[ROCm][Kimi-K3] add VLLM_ROCM_USE_AITER_MOE_SITUV2_A8W4 for the a8w4 …

[fanxingran]

…MoE path

vllm-project/vllm#50582 renamed the switch; the pinned kimi-k3 image still reads AITER_SITUV2_A8W4, so set both. Mixing them corrupts MoE output.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview	Aug 4, 2026 7:17am

[gemini-code-assist]

Code Review

This pull request updates the configuration for the Kimi-K3 model in models/moonshotai/Kimi-K3.yaml by adding the VLLM_ROCM_USE_AITER_MOE_SITUV2_A8W4: "1" environment variable. This ensures compatibility across both newer vLLM versions and older pinned images, preventing silent MoE output corruption. The documentation guide for AMD hardware is also updated to reflect this change and warn against mixed configurations. There are no review comments, so I have no additional feedback to provide.

[fanxingran]

cc @Fangzhou-Ai @hongxiayang — sorry to ping you directly, and please
correct me if I've misread anything here.
vllm-project/vllm#50582 renamed the a8w4 SiTU MoE switch to
VLLM_ROCM_USE_AITER_MOE_SITUV2_A8W4, while the pinned kimi-k3 image (Jul 27)
still reads only AITER_SITUV2_A8W4. This PR keeps both so either build lands on
the same fp8 kernels. As far as I can tell, the old name alone on a newer build is
silently wrong: vLLM preshuffles w13 separated while AITER picks the _gui
interleaved kernels.
Verified on 8x MI355X (TP8): kernels resolve to
flydsl_moe1_afp8_wfp4_bf16_*_gui_fp8 and the output looks clean. Happy to drop
the AITER_* pair once the image is refreshed — please let me know if you'd
prefer that now, or if there's a validation you'd like me to run first. Thanks
for taking a look!