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 14, 23:45 UTC
[Bugfix] Fix granite speech shape validation (#21762)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Jul 29, 2025 at 02:19 UTC
e17a4d3bf9cffe32ec308a5979790732818e4919
1 file changed
+4
-3
vllm/model_executor/models/granite_speech.py
+4
-3
@@ -64,14 +64,15 @@ class GraniteSpeechAudioInputs(TensorSchema):
64
65
Dimensions:
66
- b: Batch size
67
- - nf: Number of audio features (variable length)
67
+ - fi: Number of input features from the Mel spectrogram.
68
+ - fo: Number of output features, i.e. the embedding size.
69
- 160: Fixed feature dimension for Mel spectrogram features
70
"""
71
71
- input_features: Annotated[torch.Tensor, TensorShape("b", "nf", 160)]
72
+ input_features: Annotated[torch.Tensor, TensorShape("b", "fi", 160)]
73
"""Audio input features."""
74
74
- input_features_mask: Annotated[torch.Tensor, TensorShape("b", "nf")]
75
+ input_features_mask: Annotated[torch.Tensor, TensorShape("b", "fo")]
76
"""Mask for variable length audio features."""
77
78
audio_embed_sizes: Annotated[list[int], TensorShape("b")]