[Revert] MXFP4 + block-FP8 Humming support after H100 correctness failure

[khluu]

Why this revert

Reverts #51332 / ca29cd48ba05 after the first exact post-merge H100 evaluation of the newly added MXFP4 + block-FP8 path produced invalid outputs:

• failing H100 job: https://buildkite.com/vllm/ci/builds/85561#01a03aba-6c3e-4b98-a0f5-40d420e14882
• exact test: test_gsm8k_correctness[Qwen3-30B-A3B-MXFP4A16-humming-act-fp8-block]
• result: GSM8K 0.0000, invalid rate 0.994, versus required floor 0.7800; job summary 1 failed / 9 passed
• same-build B200 control: https://buildkite.com/vllm/ci/builds/85561#01a03aba-6c3b-42ae-a34a-2756f41934bc passed the exact new config at 0.8863, invalid rate 0.000

This isolates an SM90/H100 correctness regression in the newly merged path. The server remained alive and evaluated all 1,319 prompts, so this is not an agent, startup, or model-download failure.

Coverage gap

The exact H100 Humming Activation job did not run before merge:

• PR [Quantization][Humming] Support MXFP4 weight + block-FP8 activation for MoE #51332 exact-head #85460: blocked / not selected — no test ran
• first post-merge main #85493: blocked / not selected — no test ran
• current main #85552: blocked / not selected — no test ran
• latest-merge PR [Core] Disable fuse_allreduce_rms under VLLM_BATCH_INVARIANT (non-deterministic under TP) #51292 exact #85498: blocked / not selected — no test ran

The daily #85561 run is the first exact H100 execution of this new config.

Validation

Local:

• git diff --cached --check before commit
• .venv/bin/python -m compileall -q vllm/model_executor/layers/fused_moe vllm/model_executor/layers/quantization/utils/humming_utils.py
• uvx pre-commit run — all applicable hooks passed

Exact hardware validation at head d2ba916fe024774325baaab418d0af0352b53192:

• H100 Humming Activation: https://buildkite.com/vllm/ci/builds/85565#01a03b02-0ca1-477d-ad90-2e779061aa0b — 9 passed / 16 warnings in 31m37s
• B200 Humming Activation control: https://buildkite.com/vllm/ci/builds/85565#01a03b02-0c9f-4cfc-b122-fd1efc00e306 — FP8 9 passed / 16 warnings and INT8 6 passed / 16 warnings

The proposed revert is exact-green on both required architectures. Keep this PR draft until human review; CI-incident closure still requires merge plus an exact post-merge main H100 pass.

Duplicate-work check

No open PR references #51332 as a revert/fix, and no open PR or issue matched the exact MXFP4 + block-FP8 H100 failure. This revert is intentionally conservative until the feature can restore H100 correctness with exact pre-merge coverage.

AI assistance was used for CI triage, preparing the revert, and drafting this description. Kevin Luu reviewed the exact diff and validation evidence before submission.

[khluu]

Exact hardware validation launched: Buildkite #85565, with only lm-eval-humming-act-h100 and lm-eval-humming-act-b200 selected. H100 is the decisive failing-family gate; B200 is the same-config control. Keep this draft until both exact jobs pass.

[khluu]

Exact hardware validation is terminal green at d2ba916fe024774325baaab418d0af0352b53192:

• H100 decisive gate: https://buildkite.com/vllm/ci/builds/85565#01a03b02-0ca1-477d-ad90-2e779061aa0b — 9 passed / 16 warnings in 31m37s
• B200 control: https://buildkite.com/vllm/ci/builds/85565#01a03b02-0c9f-4cfc-b122-fd1efc00e306 — FP8 9 passed / 16 warnings, INT8 6 passed / 16 warnings

The original H100-only failure remains https://buildkite.com/vllm/ci/builds/85561#01a03aba-6c3e-4b98-a0f5-40d420e14882: exact test_gsm8k_correctness[Qwen3-30B-A3B-MXFP4A16-humming-act-fp8-block], 0.0000 accuracy / 99.4% invalid outputs against the 0.7800 floor.

This validates the proposed revert; keeping the PR draft for human review. The CI incident remains open until merge plus an exact post-merge main H100 pass.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @khluu.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork