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, 09:00 UTC
[Misc] Delay deprecation of CommonAttentionMetadata properties (#32074)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Jan 10, 2026 at 00:06 UTC
da6709c9fe6965b7348692576ffadeee8439388e
1 file changed
+3
-3
vllm/v1/attention/backends/utils.py
+3
-3
@@ -96,7 +96,7 @@ class CommonAttentionMetadata:
96
dcp_local_seq_lens_cpu: torch.Tensor | None = None
97
"""Sequence lengths of the local rank in decode context parallelism world"""
98
99
- # WARNING: Deprecated fields. Will be removed in a future release (v0.14.0)
99
+ # WARNING: Deprecated fields. Will be removed in a future release (v0.15.0)
100
_seq_lens_cpu: torch.Tensor | None = None
101
_num_computed_tokens_cpu: torch.Tensor | None = None
102
@@ -107,7 +107,7 @@ class CommonAttentionMetadata:
107
"""
108
Prefer using device seq_lens directly to avoid implicit H<>D sync.
109
If a CPU copy is needed, use `seq_lens.cpu()` instead.
110
- Will be removed in a future release (v0.14.0)
110
+ Will be removed in a future release (v0.15.0)
111
"""
112
)
113
def seq_lens_cpu(self) -> torch.Tensor:
@@ -121,7 +121,7 @@ class CommonAttentionMetadata:
121
Prefer using device seq_lens directly to avoid implicit H<>D sync which breaks full
122
async scheduling. If a CPU copy is needed, it can be derived from
123
query_start_loc_cpu and seq_lens.
124
- Will be removed in a future release (v0.14.0)
124
+ Will be removed in a future release (v0.15.0)
125
"""
126
)
127
def num_computed_tokens_cpu(self) -> torch.Tensor: