From 20405bfb15546bb988f425a3c95fef9ac59104c7 Mon Sep 17 00:00:00 2001 From: bastefaniak Date: Fri, 14 Aug 2026 11:46:46 +0200 Subject: [PATCH] [Bugfix] Fix Cosmos3-Edge processor after transformers 5.15 release (#51989) Signed-off-by: Bartosz Stefaniak --- tests/models/registry.py | 9 ++++++++- .../processors/cosmos3_edge.py | 16 ++++++++++++---- 2 files changed, 20 insertions(+), 5 deletions(-) diff --git a/tests/models/registry.py b/tests/models/registry.py index 9db20e617e..62cb289e4f 100644 --- a/tests/models/registry.py +++ b/tests/models/registry.py @@ -825,7 +825,14 @@ _MULTIMODAL_EXAMPLE_MODELS = { "Cosmos3EdgeForConditionalGeneration": _HfExamplesInfo( "nvidia/Cosmos3-Edge", max_model_len=4096, - is_available_online=False, + min_transformers_version="5.15", + use_original_num_layers=True, + hf_overrides={ + "text_config": { + "num_hidden_layers": 2, + "hybrid_override_pattern": "*-", + } + }, ), "DeepseekVLV2ForCausalLM": _HfExamplesInfo( "deepseek-ai/deepseek-vl2-tiny", diff --git a/vllm/transformers_utils/processors/cosmos3_edge.py b/vllm/transformers_utils/processors/cosmos3_edge.py index 2644c6b17e..95ca2cccd5 100644 --- a/vllm/transformers_utils/processors/cosmos3_edge.py +++ b/vllm/transformers_utils/processors/cosmos3_edge.py @@ -8,12 +8,16 @@ import torch from torchvision.transforms import InterpolationMode from transformers import AutoTokenizer from transformers.feature_extraction_utils import BatchFeature -from transformers.image_utils import ChannelDimension, PILImageResampling, SizeDict +from transformers.image_utils import ( + ChannelDimension, + PILImageResampling, + SizeDict, + get_image_size, +) from transformers.models.qwen3_vl.processing_qwen3_vl import Qwen3VLProcessor from transformers.models.qwen3_vl.video_processing_qwen3_vl import ( Qwen3VLVideoProcessor, Qwen3VLVideoProcessorInitKwargs, - get_image_size, smart_resize, ) from transformers.models.siglip2.image_processing_siglip2 import ( @@ -22,6 +26,7 @@ from transformers.models.siglip2.image_processing_siglip2 import ( ) from transformers.processing_utils import Unpack from transformers.utils import TensorType +from transformers.video_processing_utils import BaseVideoProcessor from transformers.video_utils import group_videos_by_shape, reorder_videos @@ -225,8 +230,11 @@ class Cosmos3EdgeVideoProcessor(Qwen3VLVideoProcessor): max_pixels=size.longest_edge, ) stacked_videos = stacked_videos.view(B * T, C, H, W) - stacked_videos = self.resize( - stacked_videos, + # The target size is already computed, so bypass Qwen3-VL's + # dynamic resize, which expects an unflattened video tensor. + stacked_videos = BaseVideoProcessor.resize( + self, + image=stacked_videos, size=SizeDict(height=resized_height, width=resized_width), resample=resample, )