[Quantization] Add field to skip unquantized modules for GPTQ config#25455
Merged
Isotr0py merged 9 commits intovllm-project:mainfrom Sep 26, 2025
Merged
[Quantization] Add field to skip unquantized modules for GPTQ config#25455Isotr0py merged 9 commits intovllm-project:mainfrom
Isotr0py merged 9 commits intovllm-project:mainfrom
Conversation
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py
requested review from
ProExpertProg,
WoosukKwon,
hmellor,
houseroad,
mgoin,
robertgshaw2-redhat,
sighingnow,
simon-mo,
tlrmchlsmth,
yewentao256 and
youkaichao
as code owners
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request introduces a more robust and cleaner way to handle unquantized modules in GPTQ models by adding the modules_in_block_to_quantize field to the configuration. This replaces the previous hacky approach and resolves existing issues, which is a great improvement for maintainability. The implementation correctly infers the modules to quantize from the model's checkpoint metadata when not explicitly provided. I have one suggestion regarding code duplication between GPTQConfig and GPTQMarlinConfig that would further enhance the maintainability of this new logic.
Comment on lines
+241
to
+261
| def apply_vllm_mapper(self, hf_to_vllm_mapper): | ||
| if self.modules_in_block_to_quantize is not None: | ||
| self.modules_in_block_to_quantize = hf_to_vllm_mapper.apply_list( | ||
| self.modules_in_block_to_quantize) | ||
|
|
||
| def maybe_update_config(self, | ||
| model_name: str, | ||
| revision: Optional[str] = None): | ||
| if self.modules_in_block_to_quantize: | ||
| return | ||
|
|
||
| unquant_dtypes = [torch.float16, torch.bfloat16, torch.float32] | ||
| metadata = get_safetensors_params_metadata(model_name, | ||
| revision=revision) | ||
| quant_layers: set[str] = { | ||
| param_name.rsplit(".", 1)[0] | ||
| for param_name, info in metadata.items() | ||
| if (dtype := info.get('dtype', None)) | ||
| and _SAFETENSORS_TO_TORCH_DTYPE[dtype] not in unquant_dtypes | ||
| } | ||
| self.modules_in_block_to_quantize = list(quant_layers) |
Contributor
There was a problem hiding this comment.
The methods apply_vllm_mapper and maybe_update_config are identical to the ones in vllm/model_executor/layers/quantization/gptq.py. This code duplication can lead to maintenance issues, where a change in one file might not be propagated to the other, potentially causing bugs.
To improve maintainability, I recommend refactoring this shared logic into a common base class or a mixin. For example, you could create a GPTQConfigMixin class to house these methods, and then have both GPTQConfig and GPTQMarlinConfig inherit from it. This would ensure that the logic for handling modules_in_block_to_quantize is defined in a single place.
Collaborator
|
Can you check and clean up other models that contain |
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Member
Author
Removed |
jeejeelee
approved these changes
Sep 24, 2025
Collaborator
jeejeelee
left a comment
jeejeelee
left a comment
There was a problem hiding this comment.
Thank you for the improvement. Perhaps we can apply this logic to other quantitative methods
github-actions
bot
added
the
ready
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Contributor
|
Hi @Isotr0py , the following PR breaks model loading: jart25/Qwen3-Next-80B-A3B-Instruct-Int4-GPTQ The model has been generated using Autoround-GPTQ, like the previous ones |
Member
Author
|
Ooops, sorry for breaking this. Let me fix it. 😅 |
Contributor
|
@Isotr0py Don't worry, development |
Member
Author
|
@JartX I double checked the "modules_in_block_to_quantize": [
[
+++ "self_attn.q_proj",
+++ "self_attn.k_proj",
+++ "self_attn.v_proj",
+++ "self_attn.o_proj",
"linear_attn.in_proj_qkvz",
"linear_attn.in_proj_ba",
"linear_attn.out_proj",With above changes in |
Contributor
|
@Isotr0py You're right, it works, and Autoround doesn't declare them. Does that mean it used to work “by chance”? I'm going to open an issue in Autoround. |
Member
Author
That's right. The |
2 tasks
pdasigi
pushed a commit
to pdasigi/vllm
that referenced
this pull request
Oct 2, 2025
…llm-project#25455) Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
yewentao256
pushed a commit
that referenced
this pull request
Oct 3, 2025
…25455) Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn> Signed-off-by: yewentao256 <zhyanwentao@126.com>
choprahetarth
pushed a commit
to Tandemn-Labs/vllm
that referenced
this pull request
Oct 11, 2025
…llm-project#25455) Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
5 tasks
lywa1998
pushed a commit
to lywa1998/vllm
that referenced
this pull request
Oct 20, 2025
…llm-project#25455) Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
alhridoy
pushed a commit
to alhridoy/vllm
that referenced
this pull request
Oct 24, 2025
…llm-project#25455) Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
rtourgeman
pushed a commit
to rtourgeman/vllm
that referenced
this pull request
Nov 10, 2025
…llm-project#25455) Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
5 tasks
5 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Purpose
Qwen/Qwen2-VL-2B-Instruct-GPTQ-Int4andQwen/Qwen3-30B-A3B-GPTQ-Int4have unquantized layers, but there is no conifg to skip creating unquantized weights. And we're using_maybe_ignore_quant_configto patch model implementation in a hacky way.modules_in_block_to_quantizefrom TransformersGPTQConfig(https://huggingface.co/docs/transformers/v4.56.2/en/main_classes/quantization#transformers.GPTQConfig), and automatically infer it from checkpoint metadata. So that we no longer need to do hacky patching for GPTQ models now.Test Plan
Test Result
Qwen3-MoE models with/without MoE gate quantized can all generate normal outputs without GPTQ patch now.
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.