[Bug][Help Wanted]: PPLX + vLLM CUTLASS FP8 Gives Incorrect Responses

Your current environment

• B200

Issue occurs on:

• v0.13.0
• main

not sure the earliest version

🐛 Describe the bug

I get incorrect results with PPLX + vLLM CUTLASS

GPUS := "2"
PORT := "8004"
MODEL := "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic"

launch_dp_ep_pplx:
	VLLM_USE_FLASHINFER_MOE_FP4=0 \
	VLLM_USE_FLASHINFER_MOE_FP8=0 \
	VLLM_USE_DEEP_GEMM=0 \
	VLLM_FLASHINFER_MOE_BACKEND=throughput \
	chg run --gpus {{GPUS}} -- vllm serve {{MODEL}} -dp {{GPUS}} --enable-expert-parallel --port {{PORT}} --enforce-eager --all2all-backend pplx --max-model-len 8192


eval:
	lm_eval \
		--model local-completions \
		--tasks gsm8k \
		--model_args "model={{MODEL}},base_url=http://localhost:{{PORT}}/v1/completions,num_concurrent=1000,tokenized_requests=False" --limit 1000

local-completions (model=RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic,base_url=http://localhost:8004/v1/completions,num_concurrent=1000,tokenized_requests=False), gen_kwargs: (None), limit: 1000.0, num_fewshot: None, batch_size: 1
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.307|±  |0.0146|
|     |       |strict-match    |     5|exact_match|↑  |0.280|±  |0.0142|

expect 90%

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.