[ROCm][DSv4] Functional fixes for DeepSeek V4 on MI300X (gfx942)#42893

Draft

maeehart wants to merge 9 commits into

vllm-project:mainfrom

maeehart:mahartik/dsv4-rocm-mi300x-fixes

Commits on May 25, 2026

[ROCm][DSv4] MI300X (gfx942) support for DeepSeek V4

authored and
maeehart
committed
[ROCm][DSv4] Use FNUZ FP8 on gfx942 in fused KV insert kernel

authored and
maeehart
committed
[ROCm][DSv4] Use tl.float8e4b8 for FNUZ on MI300X sparse MLA kernels
maeehart
committed
[ROCm][DSv4] Zero ROCm sparse-MLA prefill KV workspace

maeehart
and
cursoragent
committed
[ROCm][DSv4] Propagate FNUZ vs OCP gating to ROCm prefill+decode paths

authored and
maeehart
committed
[ROCm][DSv4] Revert turboquant fp8e4b15 -> fp8e4b8 changes (NVIDIA-only path)

maeehart
and
cursoragent
committed
[ROCm][DSv4][gfx942] Vendor fp8_mqa_logits kernel to avoid AITER upgrade

maeehart
and
cursoragent
committed
[ROCm][DSv4] Use FNUZ FP8_MAX=224 to match vLLM convention (review #42893)

maeehart
and
cursoragent
committed
[ROCm][DSv4] Tighten KV-workspace zero comment + FP8 encoding notes (review #42893)

maeehart
and
cursoragent
committed