[ROCm] Enable FlyDSL w4a16 MoE for Kimi INT4

[amd-asalykov]

Replace default triton w4a16 MoE kernel with more performant FlyDSL implementation for Kimi INT4 MI355X

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview, Comment	Jun 15, 2026 8:35pm

[gemini-code-assist]

Code Review

This pull request updates the ROCm deployment configuration for Kimi-K2.5 in the markdown documentation by removing obsolete environment variables and adding new CLI arguments (--moe-backend flydsl and --compilation-config). The review feedback points out that these changes should be consistently applied to other deployment configurations, such as the Docker run command and YAML configuration files. Additionally, a duplicated --mm-encoder-tp-mode data argument was identified in the updated command block.

I am having trouble creating individual review comments. Click here to see my feedback.

moonshotai/Kimi-K2.5.md (242-243)

The environment variables VLLM_ROCM_QUICK_REDUCE_QUANTIZATION and VLLM_ROCM_USE_AITER_RMSNORM have been removed here, but they are still present in the AMD (ROCm) Docker run command (lines 105-106) and the YAML configuration (models/moonshotai/Kimi-K2.5.yaml lines 103-104). To maintain consistency across all deployment methods, please update those sections as well.

moonshotai/Kimi-K2.5.md (250-251)

The new arguments --moe-backend flydsl and --compilation-config are added here, but they are missing from the AMD (ROCm) Docker run command (lines 108-118) and the YAML configuration (models/moonshotai/Kimi-K2.5.yaml). Please update those configurations to enable the FlyDSL MoE backend consistently. Additionally, the --mm-encoder-tp-mode data argument is duplicated in this command block (on line 244 and line 249). One of them should be removed.