@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 9279b9f83d
Fork 0

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, 03:15 UTC

[Bugfix] Fix max image size for LLaVA-Onevision (#11769)

Signed-off-by: Roger Wang <ywang@roblox.com>

Roger Wang committed Jan 6, 2025 at 05:48 UTC 9279b9f83dd3aa5bb3d3ce57bf92d9361755d164
1 file changed +18 -2
vllm/model_executor/models/llava_onevision.py
+18 -2
@@ -19,8 +19,8 @@ from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
19 from vllm.model_executor.sampling_metadata import SamplingMetadata
20 from vllm.multimodal import MULTIMODAL_REGISTRY
21 from vllm.multimodal.inputs import MultiModalKwargs, NestedTensors
22 -from vllm.multimodal.parse import (MultiModalDataItems, VideoEmbeddingItems,
23 - VideoProcessorItems)
22 +from vllm.multimodal.parse import (ImageSize, MultiModalDataItems,
23 + VideoEmbeddingItems, VideoProcessorItems)
24 from vllm.multimodal.processing import MultiModalFieldConfig, PromptReplacement
25 from vllm.multimodal.profiling import BaseProfilingInfo, ProcessorInputs
26 from vllm.sequence import IntermediateTensors
@@ -170,6 +170,22 @@ class LlavaOnevisionProcessingMixin(LlavaNextProcessingMixin):
170 class LlavaOnevisionProfilingInfo(LlavaOnevisionProcessingMixin,
171 BaseLlavaProfilingInfo):
172
173 + def _get_image_size_with_most_features(self) -> ImageSize:
174 + hf_config = self._get_hf_config()
175 + largest_feature_size, largest_feature_pinpoint = 0, None
176 + for (height, width) in hf_config.image_grid_pinpoints:
177 + feat_size = self._get_num_image_tokens(image_width=width,
178 + image_height=height)
179 + if feat_size > largest_feature_size:
180 + largest_feature_size = feat_size
181 + largest_feature_pinpoint = ImageSize(width=width,
182 + height=height)
183 +
184 + if largest_feature_size == 0 or largest_feature_pinpoint is None:
185 + raise ValueError("Cannot have a largest feature size of 0!")
186 +
187 + return largest_feature_pinpoint
188 +
189 def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
190 return {"image": None, "video": None}
191