Refactor and clean up hf_ptq.py

[shengliangxu]

What does this PR do?

Refactor and clean up hf_ptq.py

This script has several separate logic and the code of them are entangled, making it really hard to add new features

Refactor them so that we separate these logics:

1. sparsity, all logic go to sparsity_main. TODO: we may actually move this logic out to a separate script

2. quantize, all logic go to quantize_main.

   2.1 plain quantization with a single quantization format

   2.2 auto quantization

In the quantization pipeline, separate the pipeline to:

1. model loading
2. calibrate dataset loading
3. pre-quantize processing
4. actual quantize
5. post-quantize processing
6. quantized model export

Testing

tested the mono quantization:

python examples/llm_ptq/hf_ptq.py \
    --pyt_ckpt_path=Qwen/Qwen3-8B \
    --export_path=qwen3-8B_fp8 \
    --qformat=fp8 \
    --kv_cache_qformat=fp8 \
    --calib_size=16 \
    --batch_size=0 \
    --trust_remote_code \
    --export_fmt=hf

and deployed to vLLM and TRTLLM, validated accuracy using lm_eval

tested auto quantize:

python examples/llm_ptq/hf_ptq.py \
    --qformat=nvfp4,fp8 \
    --auto_quantize_score_size 128 \
    --auto_quantize_bits 5.0 \
    --auto_quantize_checkpoint llama-8B-auto-quantize-checkpoint \
    --pyt_ckpt_path=meta-llama/Meta-Llama-3-8B \
    --export_path=llama-8B_auto_quantize
    --kv_cache_qformat=fp8 \
    --calib_size=16 \
    --batch_size=0 \
    --trust_remote_code \
    --export_fmt=tensorrt_llm

and compared exported files

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 74.78%. Comparing base (cd0d185) to head (defa50a).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #665      +/-   ##
==========================================
- Coverage   74.80%   74.78%   -0.02%     
==========================================
  Files         192      192              
  Lines       18814    18814              
==========================================
- Hits        14073    14070       -3     
- Misses       4741     4744       +3     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[cjluo-nv]

@shengliangxu is this intended?

[shengliangxu]

yes, this None is for auto_quantize, the target function does not apply to auto_quantize at all, so removed this useless arg.

[cjluo-nv]

nit: maybe let's call default_quantize or single_precision_quantize?

[shengliangxu]

It's not necessary single precision, just a single config. We can have mixed precision even using single config.

default_quantize does not sounds a bit too general and a bit tedious. How about mono_quantize? meaning quantize using a single config

[shengliangxu]

auto_quantize vs mono_quantize, quite symmetric

[cjluo-nv]

Thanks @shengliangxu for the refactoring. Could you also validate that the checkpoint before and after this change is the same?

[shengliangxu]

Thanks @shengliangxu for the refactoring. Could you also validate that the checkpoint before and after this change is the same?

done, updated the description.