vllm
[ROCm][Perf] QSA: run FP8 K/V sparse attention as FP8 MFMA on gfx942 and gfx950
#69266
Open

[ROCm][Perf] QSA: run FP8 K/V sparse attention as FP8 MFMA on gfx942 and gfx950 #69266

mjkvaak-amd
mjkvaak-amd [ROCm][Perf][Qwen4Exp] Pad BLOCK_M and retune sparse GQA BF16 profile…
af8c32d0
mjkvaak-amd [ROCm][Perf][Qwen4Exp] Bound QSA indexer logits to the longest sequence
9f87f0d7
mjkvaak-amd [ROCm][Perf][Qwen4Exp] Score single-row QSA decode per row on CDNA4 too
d718fbee
mjkvaak-amd [ROCm][Qwen4Exp] Test the CDNA4 row profile, column bound and padded GQA
09e86026
mjkvaak-amd [ROCm][Qwen4Exp] Address review: bf16 profile gate, builder-fed index…
1f98e674
mjkvaak-amd [ROCm][Perf] QSA: run FP8 K/V sparse attention as FP8 MFMA on CDNA3/C…
a32001f5
mergify mergify added qwen
mergify mergify added rocm

Login to write a write a comment.

Login via GitHub

Reviewers
No reviews
Assignees
No one assigned
Labels
Milestone