From f09c52a587f205963d3a2b8f1c42820a831ee4b0 Mon Sep 17 00:00:00 2001 From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com> Date: Sun, 13 Sep 2026 12:56:02 -0400 Subject: [PATCH] [ROCm][Performance] Avoid blocking MiniMax M3 scalar upload (#56170) Signed-off-by: fai Co-authored-by: OpenAI Codex Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> --- vllm/models/minimax_m3/common/sparse_attention.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/vllm/models/minimax_m3/common/sparse_attention.py b/vllm/models/minimax_m3/common/sparse_attention.py index 50c03dd9bf..f385f82067 100644 --- a/vllm/models/minimax_m3/common/sparse_attention.py +++ b/vllm/models/minimax_m3/common/sparse_attention.py @@ -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