[ROCm][Perf] Enable fused indexer-Q RoPE+quant kernel for DeepSeek/GLM sparse attention - #47335

Open

xaguilar-amd wants to merge 4 commits into

vllm-project:mainfrom

xaguilar-amd:rocm-enable-fused-indexer-q

Commits on Aug 21, 2026

Enable fused indexer-Q RoPE+quant kernel
xaguilar-amd
committed
d3f91d9
Replace some cuda() and rocm() checks with current_platform.is_cuda_alike()
xaguilar-amd
committed
1fa6a88
Applied ruff to the files to avoid pre-commit failures
xaguilar-amd
committed
02fdf57

Commits on Sep 17, 2026

Merge branch 'main' into rocm-enable-fused-indexer-q
akii96
authored
3cd0a99