MCPcopy Create free account
hub / github.com/NVIDIA/TensorRT-LLM / LlavaOnevisionVisionWrapper

Class LlavaOnevisionVisionWrapper

tensorrt_llm/tools/multimodal_builder.py:574–588  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

572 args.device, torch.float16)[0]
573
574 class LlavaOnevisionVisionWrapper(torch.nn.Module):
575
576 def __init__(self, vision_tower, projector, config):
577 super().__init__()
578 self.vision_tower = vision_tower
579 self.projector = projector
580 self.config = config
581
582 def forward(self, pixel_values):
583 image_features = self.vision_tower(pixel_values,
584 output_hidden_states=True)
585 selected_image_feature = image_features.hidden_states[
586 self.config.vision_feature_layer]
587 image_features = self.projector(selected_image_feature)
588 return image_features # (sigma(bs, patches_i), 729, c)
589
590 model = LlavaOnevisionForConditionalGeneration.from_pretrained(
591 args.model_path, dtype=torch.float16)

Callers 1

build_llava_engineFunction · 0.85

Calls

no outgoing calls

Tested by

no test coverage detected