Cosmos 3 FP8: wrapping pipe.transformer (e.g. torch.compile) silently disables the checkpoint's W8A16 step policy

オープン
#15,169 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
4/5
見積もり時間
3〜5日
初心者へのやさしさ
42/100
issue の種類
バグ
明瞭さ
おおむね明確
活発さ
活発
技術スタック
python, pytorch

調査の方向性

Start in src/diffusers/pipelines/cosmos/mixed_precision.py: the loader lookup getattr(type(module), "load_config", None) in _on_disk_quantization_config() is the cause, and resolve() falls back to format="none" when it fails. Try the suggested fix of looking up load_config on the instance, or unwrapping with unwrap_module in src/diffusers/utils/torch_utils.py. Done when Cosmos3MixedPrecisionConfig.resolve(torch.compile(model)) equals resolve(model) on a small model whose config declares the ModelOpt FP8 policy, with a regression test added.

索引モデルが issue の本文から書いたものです。

説明

Describe the bug

For the ModelOpt FP8 Cosmos 3 checkpoints, the per-step W8A16/W8A8 policy (quantization_config.runtime.diffusion_step_policy in transformer/config.json, added in #14664) is resolved from self.transformer on every pipeline call:

https://github.com/huggingface/diffusers/blob/086bf9578c0f4acbc66e48cd1e7cc26befd9e10f/src/diffusers/pipelines/cosmos/pipeline_cosmos3_omni.py#L1709-L1715

Cosmos3MixedPrecisionConfig.resolve() → quantization_config_from_module() → _on_disk_quantization_config(), which looks up the config loader on the module's class:

https://github.com/huggingface/diffusers/blob/086bf9578c0f4acbc66e48cd1e7cc26befd9e10f/src/diffusers/pipelines/cosmos/mixed_precision.py#L443-L464

loader = getattr(type(module), "load_config", None)

The in-memory config does not carry quantization_config.runtime (see the docstring of quantization_config_from_module, L409-L426), so this on-disk read is the only source of the policy. When pipe.transformer is a wrapper, type(module) is the wrapper class, which has no load_config. The on-disk config is never read, _parsed_checkpoint_policy() returns None, and resolve() returns format="none" (L157-L158). Every step then runs W8A8 instead of the checkpoint's W8A16 first/last steps, with no warning. The Cosmos 3 docs note that all-W8A8 multi-step video can flicker (cosmos3.md L50-L54).

pipe.transformer = torch.compile(pipe.transformer), the pattern shown in several docs pages (e.g. optimization/cache_dit.md L260, quantization/autoround.md L102), is such a wrapper (OptimizedModule). In-place pipe.transformer.compile() and compile_repeated_blocks() (used on the Cosmos 3 page) keep the class and are not affected. The line is unchanged on main: mixed_precision.py#L450.

Reproduction

The policy is printed with the same resolve() call the pipeline makes, before any generation:

import modelopt.torch.quantization.backends.fp8_per_tensor_gemm  # noqa: F401 - registers the FP8 GEMM
import torch
from diffusers import Cosmos3OmniPipeline
from diffusers.pipelines.cosmos.mixed_precision import Cosmos3MixedPrecisionConfig
from modelopt.torch.opt import enable_huggingface_checkpointing

enable_huggingface_checkpointing()
pipe = Cosmos3OmniPipeline.from_pretrained("nvidia/Cosmos3-Nano", revision="fp8", dtype=torch.bfloat16)
pipe.vae.to("cuda")
pipe.transformer.enable_group_offload(
    onload_device=torch.device("cuda"), offload_device=torch.device("cpu"),
    offload_type="block_level", num_blocks_per_group=1, use_stream=True,
)

pipe.transformer = torch.compile(pipe.transformer, backend="eager")  # the only difference between the two runs
print(f"type(pipe.transformer) = {type(pipe.transformer).__name__}")
print(f"resolved policy = {Cosmos3MixedPrecisionConfig.resolve(pipe.transformer)}")
Logs

Without the torch.compile line:

type(pipe.transformer) = Cosmos3OmniTransformer
resolved policy = Cosmos3MixedPrecisionConfig(format='fp8', first_steps=3, last_steps=3, reasoner_policy='high_precision', overlap='a16')

With it:

type(pipe.transformer) = OptimizedModule
resolved policy = Cosmos3MixedPrecisionConfig(format='none', first_steps=3, last_steps=3, reasoner_policy='high_precision', overlap='a16')

Effect on generation (T2V 848×480, 49 frames, 8 steps, seed 123, transformer with enable_group_offload(offload_type="block_level", num_blocks_per_group=1, use_stream=True) to fit a 16 GB GPU):

pipe.transformer call resolved policy mixed-precision runtime installed on the transformer
Cosmos3OmniTransformer default fp8, first/last 3 steps W8A16 yes (252 reasoner + 252 generation linears)
Cosmos3OmniTransformer mixed_precision_format="none" none no; different video from the default run
OptimizedModule (torch.compile(pipe.transformer, backend="eager")) default none — (see below)

With the compiled ModelOpt FP8 transformer, generation currently stops earlier inside Dynamo, so the end-to-end output of the compiled run cannot be shown; the resolved policy above is what the pipeline applies to it. Any wrapper that forwards attribute access and does run reaches the none schedule.

Dynamo traceback of the compiled run (separate issue, context only)
Traceback (most recent call last):
  File "repro.py", line 73, in <module>
    video = pipe(
            ^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/utils/_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/diffusers/pipelines/cosmos/pipeline_cosmos3_omni.py", line 1756, in __call__
    preds_vision, preds_sound, preds_action = self.transformer(
                                              ^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/_dynamo/eval_frame.py", line 414, in __call__
    return super().__call__(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1786, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/_dynamo/eval_frame.py", line 832, in compile_wrapper
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1775, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1786, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/diffusers/hooks/hooks.py", line 217, in new_forward
    output = function_reference.forward(*args, **kwargs)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/diffusers/hooks/hooks.py", line 216, in new_forward
    args, kwargs = function_reference.pre_forward(module, *args, **kwargs)
  [... torch/_dynamo frames ...]
  File "/usr/local/lib/python3.12/dist-packages/modelopt/torch/opt/dynamic.py", line 81, in __len__
    return len(self._callbacks)
               ^^^^^^^^^^^^^^^
torch._dynamo.exc.InternalTorchDynamoError: AttributeError: '_FoldedCallback' object has no attribute '_callbacks'

from user code:
   File "/usr/local/lib/python3.12/dist-packages/diffusers/hooks/hooks.py", line 217, in torch_dynamo_resume_in_new_forward_at_216
    output = function_reference.forward(*args, **kwargs)
  File "/usr/local/lib/python3.12/dist-packages/diffusers/models/transformers/transformer_cosmos3.py", line 745, in forward
    packed_tokens_vision = self.proj_in(packed_tokens_vision)
  File "/usr/local/lib/python3.12/dist-packages/modelopt/torch/quantization/nn/modules/quant_linear.py", line 240, in forward
    return super().forward(input, *args, **kwargs)
  File "/usr/local/lib/python3.12/dist-packages/modelopt/torch/quantization/nn/modules/quant_module.py", line 240, in forward
    return super().forward(input, *args, **kwargs)
  File "/usr/local/lib/python3.12/dist-packages/modelopt/torch/quantization/nn/modules/quant_module.py", line 192, in forward
    output = super().forward(input, *args, **kwargs)
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/linear.py", line 134, in forward
    return F.linear(input, self.weight, self.bias)
  File "/usr/local/lib/python3.12/dist-packages/modelopt/torch/opt/dynamic.py", line 782, in __getattr__
    return manager.get_da_cb(name)(self, value)
Suggested fix

Either of these, plus a regression test (resolve(torch.compile(model)) == resolve(model) on a small model whose config declares the ModelOpt FP8 policy):

  1. Look the loader up on the instance: loader = getattr(module, "load_config", None). load_config is a ConfigMixin classmethod, so this returns the same bound classmethod, and wrappers that forward attribute access (OptimizedModule.__getattr__ → _orig_mod) work too.
  2. Or unwrap first with the existing helper unwrap_module, as other Diffusers code does for compiled modules. This covers torch.compile only.

Optionally, a warning when a ModelOpt FP8 backend is detected but the on-disk policy could not be read, so the fallback to all-W8A8 is visible. Happy to open a PR.

System Info
  • diffusers 0.41.0, nvidia-modelopt 0.44.0, transformers 5.20.2, accelerate 1.15.0
  • torch 2.9.0a0+145a3a7 (container nvcr.io/nvidia/pytorch:25.10-py3), CUDA 13.0, Python 3.12
  • NVIDIA GeForce RTX 5080 16 GB, driver 620.62
主要言語
Python
スター
34.9k
フォーク
7.5k
平均マージ
3日 16時間
マージ済み PR(30日)
65

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

huggingface/diffusers のほかの issue

huggingface/diffusers の issue をすべて見る

似ている issue

Python の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。