From efed8a1e8345de2b4398a48243301d6a5b1a361e Mon Sep 17 00:00:00 2001 From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com> Date: Fri, 17 Jul 2026 13:24:59 -0400 Subject: [PATCH] [ROCm][Perf][DSV4] Improve sparse decode reduction occupancy on gfx950 (#48788) Signed-off-by: fai --- vllm/v1/attention/ops/rocm_aiter_mla_sparse.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py b/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py index c3faee4024..96f7693430 100644 --- a/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py +++ b/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py @@ -2205,7 +2205,7 @@ def _rocm_sparse_attn_decode_ragged_triton( num_warps=4, ) - _sparse_attn_decode_reduce_kernel[(num_queries, heads_blocks)]( + _sparse_attn_decode_reduce_kernel[(num_queries, num_heads)]( part_m, part_l, part_acc, @@ -2221,7 +2221,7 @@ def _rocm_sparse_attn_decode_ragged_triton( num_heads, HAS_ATTN_SINK=has_attn_sink, COMB_DIM=comb_dim, - BLOCK_H=block_h, + BLOCK_H=1, NUM_SPLITS=num_splits, SPLITS_PAD=triton.next_power_of_2(num_splits), num_warps=4,