DeepSeek V4 ROCm: bring-up + throughput tuning on MI300X (11 atomic commits)#9

Closed

fergusfinn wants to merge 11 commits into

mainfrom

codex/deepseek-v4-rocm-bringup

Commits on May 22, 2026

[ROCm] DeepSeek V4: HIP-graph-safe paged MQA + sparse MLA fallbacks
fergusfinn
committed
[ROCm] DeepSeek V4: SWA K-cache layout
fergusfinn
committed
[ROCm] DeepSeek V4: fix compressor + KV cache fnuz FP8 format
fergusfinn
committed
[ROCm] DeepSeek V4: skip sparse top-k when full window already covers seqlen
fergusfinn
committed
[ROCm][MoE][MXFP4] DeepSeek V4 no-LoRA path: routing + direct W2 reduce
fergusfinn
committed
[MoE][Bugfix] mask MXFP4 bitmatrix padding lanes by logical block size
fergusfinn
committed
[ROCm] MLA decode: static metadata for HIP-graph capture
fergusfinn
committed
[ROCm] sparse MLA: pass output buffer through, avoid copy
fergusfinn
committed
[ROCm] sparse MLA decode: cache static bf16 projection weights
fergusfinn
committed
[ROCm] sparse MLA decode: launch-shape + occupancy tuning
fergusfinn
committed
[ROCm][MoE][MXFP4] OGS tile / ramp / epilogue tuning for DSV4
fergusfinn
committed