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 11, 00:00 UTC
[Bugfix] Enable `dynamic_dims` for different embeds shape (#31223)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Dec 24, 2025 at 10:15 UTC
dd424571c8edaf7c68fe4ff400da8ef9f26a1e48
3 files changed
+3
-3
vllm/model_executor/models/audioflamingo3.py
+1
-1
@@ -111,7 +111,7 @@ class AudioFlamingo3EmbeddingInputs(TensorSchema):
111
112
audio_embeds: Annotated[
113
list[torch.Tensor],
114
- TensorShape("bn", "naf", "hs"),
114
+ TensorShape("bn", "naf", "hs", dynamic_dims={"naf"}),
115
]
116
117
vllm/model_executor/models/minicpmv.py
+1
-1
@@ -139,7 +139,7 @@ class MiniCPMVImageEmbeddingInputs(TensorSchema):
139
type: Literal["image_embeds"]
140
image_embeds: Annotated[
141
torch.Tensor | list[torch.Tensor],
142
- TensorShape("bn", "ns", "hs"),
142
+ TensorShape("bn", "ns", "hs", dynamic_dims={"ns"}),
143
]
144
145
vllm/model_executor/models/qwen2_audio.py
+1
-1
@@ -101,7 +101,7 @@ class Qwen2AudioEmbeddingInputs(TensorSchema):
101
102
audio_embeds: Annotated[
103
list[torch.Tensor],
104
- TensorShape("bn", "naf", "hs"),
104
+ TensorShape("bn", "naf", "hs", dynamic_dims={"naf"}),
105
]
106
107