refine transform config api.

[lkk12014402]

#1577

Description

refine transform config api, which supports str|dict|TransformConfig|None

[lkk12014402]

TODO: add document

[Copilot]

Pull request overview

Refines the transform config API so callers can pass a str | dict | TransformConfig | None, and normalizes inputs before applying transforms.

Changes:

• Added a shared normalization helper to validate/standardize transform config inputs.
• Updated apply_transform to accept multiple config input types plus an optional quantization scheme.
• Updated compressor initialization to accept and normalize transform_config directly.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 6 comments.

File	Description
auto_round/experimental/transform/helper.py	Adds _normalize_transform_config for validating and converting transform_config inputs into a normalized dict.
auto_round/experimental/transform/apply.py	Expands apply_transform signature and normalizes user inputs before building TransformConfig.
auto_round/compressors/base.py	Adds transform_config parameter and normalizes it during compressor initialization.

[wenhuach21]

As also mentioned before, these two args in transform_config should not be exposed to users.

  location: str = Field(default="weight", exclude=True)

  # apply transform inside modules for nvfp4, autoround tuning etc.
  need_calibration: bool = Field(default=False, exclude=True)

[wenhuach21]

and this one

 # required, currently only supports mxfp4
    quant_scheme: str = Field(..., description="Quantization scheme. Currently supports 'MXFP4/MXFP8'.")

why users need to pass the scheme once again?

[lkk12014402]

and this one

 # required, currently only supports mxfp4
    quant_scheme: str = Field(..., description="Quantization scheme. Currently supports 'MXFP4/MXFP8'.")

why users need to pass the scheme once again?

need check the supported quantization scheme with transform

[lkk12014402]

As also mentioned before, these two args in transform_config should not be exposed to users.

  location: str = Field(default="weight", exclude=True)

  # apply transform inside modules for nvfp4, autoround tuning etc.
  need_calibration: bool = Field(default=False, exclude=True)

the arg location determines where the Hadamard transform is applied, like weight or activation, we need set the parameter when need transform activation

the arg need_calibration is for calibration (like nvfp4) and tuning (iters > 200)

[wenhuach21]

For API design, please take a position of uers, not developer

and this one

 # required, currently only supports mxfp4
    quant_scheme: str = Field(..., description="Quantization scheme. Currently supports 'MXFP4/MXFP8'.")

why users need to pass the scheme once again?

need check the supported quantization scheme with transform

I know, this is not the right logic. you should check it in another place

[wenhuach21]

@chensuyue @XuehaoSun To support Hadamard, a safetensor file has been added to the main branch. I assume we will not include this file in the release package by default. Please help include it if Kaokao thinks it is a better option than generating it at runtime.

[xin3he]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[chensuyue]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[chensuyue]

@chensuyue @XuehaoSun To support Hadamard, a safetensor file has been added to the main branch. I assume we will not include this file in the release package by default. Please help include it if Kaokao thinks it is a better option than generating it at runtime.

Added into release package.