[IlyasMoutawwakil]

What.

• Mxfp8Config and Mxfp8HfQuantizer add FP8 weights with a UE8M0 scale per 32, as MiniMax-M3 ships them. They run
  W8A8 or weight-only. tiny-random-MiniMaxM3-MXFP8 is the first user.
• quant_method="mxfp8" routes here instead of the deprecated FineGrainedFP8Config path, which stays frozen.
• The FineGrainedFP8Config deprecation points at Mxfp8Config for a [1, 32] UE8M0 block.
• read_mxfp8 reads MiniMax's ignored_layers skip-list as modules_to_not_convert, e.g.
  MiniMaxAI/MiniMax-M3-MXFP8. The deprecated config keeps its own alias.
• MiniMax-M3-VL docs point at the public checkpoints. Its parallel fixture comes from its model tester and pins the
  routing, so a near-tie in the router does not flip under sharding.
• The shared tests extend their format lists with MXFP8.
• The mixed-plan grid (MIXED_PAIRS) adds MXFP8 pairs. One runs MXFP4 experts with MXFP8 activations (W4A8) beside
  weight-only MXFP8 dense layers. MXFP8 weights with MXFP4 activations (W8A4) are not supported.
• The FINEGRAINED test spec runs the tiny-checkpoint and load-path tests on a mixed plan. Its model is DeepSeek-V4,
  with block-FP8 layers and MXFP4 experts, and its fixtures exist since 🚨 MXFP4 on the fine-grained machinery, and DeepSeek-V4's FP8 + MXFP4 plan #49778. It sits here with the other spec
  changes to the shared tester.

Depends on. #49779.

How tested. tests/quantization/mxfp8 and the shared fine-grained 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: finegrained, finegrained_fp8, mxfp8