PanoVLM-500M

PanoVLM is a linear-attention vision-language model: a FastViT-HD vision encoder (timm) feeding a PanoLM linear-attention causal LM through a lightweight projector.

• Type: Vision-Language (image-text-to-text) Model
• LM: PanoLM-380M
• Vision encoder: FastViT-HD (timm), NCHW input
• Projector: BitLinear (encoder dim → LM dim)
• Default image resolution: 1024×1024 (pad-resized; only the resolution is meant to be changed)

Parameters

Component	Parameters
PanoLM LM	~387 M
FastViT-HD vision encoder	~123 M
Projector	~3 M
Total	~513 M

Requirements

torch==2.12.0
transformers==5.8.1
flash-linear-attention==0.5.0
timm==1.0.25

Usage

Replace <repo_id> with the HF Hub identifier.

from transformers import AutoModelForImageTextToText, AutoProcessor
from PIL import Image
import requests

repo_id = "<repo_id>"
model = AutoModelForImageTextToText.from_pretrained(
    repo_id, trust_remote_code=True,
).cuda()  # fla's RMSNorm uses Triton kernels that only run on CUDA tensors.
processor = AutoProcessor.from_pretrained(repo_id, trust_remote_code=True)

url = "https://llava-vl.github.io/static/images/view.jpg"
image = Image.open(requests.get(url, stream=True).raw)

# PanoVLM's chat template wraps string content, so put the <|image|> placeholder
# inline in the message text (the processor expands it into the image tokens).
# Keep the space after <|image|>: the HF tokenizer, unlike the training tokenizer,
# does not implicitly insert one at the special-token boundary.
messages = [{"role": "user", "content": "<|image|> Is there a boat in the image?"}]
prompt = processor.tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=False,
)
inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)

out = model.generate(**inputs, max_new_tokens=512)
print(processor.decode(out[0], skip_special_tokens=True))