[ICLR 2024 & ECCV 2024] The All-Seeing Projects: Towards Panoptic Visual Recognition&Understanding and General Relation Comprehension of the Open World"
The GPU memory consumption of the model was too high, so I converted it to a LLAMA.CPP file. The GPU memory usage is fine.
However, due to the nature of the model converted to llama.cpp in the model inference step, we need to convert the input parameter format. If there are any llama.cpp experts, we would appreciate it if you could tell us how to convert it.
# all-seeing/all-seeing-v2/llava/eval/model_vqa_loader_vocab_rank.py line 156 :
# model ( == ASMv2.gguf )
# Below are the source code locations that need to be converted
with torch.inference_mode():
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
images=image_tensor.to(dtype=torch.float16, device=args.device, non_blocking=True),
).logits
Using the example provided by llama.cpp, I'm done verifying that the model performs the behavior I want.
However, I'm not going to change the issue to solved in case you have more official answers.
The GPU memory consumption of the model was too high, so I converted it to a LLAMA.CPP file. The GPU memory usage is fine. However, due to the nature of the model converted to llama.cpp in the model inference step, we need to convert the input parameter format. If there are any llama.cpp experts, we would appreciate it if you could tell us how to convert it.