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 8, 22:45 UTC
[Misc] Delay deprecation of CommonAttentionMetadata properties (#33801)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Feb 4, 2026 at 09:41 UTC
0e922986222d9c25ce320fbdd0aff20279c2cb93
1 file changed
+2
-2
vllm/v1/attention/backend.py
+2
-2
@@ -347,7 +347,7 @@ class CommonAttentionMetadata:
347
"""
348
Prefer using device seq_lens directly to avoid implicit H<>D sync.
349
If a CPU copy is needed, use `seq_lens.cpu()` instead.
350
- Will be removed in a future release (v0.15.0)
350
+ Will be removed in a future release, please migrate as soon as possible.
351
"""
352
)
353
def seq_lens_cpu(self) -> torch.Tensor:
@@ -361,7 +361,7 @@ class CommonAttentionMetadata:
361
Prefer using device seq_lens directly to avoid implicit H<>D sync which breaks full
362
async scheduling. If a CPU copy is needed, it can be derived from
363
query_start_loc_cpu and seq_lens.
364
- Will be removed in a future release (v0.15.0)
364
+ Will be removed in a future release, please migrate as soon as possible.
365
"""
366
)
367
def num_computed_tokens_cpu(self) -> torch.Tensor: