[ROCm][DSv3.2] Eliminate per-decode FillFunctor launches in sparse-MLA hot loop (#44527)

Signed-off-by: Frida Andersson <fanderss@amd.com>
This commit is contained in:
frida-andersson
2026-07-28 10:12:20 +08:00
committed by GitHub
parent 1206891822
commit 02b6ecf07c
@@ -429,7 +429,6 @@ def rocm_fp8_paged_mqa_logits(
(out_logits,) = current_workspace_manager().get_simultaneous(
((batch_size * next_n, max_model_len), torch.float32),
)
out_logits.fill_(float("-inf"))
deepgemm_fp8_paged_mqa_logits(
q_fp8,
kv_cache_fp8,
@@ -715,7 +714,6 @@ def rocm_aiter_sparse_attn_indexer(
scale_fmt,
)
topk_indices_buffer[: hidden_states.shape[0]] = -1
if has_prefill:
prefill_metadata = layer_attn_metadata.prefill
assert prefill_metadata is not None