[ROCm][DSv3.2] Eliminate per-decode FillFunctor launches in sparse-MLA hot loop (#44527)
Signed-off-by: Frida Andersson <fanderss@amd.com>
frida-andersson committed
Jul 28, 2026 at 04:12 UTC
02b6ecf07cf6a3e6dd395f73af5d1904312165cf
1 file changed
-2
vllm/v1/attention/ops/rocm_aiter_mla_sparse.py
-2
@@ -429,7 +429,6 @@ def rocm_fp8_paged_mqa_logits(
429
(out_logits,) = current_workspace_manager().get_simultaneous(
430
((batch_size * next_n, max_model_len), torch.float32),
431
)
432
- out_logits.fill_(float("-inf"))
432
deepgemm_fp8_paged_mqa_logits(
433
q_fp8,
434
kv_cache_fp8,
@@ -715,7 +714,6 @@ def rocm_aiter_sparse_attn_indexer(
714
scale_fmt,
715
)
716
718
- topk_indices_buffer[: hidden_states.shape[0]] = -1
717
if has_prefill:
718
prefill_metadata = layer_attn_metadata.prefill
719
assert prefill_metadata is not None