Support loading for static quant weight fp8 act fp8

[yiliu30]

• Add a new backend auto_round:torch_fp8_static for loading and inference w8afp8
• Enhance the QuantizationScheme to support dict-style access.
• Wrap layer_config dict as QuantizationScheme, and propagate it to the backend check

[Copilot]

Pull Request Overview

This PR adds support for loading static quantized models with FP8 weights and FP8 activations by implementing a new quantized linear layer class and updating the model conversion infrastructure.

Key changes:

• Implemented WeightFP8ActFP8StaticQuantLinear class for handling FP8 weight and activation quantization
• Updated model conversion logic to detect and handle FP8 static quantization configurations
• Enhanced test coverage to verify both export and loading functionality for static FP8 quantization

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
test/test_cpu/test_export.py	Extended test to verify loading of static FP8 quantized models and renamed test method
auto_round/inference/convert_model.py	Added support for act_dynamic parameter and FP8 static quantization detection in model conversion
auto_round/inference/backend.py	Added FP8 static quantization detection function and updated dynamic import logic
auto_round/export/export_to_autoround/export_to_fp8_woq.py	Implemented new WeightFP8ActFP8StaticQuantLinear class with quantization/dequantization methods

[wenhuach21]

This PR is unnecessary for now, you need to work with Heng to fix the FP8 --format fake bug and the evaluation should be fine

[yiliu30]

This PR is unnecessary for now, you need to work with Heng to fix the FP8 --format fake bug and the evaluation should be fine

@wenhuach21 The purpose of this PR is to support loading an existing qmodel from disk and then evaluating its accuracy.
To my understanding, this was not supported by --format fake.

cc @n1ck-guo

[wenhuach21]

This PR is unnecessary for now, you need to work with Heng to fix the FP8 --format fake bug and the evaluation should be fine

@wenhuach21 The purpose of this PR is to support loading an existing qmodel from disk and then evaluating its accuracy. To my understanding, this was not supported by --format fake.

cc @n1ck-guo

Yes, but the primary purpose is for evaluation, which the fake model should cover well #731. This is not a product feature, and it involves changes to critical product code. As discussed earlier, please hold this PR for now, or move the code elsewhere without modifying the important HF model inference code.

[yiliu30]

@wenhuach21 Please help review it again, thanks.

[yiliu30]

Hi @n1ck-guo @WeiweiZhang1, I wrapped the layer_config dict as a QuantizationScheme, and propagated it to the backend check. Please take a look and let me know if I missed any corner cases. Thanks.

[wenhuach21]

Since this PR modifies the core Transformers inference code, I’d prefer to merge it after the 0.6.1 release.

[yiliu30]

This feature is targeted for the INC 3.6, and several loading-related features depend on it, such as NVFP4/MXFP8/MXFP4, and possibly non-transformers model loading @mengniwang95 .

What’s the release plan for version 0.6.1? Could we create the 0.6.1 RC branch first and merge this PR into main as soon as possible?

cc @thuang6 @chensuyue

[wenhuach21]

will start pre-release test after merging pr 781