[randomvariable]

Purpose

import vllm.model_executor.kernels.linear.scaled_mm raised:

ModuleNotFoundError: No module named
  'vllm.model_executor.kernels.linear.scaled_mm.b12x_block'

which aborted collection of every test touching the package — including tests/v1/attention/test_flashinfer_sparse_mla_sm120_api.py, whose failure had been misread as an environment limitation.

Root cause: upstream 4ab5e5012a (#52368) consolidated scaled_mm/b12x_block.py and scaled_mm/b12x_tensor.py into scaled_mm/b12x.py, and replaced the standalone warmup_b12x_*_linear functions with a get_b12x_warmup_unit provider protocol. The rebase took the new module and the file deletion but left the old import paths and several duplicated registrations behind.

Pulling that thread surfaced five further defects, all from the same half-applied refactor.

Defects fixed

1. Dead import path (the reported P0). scaled_mm/__init__.py imported from the deleted b12x_block. Pointed at b12x, matching linear/__init__.py, which was already correct.

2. Duplicate registrations silently reordering kernel priority. Three lists carried the same kernel twice, because the fork registered it before upstream did and the rebase kept both entries — with the duplicate at the head of the list, so it took priority:

List	Duplicated kernel	Effect of the stale head entry
_POSSIBLE_FP8_KERNELS	MarlinFP8ScaledMMLinearKernel	Marlin ahead of FlashInfer and Cutlass
_POSSIBLE_NVFP4_KERNELS	B12xNvFp4LinearKernel	B12X ahead of every FlashInfer path
_POSSIBLE_MXFP4_KERNELS	B12xMxFp4LinearKernel	B12X ahead of FlashInfer/Marlin/Humming

Blame confirms the provenance: the head entries are fork commits e080b6ed0e (2026-07-21) and the pre-vllm-project#52182 Marlin position; the lower entries are upstream 3c79b1a8bf (vllm-project#52016) and bca7bea240 (vllm-project#52182). Upstream's positions are the maintained ones and the registration-priority tests already pin them, so the head duplicates are dropped.

Selection via explicit --linear-backend resolves through _LINEAR_BACKEND_KERNEL_MAP, not list order — verified by the test that monkeypatches _get_linear_backend — so the qualified B12X serving path is unaffected.

3. _POSSIBLE_MXFP6_KERNELS defined twice, identically. mypy rejects the redefinition as soon as the file is touched.

4. mxfp8/b12x.py missing three imports. It referenced b12x_warmup_token_counts, Iterable and Any without importing any of them, so warmup_b12x_mxfp8_linear raised NameError as soon as it found a layer to warm.

5. Three test files importing deleted symbols. test_b12x_linear.py imported from both b12x_block and b12x_tensor; the latter's two symbols were unused and one (_b12x_tensor_fp8_linear) no longer exists anywhere. test_b12x_{mxfp4,nvfp4}_linear.py imported the deleted standalone warmup functions.

Test migration

The FP4 warmup tests moved onto the provider protocol rather than being deleted: the dedupe contract now lives in unit.key, so they assert that two layers with the same weight signature produce equal keys and a differing one does not, then compile the distinct units. The per-call weight/alpha/bias assertions carry over unchanged, so the same behaviour is pinned — only the entry point moved. Marker assertions (layer.b12x_mxfp4_linear) became provider-identity assertions, matching what process_weights_after_loading now sets.

Test Plan

.venv/bin/python -m pytest tests/model_executor/kernels/ -q
.venv/bin/python -m pytest tests/v1/attention/test_flashinfer_sparse_mla_sm120_api.py -q
.venv/bin/pre-commit run --files <the seven changed files>

Duplicate check:

for name in ("_POSSIBLE_FP8_KERNELS", "_POSSIBLE_FP8_BLOCK_KERNELS",
             "_POSSIBLE_MXFP8_KERNELS", "_POSSIBLE_MXFP4_KERNELS",
             "_POSSIBLE_NVFP4_KERNELS", "_POSSIBLE_MXFP6_KERNELS"):
    for platform, kernels in getattr(L, name).items():
        assert len({k.__name__ for k in kernels}) == len(kernels)

Test Result

• tests/model_executor/kernels/: 49 passed — previously 3 of 4 files could not be collected
• tests/v1/attention/test_flashinfer_sparse_mla_sm120_api.py: 5 passed — previously uncollectable
• duplicate check: 0 duplicate lists remaining (was 4)
• pre-commit: exit 0 (ruff, ruff-format, mypy 3.10)

Known remaining drift

Not fixed here, deliberately, because resolving it needs on-device evidence rather than a guess:

scaled_mm/b12x_tensor.py survives as a duplicate of b12x.py's B12xTensorFP8ScaledMMLinearKernel, and warmup/kernel_warmup.py still calls its standalone warmup_b12x_tensor_fp8_linear while also calling b12x_warmup. Both select layers by the b12x_tensor_fp8_packed_weight attribute that b12x.py sets, so the tensor-FP8 path is plausibly warmed twice. Confirming that — and removing the stale file — wants a GB10 warmup log, so it is left for the image build.

AI assistance was used for this change.