diff --git a/vllm/v1/attention/backends/rocm_attn.py b/vllm/v1/attention/backends/rocm_attn.py index 331b5c3fb3..f73e558ec2 100644 --- a/vllm/v1/attention/backends/rocm_attn.py +++ b/vllm/v1/attention/backends/rocm_attn.py @@ -103,11 +103,9 @@ class RocmAttentionMetadataBuilder(AttentionMetadataBuilder[RocmAttentionMetadat # slow, so here we set it to 1. attn_metadata.seq_lens.fill_(1) - # Here we set the query start locs to 0. This is to - # cover up an invalid memory access in the prefix_prefil kernel - # that we run into during graph capture (#25985) + # Zero device query start locations to avoid invalid memory access in + # the prefix prefill kernel during graph capture (#25985). common_attn_metadata.query_start_loc.zero_() - common_attn_metadata.query_start_loc_cpu.zero_() return attn_metadata