llava-hf

/

llava-interleave-qwen-7b-hf

like 23

Follow

Llava Hugging Face 111

Hey! If you're doing inference w/o images, in your template conversation you should remove the {"type": "image"},. Otherwise the model will expect as many images as there were {"type": "image"} fields in the conversation

It seems that the error still occurred after removing {"type": "image"}

import requests
from PIL import Image

import torch
from transformers import AutoProcessor, LlavaForConditionalGeneration

model_id = "llava-hf/llava-interleave-qwen-7b-hf"
model = LlavaForConditionalGeneration.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
    low_cpu_mem_usage=True, 
).to(0)

processor = AutoProcessor.from_pretrained(model_id)

# Define a chat histiry and use `apply_chat_template` to get correctly formatted prompt
# Each value in "content" has to be a list of dicts with types ("text", "image") 
conversation = [
    {

      "role": "user",
      "content": [
          {"type": "text", "text": "What are these?"}
        ]
    },
]
prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)

inputs = processor(prompt, None, return_tensors='pt').to(0, torch.float16)

output = model.generate(**inputs, max_new_tokens=200, do_sample=False)
generated_ids = [output_id[len(input_id):] for input_id, output_id in zip(inputs.input_ids, output)]
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

print(generated_text)

Traceback (most recent call last):
  File "/data2/zhuang/test.py", line 33, in <module>
    inputs = processor(prompt, return_tensors='pt').to(0, torch.float16)
  File "/data2/zhuang/transformers/src/transformers/feature_extraction_utils.py", line 230, in to
    if torch.is_floating_point(v):
TypeError: is_floating_point(): argument 'input' (position 1) must be Tensor, not NoneType

Ah, I see what's the issue. Seems like Llava was't used before for text-only setting or was used only on "cpu". I will make a PR to fix it soon, imo we can support text-only use-cases similar to other VLMs in the library.

You can also pop "pixel_values" to test it before the fix is made, by adding

inputs = processor(prompt, None, return_tensors='pt')
inputs.pop("pixel_values") # it is set to None, so we should remove it
inputs = inputs.to(0) # and now we can safely move tensors to cuda, no need for bf16 as we have no floating tensors anymore

Thanks !!