From 43196f2458bebb822cd825c0370b6f157b66f293 Mon Sep 17 00:00:00 2001 From: Keyi Li <94494390+JasonKeyiL@users.noreply.github.com> Date: Fri, 28 Aug 2026 19:00:46 -0700 Subject: [PATCH] [Perf][MLA Sparse] Pin req_id_per_token before non_blocking H2D on XPU and ROCm (#54295) Signed-off-by: Keyi Li Co-authored-by: Keyi Li Co-authored-by: Claude Opus 4.7 --- vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py | 3 ++- vllm/v1/attention/backends/mla/xpu_mla_sparse.py | 4 ++-- 2 files changed, 4 insertions(+), 3 deletions(-) diff --git a/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py b/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py index 3736a38f5f..54f716ea50 100644 --- a/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py +++ b/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py @@ -17,6 +17,7 @@ from vllm.model_executor.layers.attention.mla_attention import ( ) from vllm.platforms import current_platform from vllm.triton_utils import tl, triton +from vllm.utils.torch_utils import np_to_pinned_tensor from vllm.v1.attention.backend import ( AttentionBackend, AttentionCGSupport, @@ -515,7 +516,7 @@ class ROCMAiterMLASparseMetadataBuilder( self._prev_req_extent = new_req_extent self._prev_indices_extent = new_indices_extent self.req_id_per_token_buffer[:new_req_extent].copy_( - torch.from_numpy(req_id_per_token), non_blocking=True + np_to_pinned_tensor(req_id_per_token), non_blocking=True ) query_lens = ( common_attn_metadata.query_start_loc[1:] diff --git a/vllm/v1/attention/backends/mla/xpu_mla_sparse.py b/vllm/v1/attention/backends/mla/xpu_mla_sparse.py index 8179534468..8aaa4749d5 100644 --- a/vllm/v1/attention/backends/mla/xpu_mla_sparse.py +++ b/vllm/v1/attention/backends/mla/xpu_mla_sparse.py @@ -13,7 +13,7 @@ from vllm.logger import init_logger from vllm.model_executor.layers.attention.mla_attention import ( get_mla_dims, ) -from vllm.utils.torch_utils import is_quantized_kv_cache +from vllm.utils.torch_utils import is_quantized_kv_cache, np_to_pinned_tensor from vllm.v1.attention.backend import ( AttentionBackend, AttentionCGSupport, @@ -158,7 +158,7 @@ class XPUMLASparseMetadataBuilder(AttentionMetadataBuilder[XPUMLASparseMetadata] # Zero-fill for cudagraphs self.req_id_per_token_buffer.fill_(0) self.req_id_per_token_buffer[: req_id_per_token.shape[0]].copy_( - torch.from_numpy(req_id_per_token), non_blocking=True + np_to_pinned_tensor(req_id_per_token), non_blocking=True ) req_id_per_token = self.req_id_per_token_buffer[:num_tokens]