[ROCm][Perf][DSV4] Improve sparse decode reduction occupancy on gfx950 (#48788)
Signed-off-by: fai <fangzhouai@gmail.com>
Fangzhou Ai committed
Jul 17, 2026 at 13:24 UTC
efed8a1e8345de2b4398a48243301d6a5b1a361e
1 file changed
+2
-2
vllm/v1/attention/ops/rocm_aiter_mla_sparse.py
+2
-2
@@ -2205,7 +2205,7 @@ def _rocm_sparse_attn_decode_ragged_triton(
2205
num_warps=4,
2206
)
2207
2208
- _sparse_attn_decode_reduce_kernel[(num_queries, heads_blocks)](
2208
+ _sparse_attn_decode_reduce_kernel[(num_queries, num_heads)](
2209
part_m,
2210
part_l,
2211
part_acc,
@@ -2221,7 +2221,7 @@ def _rocm_sparse_attn_decode_ragged_triton(
2221
num_heads,
2222
HAS_ATTN_SINK=has_attn_sink,
2223
COMB_DIM=comb_dim,
2224
- BLOCK_H=block_h,
2224
+ BLOCK_H=1,
2225
NUM_SPLITS=num_splits,
2226
SPLITS_PAD=triton.next_power_of_2(num_splits),
2227
num_warps=4,