biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Attention][3/n] Remove usage of deprecated seq_lens_cpu and num_computed_tokens_cpu CommonAttentionMetadata properties (#31850)

Browse Source 
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
This commit is contained in:
Lucas Wilkinson
2026-01-07 00:31:34 -05:00
committed by GitHub
parent 6409004b26
commit c7a79d41a0
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/v1/attention/backends/rocm_aiter_fa.py
View File

@@ -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:

2
vllm/v1/attention/backends/rocm_attn.py
View File

@@ -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