[ROCm][Perf][DSV4] Enable split sparse decode on gfx942 (#46275)

Signed-off-by: Tuukka Sarvi <tuukka.sarvi@amd.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Tuukka Sarvi
2026-07-16 12:28:20 +00:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent d803b44dbe
commit b8168e33e0
2 changed files with 44 additions and 41 deletions
@@ -2104,7 +2104,7 @@ def _rocm_sparse_attn_decode_ragged_triton(
comb_dim = nope_head_dim + rope_head_dim
is_fnuz = current_platform.is_fp8_fnuz()
if not _ON_GFX950: # Fallback path for un-tuned architectures.
if not (_ON_GFX942 or _ON_GFX950): # Fallback path for un-tuned architectures.
block_k = 16 if head_dim >= 256 else 32
_sparse_attn_decode_ragged_kernel[(num_queries, heads_blocks)](
q,