Hcompany

/

Holo-3.1-9B

Like 29

Follow

H company 908

The model Holo-3.1-9B generation looks kinda broken right now. Here's an exemple of the code I'm running and the generation from the model :

from transformers import AutoModelForImageTextToText, AutoProcessor
from transformers.models.qwen2_vl.image_processing_qwen2_vl import smart_resize

### Loading model
holo_model = "Hcompany/Holo-3.1-9B"
model = AutoModelForImageTextToText.from_pretrained(
    holo_model,
    # torch_dtype="auto",
    # attn_implementation="flash_attention_2",
    device_map="auto",
)
model.eval()
processor = AutoProcessor.from_pretrained(holo_model)

### Resizing image
from PIL import ImageDraw
original_image = Image.fromarray(video[0])
min_pixels = 56 * 56
max_pixels = 28 * 28 * 1280

image_processor = processor.image_processor
resized_height, resized_width = smart_resize(
    original_image.height,
    original_image.width,
    factor=image_processor.patch_size * image_processor.merge_size,
    min_pixels=min_pixels,
    max_pixels=max_pixels,
)
frame_questionned = original_image.resize(size=(resized_width, resized_height), resample=None)  # type: ignore

### Generating prompt
guidelines = "Localize an element on the GUI image according to my instructions and output a click position as Click(x, y) with x num pixels from the left edge and y num pixels from the top edge."
new_messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": frame_questionned},
        {"type": "text", "text": f"{guidelines}\nLocalize the cursor."},
    ],
}]

text = processor.apply_chat_template(
    new_messages, 
    tokenize=False, 
    add_generation_prompt=True,
    enable_thinking=True
)

inputs = processor(
    text=[text],
    images=[frame_questionned],
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

### Generating answer 
generated_ids = model.generate(
    **inputs, 
    temperature=0.6,
    max_new_tokens=512,
    do_sample=True,
    top_p=0.95,
    top_k=20,
    num_return_sequences=10
)
generated_ids_trimmed = generated_ids[:, inputs.input_ids.shape[-1]:]
res = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)

This is the output I get :

'The9\n\n=").\n\n\n"),即。\n\n、 5 works"（），或者\n\n。\n=").\n\n和ch"。\n")\n\n>").>").\n=").\n .").\n\n\n\n").\n\n=").\n。=").\n=").\n=").\n=").\n [。")\n\n\n\n),\n=").\n".\n="). ,,\n</think>\n\n\n".\n}")\n\n。\n\n\n ", ,,\n</think>。）=").\n".\n}.\n",\n",\n)".\n").\n".\n ",=").\n=").\n "。\n" " "\n", ").\n " " "\n",")\n\n。\n "。", "", " ".", "", "。 "。 " ,,\n</think>）。\n " " ".\n",完璧。\n", " " "\n", " " " "\n", " ", " ",\n".").\n " " "\n " (。"),\n".\n " "。 " " "\n ".\n ".\n " " "\n ").\n " " "\n " "\n",").",\n".")\n=").\n "。\n "。\n " " ".\n ",".\n " "。", "", " " "\n " "\n ", " "\n " "\n ").\n " " "\n " "\n " "\n " "\n ", " "\n ".\n " " "\n " "\n \\", " " "\n ", ".", " " "\n " " "\n \\", " " "\n " "\n \\",需要山体"。" " ", "", "。 "。\n " "\n " "\n " "\n ". "", " "。\n".", " " "\n " "\n \\x amministr "[。",".", " "。 "e过程和\n\n。。 "。 "". "。 "。 " " ".\n " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " " "',

The same code works fine for previous models (Holo-2) and for Holo-3.1-4B