This repository contains a script for training Qwen2-VL with only using HuggingFace and Liger-Kernel.
[Phi3-Vision Finetuning]
[Llama3.2-Vision Finetuning]
[Molmo Finetune]
[Pixtral Finetune]
- [2024/11/05] Add memory efficient 8-bit training.
- [2024/09/12] đ„Now the model is trained using Liger-Kernel.
- [2024/09/11] Supports setting different learning rates to projector and vision model.
- [2024/09/11] đ„Supports multi-image and video training.
- Fine-tuning Qwen2-VL
- Deepspeed
- LoRA/QLoRA
- Full-finetuning
- Enable finetuning
vision_model
while using LoRA. - Disable/enable Flash Attention 2
- Multi-image and video training
- Training optimized with liger kernel
Install the required packages using environment.yaml
.
conda env create -f environment.yaml
conda activate qwen2
pip install qwen-vl-utils
pip install flash-attn==2.5.8 --no-build-isolation
Note: You should install flash-attn after installing the other packages.
The script requires a dataset formatted according to the LLaVA specification. The dataset should be a JSON file where each entry contains information about conversations and images. Ensure that the image paths in the dataset match the provided --image_folder
.
When using a multi-image dataset, the image tokens should all be <image>
, and the image file names should have been in a list.
Please see the example below and follow format your data.
Example for single image dataset
[
{
"id": "000000033471",
"image": "000000033471.jpg",
"conversations": [
{
"from": "human",
"value": "<image>\nWhat are the colors of the bus in the image?"
},
{
"from": "gpt",
"value": "The bus in the image is white and red."
},
{
"from": "human",
"value": "What feature can be seen on the back of the bus?"
},
{
"from": "gpt",
"value": "The back of the bus features an advertisement."
},
{
"from": "human",
"value": "Is the bus driving down the street or pulled off to the side?"
},
{
"from": "gpt",
"value": "The bus is driving down the street, which is crowded with people and other vehicles."
}
]
}
...
]
Example for multi image dataset
[
{
"id": "000000033471",
"image": ["000000033471.jpg", "000000033472.jpg"],
"conversations": [
{
"from": "human",
"value": "<image>\n<image>\nIs the perspective of the camera differnt?"
},
{
"from": "gpt",
"value": "Yes, It the perspective of the camera is different."
}
]
}
...
]
Example for video dataset
[
{
"id": "sample1",
"video": "sample1.mp4",
"conversations": [
{
"from": "human",
"value": "<video>\nWhat is going on in this video?"
},
{
"from": "gpt",
"value": "A man is walking down the road."
}
]
}
...
]
Note: Qwen2-VL uses a video as a sequential of images.
Note: With the mixed-dataset (e.g. some data in a batch have images while some don't) It only supports with zero2.
To run the training script, use the following command:
bash scripts/finetune.sh
bash scripts/finetune_8bit.sh
This script will finetune the model with 8bit-adamw and fp8 model dtype. If you run out of vram, you could use this.
If you want to train only the language model with LoRA and perform full training for the vision model:
bash scripts/finetune_lora.sh
If you want to train both the language model and the vision model with LoRA:
bash scripts/finetune_lora_vision.sh
IMPORTANT: If you want to tune the embed_token
with LoRA, You need to tune lm_head
together.
Note: Freezing LLM would only work without LoRA (including vision_model LoRA).
Training arguments
--deepspeed
(str): Path to DeepSpeed config file (default: "scripts/zero2.json").--data_path
(str): Path to the LLaVA formatted training data (a JSON file). (Required)--image_folder
(str): Path to the images folder as referenced in the LLaVA formatted training data. (Required)--model_id
(str): Path to the Qwen2-VL model. (Required)--output_dir
(str): Output directory for model checkpoints--num_train_epochs
(int): Number of training epochs (default: 1).--per_device_train_batch_size
(int): Training batch size per GPU per forwarding step.--gradient_accumulation_steps
(int): Gradient accumulation steps (default: 4).--freeze_vision_tower
(bool): Option to freeze vision_model (default: False).--freeze_llm
(bool): Option to freeze LLM (default: False).--tune_merger
(bool): Option to tune projector (default: True).--num_lora_modules
(int): Number of target modules to add LoRA (-1 means all layers).--vision_lr
(float): Learning rate for vision_model.--merger_lr
(float): Learning rate for merger(projector).--learning_rate
(float): Learning rate for language module.--bf16
(bool): Option for using bfloat16.--fp16
(bool): Option for using fp16.--min_pixels
(int): Option for minimum input tokens.--max_pixles
(int): OPtion for maximum maxmimum tokens.--lora_namespan_exclude
(str): Exclude modules with namespans to add LoRA.--max_seq_length
(int): Maximum sequence length (default: 32K).--bits
(int): Quantization bits (default: 16).--disable_flash_attn2
(bool): Disable Flash Attention 2.--report_to
(str): Reporting tool (choices: 'tensorboard', 'wandb', 'none') (default: 'tensorboard').--logging_dir
(str): Logging directory (default: "./tf-logs").--lora_rank
(int): LoRA rank (default: 128).--lora_alpha
(int): LoRA alpha (default: 256).--lora_dropout
(float): LoRA dropout (default: 0.05).--logging_steps
(int): Logging steps (default: 1).--dataloader_num_workers
(int): Number of data loader workers (default: 4).
Note: The learning rate of vision_model
should be 10x ~ 5x smaller than the language_model
.
You can train the model using a video dataset. However, Qwen2-VL processes videos as a sequence of images, so youâll need to select specific frames and treat them as multiple images for training. You can set LoRA configs and use for LoRA too.
bash scripts/finetune_video.sh
Note: When training with video, it just as multi-image so you should adjust the max_pixels
for maximum resolution and fps
based on the available VRAM.
If you run out of vram, you can use zero3_offload instead of zero3. However, using zero3 is preferred.
bash scripts/merge_lora.sh
Note: Remember to replace the paths in finetune.sh
or finetune_lora.sh
with your specific paths. (Also in merge_lora.sh
when using LoRA.)
The model supprots a wide range of resolution inputs. By default, it uses the native resolution for input.
For better performance using native or higer pixel numbers are recommended, however it takes too much memory and computation time for large images. So you could adjust the pixel numbers for it.
The model splits the image into token * 28 * 28
so you could just change the the token_num part in the script.
For example:
min_pixels = 256 * 28 * 28
max_pixels = 1280 * 28 * 28
Note: For video, the you don't have to set like this, you could just set the maximum resolution for it.
Could not load library libcudnn_cnn_train.so.8. Error: /usr/local/cuda-12.1/lib/libcudnn_cnn_train.so.8: undefined symbol: _ZN5cudnn3cnn34layerNormFwd_execute_internal_implERKNS_7backend11VariantPackEP11CUstream_stRNS0_18LayerNormFwdParamsERKNS1_20NormForwardOperationEmb, version libcudnn_cnn_infer.so.8
You could run unset LD_LIBRARY_PATH
for this error.
You could see this issue
Note: You should use the merged weight when trained with LoRA.
- Install gradio
pip install gradio
- Launch app
python -m src.serve.app \
--model-path /path/to/merged/weight
You can launch gradio based demo with this command. This can also set some other generation configs like repetition_penalty
, temperature
etc.
- Support for video data
- Add demo for multi-image and video
- Support for dyanmic truncation
This project is licensed under the Apache-2.0 License. See the LICENSE file for details.
If you find this repository useful in your project, please consider giving a â and citing:
@misc{Qwen2-VL-Finetuning,
author = {Yuwon Lee},
title = {Qwen2-VL-Finetune},
year = {2024},
publisher = {GitHub},
url = {https://github.com/2U1/Qwen2-VL-Finetune}
}
This project is based on
- LLaVA-NeXT: An amazing open-source project of LMM.
- Mipha: Open-source projcet of SMM with amazing capabilites.
- Qwen2-VL-7B-Instruct: Awesome pretrained MLLM based on Qwen2.
- Liger-Kernel: Collection of Tirton kernels designed specifically for LLM training.