[IlyasMoutawwakil]

What.

• NVFP4Config and NVFP4HfQuantizer become a block-scaled format. Weights are E2M1 with an E4M3 scale per 16 and an
  fp32 global. This moves NVFP4 onto the shared kernels and loaders that FP8 and MXFP4 already use.
• NVFP4 checkpoints, transformers saves included, use the ModelOpt key layout. The quantizer reads it per expert
  (GLM-5.2-NVFP4) and stacked per layer (vLLM).
• The gate and up globals fold through FineGrainedWeightGlobals. A gate|up stack keeps one global, so the up ratio
  moves into the down global.
• A weight-only checkpoint refuses quantized activations. It has no input_scale to run them with. The check is
  skipped when the load dequantizes, since a dequantized model runs no activation format.
• NVFP4Config keeps main's positional modules_to_not_convert, so existing callers keep working.
• The frozen integrations/nvfp4.py stays importable, with a FutureWarning. It goes in v5.25.
• MuseSpark's input-scaled RMS norm fuses into the kernels' MoE chain. It scales by (1 + weight) and rounds to
  bf16 before the norm, so it gets its own kernel name. tiny-random-Muse-Spark-NVFP4 is its first user.

Behaviour change. A pre-quantized NVFP4 checkpoint on CPU dequantizes instead of raising.

Depends on. #49778.

How tested. tests/quantization/nvfp4, the NVFP4 tests
in tests/kernels and tests/integrations/finegrained, mixed FP8/MXFP4/NVFP4 plans, Qwen4-exp.

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: muse_spark, qwen4_exp, finegrained, nvfp4