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, 05:30 UTC
Fix TensorSchema validation test for symbolic dims (#22366)
Signed-off-by: Benji Beck <benjibeck@meta.com>
Benji Beck committed
Aug 10, 2025 at 10:16 UTC
68b254d67300a1740db900a3d0ff4252424715d7
1 file changed
+18
-14
tests/standalone_tests/test_tensor_schema.py
+18
-14
@@ -4,8 +4,8 @@
4
import pytest
5
import torch
6
7
-from vllm.model_executor.models.fuyu import FuyuImagePatchInputs
7
from vllm.model_executor.models.glm4_1v import Glm4vImageEmbeddingInputs
8
+from vllm.model_executor.models.granite_speech import GraniteSpeechAudioInputs
9
from vllm.model_executor.models.phi3v import Phi3VImagePixelInputs
10
11
@@ -129,23 +129,27 @@ def test_tensor_schema_with_invalid_resolve_binding_dims():
129
130
131
def test_tensor_schema_with_list_of_symbolic_dim():
132
- flat_data = torch.stack([torch.randn(768) for _ in range(3)]) # (bn=3, fn)
133
- patches_per_image = [64, 64, 64] # len = bn = 3
134
-
135
- FuyuImagePatchInputs(
136
- flat_data=flat_data,
137
- patches_per_image=patches_per_image,
132
+ input_features = torch.randn(3, 10, 160) # (b=3, fi=10, 160)
133
+ input_features_mask = torch.randn(3, 8) # (b=3, fo=8)
134
+ audio_embed_sizes = [8, 8, 8] # len = b = 3
135
+
136
+ GraniteSpeechAudioInputs(
137
+ input_features=input_features,
138
+ input_features_mask=input_features_mask,
139
+ audio_embed_sizes=audio_embed_sizes,
140
)
141
142
143
def test_tensor_schema_with_list_of_symbolic_dim_mismatch_in_length():
142
- flat_data = torch.stack([torch.randn(768) for _ in range(4)]) # (bn=4, fn)
143
- patches_per_image = [64, 64, 64] # len = 3 ≠ bn
144
-
145
- with pytest.raises(ValueError, match="expected 'bn'=4, got 3"):
146
- FuyuImagePatchInputs(
147
- flat_data=flat_data,
148
- patches_per_image=patches_per_image,
144
+ input_features = torch.randn(4, 10, 160) # (b=4, fi=10, 160)
145
+ input_features_mask = torch.randn(4, 8) # (b=4, fo=8)
146
+ audio_embed_sizes = [8, 8, 8] # len = 3 ≠ b
147
+
148
+ with pytest.raises(ValueError, match="expected 'b'=4, got 3"):
149
+ GraniteSpeechAudioInputs(
150
+ input_features=input_features,
151
+ input_features_mask=input_features_mask,
152
+ audio_embed_sizes=audio_embed_sizes,
153
)
154
155