">,</span>
<span class="p">)</span>
<span class="n">weights</span> <span class="o">=</span> <span class="n">load_from_hf_llama</span><span class="p">(</span>
<span class="n">tllm_llama</span><span class="p">,</span>
<span class="n">hf_llama</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">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">,</span>
<span class="c1"># TODO: these shall be outside from_hugging_face too.</span>
<span class="n">use_gemm_woq_plugin</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="s2">&quot;use_gemm_woq_plugin&quot;</span><span class="p">,</span> <span class="kc">False</span><span class="p">),</span>
<span class="n">lora_config</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="s2">&quot;lora_config&quot;</span><span class="p">,</span> <span class="n">LoraConfig</span><span class="p">()),</span>
<span class="p">)</span>
<span class="n">profiler</span><span class="o">.</span><span class="n">stop</span><span class="p">(</span><span class="s2">&quot;Loading weights from HF&quot;</span><span class="p">)</span>
<span class="k">del</span> <span class="n">hf_llama</span>
<span class="n">weights</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">q_weights</span><span class="p">)</span>
<span class="n">tllm_llama</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="k">else</span><span class="p">:</span>
<span class="n">tllm_llama</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="n">q_weights</span><span class="p">)</span>
<span class="k">return</span> <span class="n">tllm_llama</span></div>
<span class="k">def</span> <span class="nf">_quantize</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">hf_model_dir</span><span class="p">,</span> <span class="n">dtype</span><span class="p">,</span> <span class="n">cfg</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
<span class="w"> </span><span class="sd">&#39;&#39;&#39;Given the quant_mode set in the Module object, read from given hf model</span>
<span class="sd"> call AMMO to generate quantization scales, and set the scales back the module parameters.</span>
<span class="sd"> &#39;&#39;&#39;</span>
<span class="c1"># use self destructed temporary path if kwargs[quantization_cache_dir] is not specified</span>
<span class="c1"># sometimes the quantization checkpoint path needs to be saved for debug purpose</span>
<span class="n">quantized_temp_dir</span> <span class="o">=</span> <span class="n">tempfile</span><span class="o">.</span><span class="n">TemporaryDirectory</span><span class="p">(</span><span class="s2">&quot;llama-quantized&quot;</span><span class="p">)</span>
<span class="n">quantized_checkpoint_path</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="s2">&quot;quantization_cache_dir&quot;</span><span class="p">,</span>
<span class="n">quantized_temp_dir</span><span class="o">.</span><span class="n">name</span><span class="p">)</span>
<span class="n">quantize_lm_head</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="s2">&quot;quantize_lm_head&quot;</span><span class="p">,</span> <span class="kc">False</span><span class="p">)</span>
<span class="n">quant_mode</span> <span class="o">=</span> <span class="n">cfg</span><span class="o">.</span><span class="n">quant_mode</span>
<span class="n">ammo_qformat</span> <span class="o">=</span> <span class="kc">None</span>
<span class="n">calib_size</span> <span class="o">=</span> <span class="kc">None</span>
<span class="k">if</span> <span class="n">quant_mode</span><span class="o">.</span><span class="n">has_fp8_qdq</span><span class="p">()</span> <span class="ow">or</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="n">ammo_qformat</span> <span class="o">=</span> <span class="s1">&#39;fp8&#39;</span>
<span class="n">calib_size</span> <span class="o">=</span> <span class="mi">512</span>
<span class="c1"># TODO: how to distinguish from quant_mode about int4_awq or int4_gptq?</span>
<span class="k">elif</span> <span class="n">quant_mode</span><span class="o">.</span><span class="n">is_int4_weight_only_per_group</span><span class="p">():</span>
<span class="n">ammo_qformat</span> <span class="o">=</span> <span class="s1">&#39;int4_awq&#39;</span>
<span class="n">calib_size</span> <span class="o">=</span> <span class="mi">32</span>
<span class="k">assert</span> <span class="n">ammo_qformat</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
<span class="c1"># local import to avoid pytest issue when importing AMMO and transformers lib</span>
<span class="kn">from</span> <span class="nn">.quantize</span> <span class="kn">import</span> <span class="n">quantize_llama_and_export</span>
<span class="n">quantize_llama_and_export</span><span class="p">(</span><span class="n">hf_model_dir</span><span class="p">,</span>
<span class="n">quantized_checkpoint_path</span><span class="p">,</span>
<span class="n">ammo_qformat</span><span class="p">,</span>
<span class="n">dtype</span><span class="p">,</span>
<span class="n">calib_size</span><span class="o">=</span><span class="n">calib_size</span><span class="p">,</span>
<span class="n">quantize_lm_head</span><span class="o">=</span><span class="n">quantize_lm_head</span><span class="p">)</span>
<span class="n">ckpt</span> <span class="o">=</span> <span class="n">Path</span><span class="p">(</span><span class="n">quantized_checkpoint_path</span><span class="p">)</span> <span class="o">/</span> <span class="s2">&quot;llama_tp1_rank0.npz&quot;</span>
<span class="k">assert</span> <span class="n">ckpt</span><span class="o">.</span><span class="n">exists</span><span class="p">(),</span> <span class="sa">f</span><span class="s2">&quot;The expecting checkpoint path </span><span class="si">{</span><span class="n">ckpt</span><span class="si">}</span><span class="s2"> does not exist&quot;</span> \
<span class="s2">&quot;it&#39;s likely quantization failed, pls check error logs&quot;</span>
<span class="n">hf_config</span> <span class="o">=</span> <span class="n">AutoConfig</span><span class="o">.</span><span class="n">from_pretrained</span><span class="p">(</span><span class="n">hf_model_dir</span><span class="p">,</span>
<span class="n">trust_remote_code</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="k">if</span> <span class="n">ammo_qformat</span> <span class="o">==</span> <span class="s1">&#39;fp8&#39;</span><span class="p">:</span>
<span class="k">return</span> <span class="n">load_from_fp8_llama</span><span class="p">(</span>
<span class="nb">str</span><span class="p">(</span><span class="n">ckpt</span><span class="p">),</span>
<span class="n">hf_config</span><span class="o">.</span><span class="n">num_hidden_layers</span><span class="p">,</span>
<span class="n">cfg</span><span class="o">.</span><span class="n">mapping</span><span class="p">,</span>
<span class="n">fp8_kv_cache</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">else</span><span class="p">:</span>
<span class="k">return</span> <span class="n">load_from_awq_llama</span><span class="p">(</span><span class="nb">str</span><span class="p">(</span><span class="n">ckpt</span><span class="p">),</span>
<span class="n">hf_config</span><span class="o">.</span><span class="n">num_hidden_layers</span><span class="p">,</span>
<span class="n">hf_config</span><span class="o">.</span><span class="n">vocab_size</span><span class="p">,</span>
<span class="n">cfg</span><span class="o">.</span><span class="n">mapping</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="c1"># llama specific setters, user shall has the chance to change the module attributes after</span>
<span class="c1"># from_hugging_face factory method created the model when these attributes is not included in the huggingface checkpoint</span>
<div class="viewcode-block" id="LLaMAForCausalLM.rotary_base">
<a class="viewcode-back" href="../../../../python-api/tensorrt_llm.models.html#tensorrt_llm.models.LLaMAForCausalLM.rotary_base">[docs]</a>
<span class="k">def</span> <span class="nf">rotary_base</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">val</span><span class="p">):</span>
<span class="k">for</span> <span class="n">decoder</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">layers</span><span class="p">:</span>
<span class="n">decoder</span><span class="o">.</span><span class="n">attention</span><span class="o">.</span><span class="n">rotary_embedding_base</span> <span class="o">=</span> <span class="n">val</span>
<span class="k">return</span> <span class="bp">self</span></div>
<div class="viewcode-block" id="LLaMAForCausalLM.rotary_scaling">
<a class="viewcode-back" href="../../../../python-api/tensorrt_llm.models.html#tensorrt_llm.models.LLaMAForCausalLM.rotary_scaling">[docs]</a>
<span class="k">def</span> <span class="nf">rotary_scaling</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">scaling_type</span><span class="p">,</span> <span class="n">factor</span><span class="p">):</span>
<span class="c1"># TODO: what if there are some other behaviors triggered by the these changes?</span>
<span class="c1"># should implement these assignment as setters of the Attention Module</span>
<span class="k">assert</span> <span class="n">scaling_type</span> <span class="ow">in</span> <span class="p">(</span><span class="s2">&quot;linear&quot;</span><span class="p">,</span> <span class="s2">&quot;dynamic&quot;</span><span class="p">),</span> <span class="sa">f</span><span class="s2">&quot;Got </span><span class="si">{</span><span class="n">scaling_type</span><span class="si">}</span><span class="s2">&quot;</span>
<span class="k">assert</span> <span class="n">factor</span> <span class="o">&gt;</span> <span class="mf">1.0</span><span class="p">,</span> <span class="sa">f</span><span class="s2">&quot;Got </span><span class="si">{</span><span class="n">factor</span><span class="si">}</span><span class="s2">&quot;</span>
<span class="k">for</span> <span class="n">decoder</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">layers</span><span class="p">:</span>
<span class="n">decoder</span><span class="o">.</span><span class="n">attention</span><span class="o">.</span><span class="n">rotary_embedding_scale_type</span> <span class="o">=</span> <span class="n">RotaryScalingType</span><span class="o">.</span><span class="n">linear</span> <span class="k">if</span> <span class="n">scaling_type</span> <span class="o">==</span> <span class="s2">&quot;linear&quot;</span> <span class="k">else</span> <span class="n">RotaryScalingType</span><span class="o">.</span><span class="n">dynamic</span>
<span class="n">decoder</span><span class="o">.</span><span class="n">attention</span><span class="o">.</span><span class="n">rotary_embedding_scale</span> <span class="o">=</span> <span class="n">factor</span>
<span class="k">return</span> <span class="bp">self</span></div>
<div class="viewcode-block" id="LLaMAForCausalLM.default_plugin_config">
<a class="viewcode-back" href="../../../../python-api/tensorrt_llm.models.html#tensorrt_llm.models.LLaMAForCausalLM.default_plugin_config">[docs]</a>
<span class="k">def</span> <span class="nf">default_plugin_config</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
<span class="n">plugin_config</span> <span class="o">=</span> <span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="n">default_plugin_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="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="o">.</span><span class="n">is_int4_weight_only_per_group</span><span class="p">():</span>
<span class="n">plugin_config</span><span class="o">.</span><span class="n">set_weight_only_groupwise_quant_matmul_plugin</span><span class="p">()</span>
<span class="k">return</span> <span class="n">plugin_config</span></div>
</div>
</pre></div>
</div>
</div>
<footer>
<hr/>
<div role="contentinfo">
<p>&#169; Copyright 2023, NVidia.</p>
</div>
Built with <a href="https://www.sphinx-doc.org/">Sphinx</a> using a
<a href="https://github.com/readthedocs/sphinx_rtd_theme">theme</a>
provided by <a href="https://readthedocs.org">Read the Docs</a>.
</footer>
</div>
</div>
</section>
</div>
<script>
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
</body>
</html>
Reference in New Issue View Git Blame Copy Permalink