Update model_loader deps and qqq quantization deps

[HandH1998]

Motivation

Update the model_loader deps and qqq quantization deps for SGLang.

Modifications

We modified the relevant code primarily according to vLLM. Thanks the vLLM team for their significant contributions. Here we list the main modifications.

• We adapted the model_loader code from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/model_loader and modified it adaptively for SGLang. The updated model_loader code is located at python/sglang/srt/model_loader.
• We added registry.py at python/sglang/srt/models/registry.py and registered all the models into class ModelRegistry. Consequently, we removed all monkey patches in python/sglang/srt/model_executor/model_runner.py.
• We have added load_config.py and device_config.py to python/sglang/srt/configs. Additionally, we removed LoraConfig, CacheConfig, ParallelConfig, and SchedulerConfig as they are set to None and are not being utilized.

[HandH1998]

There are some failures due to cannot import name 'marlin_qqq_gemm' from 'torchao.ops' (/usr/local/lib/python3.10/dist-packages/torchao/ops.py) in the CR. This issue arises because the installed version of torchao is v0.6.1, which does not support marlin_qqq_gemm. Although our marlin_qqq_gemm has been merged into the main branch of torchao, the torchao team has not yet released a new version which supports marlin_qqq_gemm.

[zhyncs]

Overall LGTM left some comments
Except for rope, vllm.distributed and quant, everything else related to vllm needs to be removed, such as some utils
BTW python/sglang/srt/models/phi3_small.py should also be handled

[HandH1998]

I have updated the code according to your review.

[zhyncs]

@HandH1998 May you change the permissions of this PR to allow maintainers to update your branch's code? This way, we can also help fix it and speed up the merging process.

[HandH1998]

@zhyncs ok, I have added the permission.

[HandH1998]

I think I have fixed all the issues in the CR. Please review the latest code.

[zhyncs]

I'll merge this PR into sgl-project:HandH1998/sgl_model_loader first for the nightly gsm8k evaluation. Then, I'll grant you permission to update the PR. @HandH1998 cc @merrymercy @Ying1123

[zhyncs]

ref https://github.com/sgl-project/sglang/actions/runs/12121646155