[khluu]

Purpose

Fix :nvidia: (4xB200) DeepSeek-V4.1 NVIDIA Kernels (nightly). It has failed 12 times in a row since #64111 merged (latest: 99485):

FAILED models/deepseek_v41/nvidia/specialized_decode/test_kernels.py::test_engram_wkv_prefetch_matches_inline[False]
FAILED models/deepseek_v41/nvidia/specialized_decode/test_kernels.py::test_engram_wkv_prefetch_matches_inline[True]
AttributeError: '_Mxfp8Linear' object has no attribute 'params_dtype'

Root cause: #64111 changed FlashInferCutedslMxfp8LinearKernel.apply_weights (vllm/model_executor/kernels/linear/mxfp8/flashinfer.py) to take the output dtype of unquantized inputs from layer.params_dtype instead of x.dtype. Real vLLM linear layers always set params_dtype, but the hand-built _Mxfp8Linear stand-in in this test did not.

Why #64111's CI didn't catch it: the step is optional: true (nightly only), so it never runs on PRs, and its source triggers don't include vllm/model_executor/kernels/linear/mxfp8/.

Fix: set params_dtype = torch.bfloat16 on the stand-in, as a real linear layer would. The test feeds bf16 inputs, so the kernel's output dtype is the same as before #64111.

Not a duplicate: no open PR touches this test or stub.

Test Plan

Not run locally (needs B200 GPUs). The step is nightly-only, so it will be confirmed by the next nightly after merge, or by a selected-step run of dsv41-nvidia-kernels-4xb200 on this PR.

Notes

AI assistance (Claude Code) was used; reviewed by the submitter.

🤖 Generated with Claude Code

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}