PaddleOCR-VL-1.6 Vietnamese Merged

Full BF16 model produced by merging the Vietnamese LoRA adapter into PaddlePaddle/PaddleOCR-VL-1.6. This repository is standalone; the LoRA adapter is not required at inference time.

Use the exact user prompt OCR: and preserve the included native multimodal chat template.

Loading and inference

The verified environment used Python 3.12, PyTorch 2.11.0+cu128, Transformers 4.55.4, and a CUDA GPU.

import os

os.environ.setdefault("CUBLAS_WORKSPACE_CONFIG", ":4096:8")

import torch
from PIL import Image
from transformers import AutoModelForCausalLM, AutoProcessor

model_id = "tieubaoca/PaddleOCR-VL-1.6-Vietnamese-Merged"
processor = AutoProcessor.from_pretrained(
    model_id,
    trust_remote_code=True,
    use_fast=False,
)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    attn_implementation="eager",
).to("cuda").eval()

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "document.png"},
            {"type": "text", "text": "OCR:"},
        ],
    }
]
prompt = processor.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
with Image.open("document.png") as source:
    image = source.convert("RGB").copy()
inputs = processor(text=[prompt], images=[image], return_tensors="pt")
inputs = {name: value.to("cuda") for name, value in inputs.items()}
prompt_length = inputs["input_ids"].shape[1]

with torch.inference_mode():
    generated_ids = model.generate(
        **inputs,
        do_sample=False,
        num_beams=1,
        max_new_tokens=256,
        use_cache=True,
    )
text = processor.batch_decode(
    generated_ids[:, prompt_length:],
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False,
)[0].strip()
print(text)

Evaluation

Deterministic greedy decoding with the native processor and chat template, 32 fixed validation samples from each of five sources, and 256 maximum new tokens.

Merged per-source metrics:

Merge verification

• All 126 LoRA-targeted tensors were merged.
• The exported weights matched the in-memory merged model after BF16 output casting within rtol=0.0005 and atol=0.0005.
• Exported and in-memory merged logits were identical on the verification fixture (max_abs_error=0.0, equal argmax).

The original PaddleOCR-VL-1.6 license is included as LICENSE.