huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-09-11 20:34:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: adjust resize case for qwen2_vl warmup

Browse Source
This commit is contained in:
David Holtz 2024-10-29 15:47:32 +00:00
parent 4f90db47be
commit 77eb07f73b
1 changed files with 8 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
server/text_generation_server/models/vlm_causal_lm.py
View File

@ -177,6 +177,14 @@ class VlmCausalLMBatch(FlashCausalLMBatch):
pass pass
elif chunk_type == "image": elif chunk_type == "image":
image = Image.open(BytesIO(chunk.image.data)) image = Image.open(BytesIO(chunk.image.data))
# qwen2_vl expects images to be greater than 20 pixels, this is for warmup since the
# default warmup image is 20x20
if config.model_type == "qwen2_vl":
if image.width <= 20:
w = image.width * 2
h = image.height * 2
image = image.resize((w, h))
if config.model_type == "llava_next": if config.model_type == "llava_next":
images.append(image) images.append(image)
else: else:
@ -185,11 +193,6 @@ class VlmCausalLMBatch(FlashCausalLMBatch):
raise RuntimeError(f"Invalid chunk type {chunk_type}") raise RuntimeError(f"Invalid chunk type {chunk_type}")
if images: if images:
if images[0][0].width <= 20:
# TODO: provide a better way to handle the issue of the prefill image being too small
images = images[0][0].resize(
(images[0][0].width * 2, images[0][0].height * 2)
)
image_inputs = processor.image_processor(images, return_tensors="pt") image_inputs = processor.image_processor(images, return_tensors="pt")
else: else:
image_inputs = None image_inputs = None