BS1-MP1M-NOVISION-LR3e4-TRAINING-V2-VLSP-EP2

qwen3vl-8b-lora

This is a LoRA adapter fine-tuned on top of Qwen/Qwen3-VL-8B-Instruct.

Model Description

This model is a fine-tuned version of Qwen3-VL-8B-Instruct using LoRA (Low-Rank Adaptation) for efficient training. The adapter weights can be merged with the base model for inference.

Training Details

Base Model

• Model: Qwen/Qwen3-VL-8B-Instruct
• Architecture: Vision-Language Model (VLM)

LoRA Configuration

• Rank (r): 64
• Alpha: 128
• Dropout: 0.05
• Target Modules: q_proj, k_proj, v_proj, o_proj
• Task Type: Causal Language Modeling

Training Hyperparameters

• Learning Rate: 1e-5
• Batch Size: 4 (per device)
• Gradient Accumulation Steps: 4
• Epochs: 2
• Optimizer: AdamW
• Weight Decay: 0
• Warmup Ratio: 0.03
• LR Scheduler: Cosine
• Max Gradient Norm: 1.0
• Model Max Length: 40960
• Max Pixels: 250880
• Min Pixels: 784

Training Infrastructure

• Framework: PyTorch + DeepSpeed (ZeRO Stage 2)
• Precision: BF16
• Gradient Checkpointing: Enabled

Usage

Requirements

pip install transformers peft torch pillow qwen-vl-utils

Loading the Model

from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from peft import PeftModel
import torch

# Load base model
base_model = Qwen2VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen3-VL-8B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(
    base_model,
    "openhay/qwen3vl-8b-lora",
    torch_dtype=torch.bfloat16
)

# Load processor
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-8B-Instruct")

Inference Example

from qwen_vl_utils import process_vision_info
from PIL import Image

# Prepare messages
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "path/to/image.jpg"},
            {"type": "text", "text": "Describe this image in detail."},
        ],
    }
]

# Prepare for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Generate
with torch.no_grad():
    generated_ids = model.generate(**inputs, max_new_tokens=512)
    
generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, 
    skip_special_tokens=True, 
    clean_up_tokenization_spaces=False
)

print(output_text[0])

Merging LoRA Weights (Optional)

If you want to merge the LoRA weights into the base model for faster inference:

from transformers import Qwen2VLForConditionalGeneration
from peft import PeftModel

# Load base model and adapter
base_model = Qwen2VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen3-VL-8B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "openhay/qwen3vl-8b-lora")

# Merge and save
merged_model = model.merge_and_unload()
merged_model.save_pretrained("./merged_model")

Limitations

• This model inherits all limitations from the base Qwen3-VL-8B-Instruct model
• Performance depends on the quality and domain of the fine-tuning dataset
• LoRA adapters may not capture all nuances that full fine-tuning would achieve

Citation

If you use this model, please cite:

@misc{qwen3vl_8b_lora,
  author = {OpenHay},
  title = {qwen3vl-8b-lora},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/openhay/qwen3vl-8b-lora}}
}

Acknowledgements

• Base model: Qwen3-VL-8B-Instruct by Alibaba Cloud
• Training framework: LLaMA-Factory or similar
• LoRA implementation: PEFT by Hugging Face