[IlyasMoutawwakil]

What.

• Mxfp4Config and Mxfp4HfQuantizer become a block-scaled format: W4A16 (the default, as GPT-OSS runs), W4A8 or
  W4A4. This moves MXFP4 onto the shared kernels and loaders that FP8 uses since FP8 on the fine-grained machinery (Fp8Config) #49777.
• The quantizer reads GPT-OSS's _blocks/_scales layout (openai/gpt-oss-20b). Mxfp4Config keeps
  main's positional arguments. The frozen integrations/mxfp4 stays for main's behaviours.
• read_fp8 builds DeepSeek-V4's plan when the FP8 checkpoint says expert_dtype="fp4": block FP8 over W4A4 MXFP4
  experts. V4.1 ships expert_dtype in its quantization config. V4 sets it on the model config, so
  DeepseekV4Config.__post_init__ moves it into the quantization config, with the "deepseek" layout. The Hub
  config (deepseek-ai/DeepSeek-V4-Flash) only says fp8, so only the model knows its scales are named scale.
• Its conversion mapping names the checkpoint's own keys, so a save writes them back.

Behaviour changes.

• DeepSeek-V4's FP4 experts on DeepGEMM need experts_implementation="deepgemm" at load. Switching to it after a
  Triton load raises, so the SM100 test sweeps leave it out. deepgemm_megamoe works as before.
• Every MXFP4 format needs compute capability 8.9, because the kernels use Triton's FP8 type. On a CUDA GPU below
  8.9, such as an A100 or a T4, Mxfp4Config loads through main's MXFP4 quantizer, kept frozen as
  Mxfp4LegacyHfQuantizer.
• A missing kernels package still dequantizes a checkpoint.
• A serialized MXFP4 quantization_config holds more fields.
• DeepSeek-V4 loads on the fine-grained path instead of the legacy one.

Breaking. 🚨 GPT-OSS MXFP4 with use_kernels=True now loads dequantized on CPU too.

• use_kernels puts the megablocks layer on GptOssMLP, and that layer reads bf16 experts. So an MXFP4 model with
  use_kernels=True dequantizes its experts, on every device. Main did this on GPU only.
• On CPU, main kept main's MXFP4 experts, which the megablocks CPU kernel reads. The fine-grained experts are not
  that class, so the CPU case follows the same rule as the GPU one.

Depends on. #49777. It needs FP8 for DeepSeek-V4.

How tested. tests/quantization/mxfp4, the DeepSeek-V4 plan tests in tests/quantization/finegrained, GPT-OSS
and DeepSeek-V4 model tests.

The comments and docstrings that say why each change exists are there for readability. Feel free to ask for any of
them to go if it's self-evident.

🤖 Generated with Claude Code

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: deepseek_v4, finegrained, fp8, mxfp4