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 16, 04:30 UTC
[Bugfix] Fix broken phi3-v mm_processor_kwargs tests (#11263)
Signed-off-by: Isotr0py <2037008807@qq.com>
Isotr0py committed
Dec 19, 2024 at 02:16 UTC
996aa70f00818933866d8cfdbbf8f131a6a63664
2 files changed
+17
-12
tests/models/decoder_only/vision_language/mm_processor_kwargs/test_phi3v.py
+5
-7
@@ -58,16 +58,14 @@ def test_max_tokens_override(get_max_phi3v_image_tokens, model: str,
58
59
@pytest.mark.parametrize("model", models)
60
@pytest.mark.parametrize(
61
- "num_crops,expected_toks_per_img,num_imgs",
61
+ "num_crops,expected_toks_per_img",
62
[
63
- (4, 757, 1),
64
- (4, 757, 2),
65
- (16, 1921, 1),
66
- (16, 1921, 2),
63
+ (4, 757),
64
+ (16, 1921),
65
# the default num_crops of phi-3.5-vision is 4
68
- (None, 757, 2),
69
- (None, 757, 2),
66
+ (None, 757),
67
])
68
+@pytest.mark.parametrize("num_imgs", [1, 2])
69
def test_processor_override(processor_for_phi3v, image_assets: _ImageAssets,
70
model: str, num_crops: Optional[int],
71
expected_toks_per_img: int, num_imgs: int):
vllm/model_executor/models/phi3v.py
+12
-5
@@ -302,11 +302,18 @@ class Phi3HDImageEmbedding(Phi3ImageEmbeddingBase):
302
return image_features_hd_newline
303
304
305
-def get_max_phi3v_image_tokens(ctx: InputContext) -> int:
306
- processor = ctx.get_hf_processor()
307
- image_processor = processor.image_processor # type: ignore
308
-
309
- return image_processor.calc_num_image_tokens_from_image_size(
305
+def get_max_phi3v_image_tokens(
306
+ ctx: InputContext,
307
+ *,
308
+ num_crops: Optional[int] = None,
309
+) -> int:
310
+ mm_processor_kwargs = {}
311
+ if num_crops:
312
+ mm_processor_kwargs["num_crops"] = num_crops
313
+
314
+ processor = ctx.get_hf_processor(**mm_processor_kwargs)
315
+
316
+ return processor.calc_num_image_tokens_from_image_size(
317
width=MAX_IMAGE_FEATURE_SIZE_WIDTH,
318
height=MAX_IMAGE_FEATURE_SIZE_HEIGHT,
319
)