From e2db0222e9b2ff37f7a1c54a1d8043e61f8b0b8d Mon Sep 17 00:00:00 2001 From: qizixi <22851944+zixi-qi@users.noreply.github.com> Date: Wed, 10 Jun 2026 15:56:49 -0700 Subject: [PATCH] [Perf][Attention] Pin MLA chunked-context metadata tensors so H2D copies are truly non-blocking (#45074) Signed-off-by: zixi-qi --- vllm/model_executor/layers/attention/mla_attention.py | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/vllm/model_executor/layers/attention/mla_attention.py b/vllm/model_executor/layers/attention/mla_attention.py index 140e071c74..b04edcc513 100644 --- a/vllm/model_executor/layers/attention/mla_attention.py +++ b/vllm/model_executor/layers/attention/mla_attention.py @@ -1664,7 +1664,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]): .unsqueeze(1) .expand(-1, num_prefills) * max_context_chunk - ) + ).pin_memory() chunk_ends = torch.min( context_lens_cpu.unsqueeze(0), chunk_starts + max_context_chunk ) @@ -1680,7 +1680,9 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]): max_token_num_over_chunk = chunk_total_token.max().item() token_to_seq_tensor_cpu = torch.zeros( - [num_chunks, max_token_num_over_chunk], dtype=torch.int32 + [num_chunks, max_token_num_over_chunk], + dtype=torch.int32, + pin_memory=True, ) range_idx = torch.arange(num_prefills, dtype=torch.int32) for i in range(num_chunks): @@ -1724,7 +1726,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]): .unsqueeze(1) .expand(-1, num_prefills) * padded_local_max_context_chunk_across_ranks - ) + ).pin_memory() local_chunk_ends = torch.min( padded_local_context_lens_cpu.unsqueeze(0), local_chunk_starts