config file based modelopt config 1/N

[shengliangxu]

What does this PR do?

Type of change: ?

new feature Jira: [OMNIML-3212]

Overview: ?

1. start a new config system using yaml/yml files. The config system adopts Hydra style override, using the defaults tag, but we implement it by ourselves by simply merging the configs using OmegaConf

2. implement the quantization configs using the new config system, but not actually in use.

3. make sure the configs from the new config system match the exisiting configs

4. pipe through config_file based hf_ptq script

5. testted hf_ptq using both builtin and extenal config file

The config files are not actually in use in this PR, except the hf_ptq.py script accepts a quantization config name corresponding to a config filename

Testing

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

python examples/llm_ptq/hf_ptq.py \
    --pyt_ckpt_path=Qwen/Qwen3-8B \
    --export_path=nvfp4_awq \
    --qformat=nvfp4_awq_full \
    --kv_cache_qformat=fp8 \
    --calib_size=16 \
    --batch_size=0 \
    --trust_remote_code \
    --export_fmt=hf

[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

❌ Patch coverage is 76.47059% with 16 lines in your changes missing coverage. Please review.
✅ Project coverage is 74.52%. Comparing base (9409412) to head (c5f657e).
⚠️ Report is 6 commits behind head on main.

Files with missing lines	Patch %	Lines
modelopt/torch/opt/config.py	31.57%	13 Missing ⚠️
modelopt/torch/quantization/config.py	93.87%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #657      +/-   ##
==========================================
- Coverage   74.66%   74.52%   -0.15%     
==========================================
  Files         183      183              
  Lines       18550    18467      -83     
==========================================
- Hits        13851    13763      -88     
- Misses       4699     4704       +5     

☔ 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.

[kevalmorabia97]

Why do we keep this in modelopt/config/quantization/ instead of modelopt/torch/quantization/config?

[realAsma]

+1 on this. Since we already have modelopt/torch/quantization/config.py, can we move them to modelopt/torch/quantization/config_yamls/ or something?

[shengliangxu]

Because I want to make the builtin configs first class citizens. The configs are data, I want to put all built-in configs in one centralized place rather than scatter them around in code folders.

And I plan to convert all modelopt configs to be config file based.

The quantization configs are now under modelopt/config/quantization/, other configs will be placed in correponding sub-folder of modelopt/config.

The actual sub-dirs of each config type may change, but overall the idea of put all of them in a central place will be kept.

[realAsma]

Overall design looks great to me.

Can we make the yaml file flexible to add more methods such as auto_quantize or sparsity? the sparsity + quantization will be very important in future Cc @kaix-nv

Should we have one more level to the yaml file based on the high level API:

quantize: # Arguments to mtq.quantize API
     quant_cfg:
          ....

     algorithm:
        ....

sparsify: # Arguments to mts.sparsify
    config:
        ....

[shengliangxu]

Overall design looks great to me.

Can we make the yaml file flexible to add more methods such as auto_quantize or sparsity? the sparsity + quantization will be very important in future Cc @kaix-nv

Should we have one more level to the yaml file based on the high level API:

quantize: # Arguments to mtq.quantize API
     quant_cfg:
          ....

     algorithm:
        ....

sparsify: # Arguments to mts.sparsify
    config:
        ....

This is just 1/N of the changes. The use of the new config system is just limited to one use case just for illustration. More broader adaption will be in follow-up changes.

[shengliangxu]

The more I think about it, the more I feel it may be worth a design review of itself. I'll put together a more thorough design doc and share it.