From c7a79d41a03f925942e8fb8bc589df4f39bcb950 Mon Sep 17 00:00:00 2001 From: Lucas Wilkinson Date: Wed, 7 Jan 2026 00:31:34 -0500 Subject: [PATCH] [Attention][3/n] Remove usage of deprecated `seq_lens_cpu` and `num_computed_tokens_cpu` CommonAttentionMetadata properties (#31850) Signed-off-by: Lucas Wilkinson --- vllm/v1/attention/backends/rocm_aiter_fa.py | 4 ++-- vllm/v1/attention/backends/rocm_attn.py | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/vllm/v1/attention/backends/rocm_aiter_fa.py b/vllm/v1/attention/backends/rocm_aiter_fa.py index b6aa0ae2b..501f197c5 100644 --- a/vllm/v1/attention/backends/rocm_aiter_fa.py +++ b/vllm/v1/attention/backends/rocm_aiter_fa.py @@ -337,7 +337,7 @@ class AiterFlashAttentionMetadataBuilder( query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu - seq_lens = common_attn_metadata.seq_lens_cpu + seq_lens = common_attn_metadata.seq_lens.cpu() query_lens_cpu = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1] @@ -367,7 +367,7 @@ class AiterFlashAttentionMetadataBuilder( if num_extends > 0: num_extends_slice = slice(num_decodes, num_decodes + num_extends) query_lens_for_extend = query_lens_cpu[num_extends_slice] - seq_lens_for_extend = common_attn_metadata.seq_lens_cpu[num_extends_slice] + seq_lens_for_extend = seq_lens[num_extends_slice] computed_kv_lens = seq_lens_for_extend - query_lens_for_extend swa_metadata = None if self.aot_sliding_window is not None: diff --git a/vllm/v1/attention/backends/rocm_attn.py b/vllm/v1/attention/backends/rocm_attn.py index 3701373f3..b9b32d07b 100644 --- a/vllm/v1/attention/backends/rocm_attn.py +++ b/vllm/v1/attention/backends/rocm_attn.py @@ -124,7 +124,7 @@ class RocmAttentionMetadataBuilder(AttentionMetadataBuilder[RocmAttentionMetadat prefix_kv_lens = torch.tensor( [common_prefix_len], dtype=torch.int32, device=self.device ) - suffix_kv_lens = common_attn_metadata.seq_lens_cpu - common_prefix_len + suffix_kv_lens = common_attn_metadata.seq_lens.cpu() - common_prefix_len suffix_kv_lens = suffix_kv_lens.to(self.device) else: cu_prefix_query_lens = None