[Bugfix] Fix Maverick correctness by filling zero to cache space in cutlass_moe

[minosfuture]

Purpose

vllm-project#19667 changed the workspace creation from torch.zeros to torch.empty. This ends up causing correctness for models using cutlass_moe, e.g. Maverick in our test case. This PR fixes the correctness issue by explicitly filling zeros in cutlass_moe.

Test Plan

lm_eval, ut

Test Result

lm_eval results:

local-chat-completions (model=meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8,base_url=http://127.0.0.1:8081/v1/chat/completions,num_concurrent=32), gen_kwargs: (None), limit: 200.0, num_fewshot: 5, batch_size: 1

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.935	±	0.0175
		strict-match	5	exact_match	↑	0.920	±	0.0192

unit test stability verified:

• without c1.fill_(0), the following one liner verifies stable failure:

for i in {1..10}; do echo $i; pytest -s tests/kernels/moe/test_cutlass_moe.py  -k "test_run_cutlass_moe_fp8 or test_cutlass_moe_8_bit_EP_large" -v  2>&1 > /dev/null && { echo "shouldn't succeed"; exit 1; } done`

• with c1.fill_(0), the following verifies stable success:

for i in {1..10}; do echo $i; pytest -s tests/kernels/moe/test_cutlass_moe.py  -k "test_run_cutlass_moe_fp8 or test_cutlass_moe_8_bit_EP_large" -v  2>&1 > /dev/null || { echo "should succeed"; exit 1; } done

(Optional) Documentation Update

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)

[yeqcharlotte]

great! any way to capture this in test_cutlass_moe?

[minosfuture]

yep, added a couple unit tests