[ROCm][Performance] Avoid blocking MiniMax M3 scalar upload (#56170)

Signed-off-by: fai <fangzhouai@gmail.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Fangzhou Ai
2026-09-13 16:56:02 +00:00
committed by GitHub
co-authored by OpenAI Codex mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent b7e0cdac5d
commit f09c52a587
@@ -314,7 +314,10 @@ class MiniMaxM3SparseMetadataBuilder(AttentionMetadataBuilder[MiniMaxM3SparseMet
prefill_cu_seqlens_k = torch.empty(
num_prefills + 1, dtype=torch.int32, device=seq_lens.device
)
prefill_cu_seqlens_k[0] = 0
if current_platform.is_rocm():
prefill_cu_seqlens_k[:1].zero_()
else:
prefill_cu_seqlens_k[0] = 0
torch.cumsum(prefill_kv_lens, dim=0, out=prefill_cu_seqlens_k[1:])
prefill_cu_seqlens_q = (
query_start_loc[num_decodes:] - num_decode_tokens