20231088/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[Bugfix] Fix phi3v batch inference when images have different aspect ratio (#7392)

Browse Source
This commit is contained in:
Isotr0py 2024-08-11 00:19:33 +08:00 committed by GitHub
parent baa240252e
commit 4c5d8e8ea9
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 24 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tests/models/test_phi3v.py
View File

@ -81,7 +81,10 @@ def run_test(
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
[
rescale_image_size(image, factor, transpose=idx)
for idx, factor in enumerate(size_factors)
],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
# NOTE: take care of the order. run vLLM first, and then run HF.

4
tests/tracing/test_tracing.py
View File

@ -114,5 +114,5 @@ def test_traces(trace_service):
SpanAttributes.LLM_LATENCY_TIME_TO_FIRST_TOKEN) == ttft
e2e_time = metrics.finished_time - metrics.arrival_time
assert attributes.get(SpanAttributes.LLM_LATENCY_E2E) == e2e_time
assert attributes.get(SpanAttributes.LLM_LATENCY_TIME_IN_SCHEDULER
) == metrics.scheduler_time
assert attributes.get(
SpanAttributes.LLM_LATENCY_TIME_IN_SCHEDULER) == metrics.scheduler_time

24
vllm/model_executor/models/phi3v.py
View File

@ -189,7 +189,7 @@ class Phi3HDImageEmbedding(Phi3ImageEmbeddingBase):
global_image_features_hd_newline = self.add_image_newline(
global_image_features_hd)
all_image_embeddings = []
batch_image_features_proj = []
# need a for loop to process each image because of different image sizes
# (patch arrangement is different for each image)
for i, img_size in enumerate(image_sizes):
@ -207,19 +207,17 @@ class Phi3HDImageEmbedding(Phi3ImageEmbeddingBase):
sub_image_features_hd)
# [sub features, separator, global features]
all_image_embeddings.append(
torch.cat([
sub_image_features_hd_newline.squeeze(
0), # (h_crop*12*(w_crop*12+1), 4096)
self.glb_GN.squeeze(0),
global_image_features_hd_newline[i],
]))
image_embeddings = torch.cat([
sub_image_features_hd_newline.squeeze(
0), # (h_crop*12*(w_crop*12+1), 4096)
self.glb_GN.squeeze(0),
global_image_features_hd_newline[i],
])
img_proj = self.img_projection(
image_embeddings.to(target_device, target_dtype))
batch_image_features_proj.append(img_proj)
image_features_proj = self.img_projection(
torch.stack(all_image_embeddings).to(target_device, target_dtype)
) # (num_images, (h_crop*12*(w_crop*12+1)+1), hidden_size)
return image_features_proj
return batch_image_features_proj
def reshape_hd_patches_2x2merge(self, image_features, h_crop, w_crop):
"""

9
vllm/multimodal/utils.py
View File

@ -90,8 +90,13 @@ def load_image_from_base64(image: Union[bytes, str]) -> Image.Image:
return _load_image_from_bytes(base64.b64decode(image))
def rescale_image_size(image: Image.Image, size_factor: float) -> Image.Image:
def rescale_image_size(image: Image.Image,
size_factor: float,
transpose: int = -1) -> Image.Image:
"""Rescale the dimensions of an image by a constant factor."""
new_width = int(image.width * size_factor)
new_height = int(image.height * size_factor)
return image.resize((new_width, new_height))
image = image.resize((new_width, new_height))
if transpose >= 0:
image = image.transpose(Image.Transpose(transpose))
return image