A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 15:15 UTC
[Attention][3/n] Remove usage of deprecated `seq_lens_cpu` and `num_computed_tokens_cpu` CommonAttentionMetadata properties (#31850)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Jan 7, 2026 at 00:31 UTC
c7a79d41a03f925942e8fb8bc589df4f39bcb950
2 files changed
+3
-3
vllm/v1/attention/backends/rocm_aiter_fa.py
+2
-2
@@ -337,7 +337,7 @@ class AiterFlashAttentionMetadataBuilder(
337
338
query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
339
340
- seq_lens = common_attn_metadata.seq_lens_cpu
340
+ seq_lens = common_attn_metadata.seq_lens.cpu()
341
342
query_lens_cpu = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
343
@@ -367,7 +367,7 @@ class AiterFlashAttentionMetadataBuilder(
367
if num_extends > 0:
368
num_extends_slice = slice(num_decodes, num_decodes + num_extends)
369
query_lens_for_extend = query_lens_cpu[num_extends_slice]
370
- seq_lens_for_extend = common_attn_metadata.seq_lens_cpu[num_extends_slice]
370
+ seq_lens_for_extend = seq_lens[num_extends_slice]
371
computed_kv_lens = seq_lens_for_extend - query_lens_for_extend
372
swa_metadata = None
373
if self.aot_sliding_window is not None:
vllm/v1/attention/backends/rocm_attn.py
+1
-1
@@ -124,7 +124,7 @@ class RocmAttentionMetadataBuilder(AttentionMetadataBuilder[RocmAttentionMetadat
124
prefix_kv_lens = torch.tensor(
125
[common_prefix_len], dtype=torch.int32, device=self.device
126
)
127
- suffix_kv_lens = common_attn_metadata.seq_lens_cpu - common_prefix_len
127
+ suffix_kv_lens = common_attn_metadata.seq_lens.cpu() - common_prefix_len
128
suffix_kv_lens = suffix_kv_lens.to(self.device)
129
else:
130
cu_prefix_query_lens = None