Alpaca Lora 4bit

Made some adjust for the code in peft and gptq for llama, and make it possible for lora finetuning with a 4 bits base model. The same adjustment can be made for 2, 3 and 8 bits.

For those who want to use pip installable version:

pip install git+https://github.com/johnsmith0031/alpaca_lora_4bit@winglian-setup_pip

Model Server

Better inference performance with text_generation_webui, about 40% faster

Simple expriment results:
7b model with groupsize=128 no act-order
improved from 13 tokens/sec to 20 tokens/sec

Step:

run model server process
run webui process with monkey patch

Example

run_server.sh

#!/bin/bash

export PYTHONPATH=$PYTHONPATH:./

CONFIG_PATH=
MODEL_PATH=
LORA_PATH=

VENV_PATH=
source $VENV_PATH/bin/activate
python ./scripts/run_server.py --config_path $CONFIG_PATH --model_path $MODEL_PATH --lora_path $LORA_PATH --groupsize=128 --quant_attn --port 5555 --pub_port 5556

run_webui.sh

#!/bin/bash

if [ -f "server2.py" ]; then
    rm server2.py
fi
echo "import custom_model_server_monkey_patch" > server2.py
cat server.py >> server2.py

export PYTHONPATH=$PYTHONPATH:../

VENV_PATH=
source $VENV_PATH/bin/activate
python server2.py --chat --listen

Note:

quant_attn only support torch 2.0+
lora support is only for simple lora with only q_proj and v_proj
this patch breaks model selection, lora selection and training feature in webui

Docker

Quick start for running the chat UI

git clone https://github.com/johnsmith0031/alpaca_lora_4bit.git
cd alpaca_lora_4bit
DOCKER_BUILDKIT=1 docker build -t alpaca_lora_4bit . # build step can take 12 min
docker run --gpus=all -p 7860:7860 alpaca_lora_4bit

Point your browser to http://localhost:7860

Results

It's fast on a 3070 Ti mobile. Uses 5-6 GB of GPU RAM.

Development

Install Manual by s4rduk4r: https://github.com/s4rduk4r/alpaca_lora_4bit_readme/blob/main/README.md
Also Remember to create a venv if you do not want the packages be overwritten.

Update Logs

Resolved numerically unstable issue
Reconstruct fp16 matrix from 4bit data and call torch.matmul largely increased the inference speed.
Added install script for windows and linux.
Added Gradient Checkpointing. Now It can finetune 30b model 4bit on a single GPU with 24G VRAM with Gradient Checkpointing enabled. (finetune.py updated) (but would reduce training speed, so if having enough VRAM this option is not needed)
Added install manual by s4rduk4r
Added pip install support by sterlind, preparing to merge changes upstream
Added V2 model support (with groupsize, both inference + finetune)
Added some options on finetune: set default to use eos_token instead of padding, add resume_checkpoint to continue training
Added offload support. load_llama_model_4bit_low_ram_and_offload_to_cpu function can be used.
Added monkey patch for text generation webui for fixing initial eos token issue.
Added Flash attention support. (Use --flash-attention)
Added Triton backend to support model using groupsize and act-order. (Use --backend=triton)
Added g_idx support in cuda backend (need recompile cuda kernel)
Added xformers support
Removed triton, flash-atten from requirements.txt for compatibility
Removed bitsandbytes from requirements
Added pip installable branch based on winglian's PR
Added cuda backend quant attention and fused mlp from GPTQ_For_Llama.
Added lora patch for GPTQ_For_Llama repo triton backend.
Usage:

from monkeypatch.gptq_for_llala_lora_monkey_patch import inject_lora_layers
inject_lora_layers(model, lora_path, device, dtype)

Added Model server for better inference performance with webui (40% faster than original webui which runs model and gradio in same process)

Requirements

gptq-for-llama
peft
The specific version is inside requirements.txt

Install

pip install -r requirements.txt

Finetune

After installation, this script can be used. Use --v1 flag for v1 model.

python finetune.py ./data.txt \
    --ds_type=txt \
    --lora_out_dir=./test/ \
    --llama_q4_config_dir=./llama-7b-4bit/ \
    --llama_q4_model=./llama-7b-4bit.pt \
    --mbatch_size=1 \
    --batch_size=2 \
    --epochs=3 \
    --lr=3e-4 \
    --cutoff_len=256 \
    --lora_r=8 \
    --lora_alpha=16 \
    --lora_dropout=0.05 \
    --warmup_steps=5 \
    --save_steps=50 \
    --save_total_limit=3 \
    --logging_steps=5 \
    --groupsize=-1 \
    --v1 \
    --xformers \
    --backend=cuda

Inference

After installation, this script can be used:

python inference.py

Text Generation Webui Monkey Patch

Clone the latest version of text generation webui and copy all the files into ./text-generation-webui/

git clone https://github.com/oobabooga/text-generation-webui.git

Open server.py and insert a line at the beginning

import custom_monkey_patch # apply monkey patch
...

Use the command to run

python server.py

monkey patch inside webui

Currently the webui support using this repo by the monkeypatch inside it.
You can simply clone this repo to ./repositories/ in the path of text generation webui.

Flash Attention

It seems that we can apply a monkey patch for llama model. To use it, simply download the file from MonkeyPatch. And also, flash-attention is needed, and currently do not support pytorch 2.0. Just add --flash-attention to use it for finetuning.

Xformers

Install

pip install xformers

Usage

from monkeypatch.llama_attn_hijack_xformers import hijack_llama_attention
hijack_llama_attention()

Quant Attention and MLP Patch

Note: Currently does not support peft lora, but can use inject_lora_layers to load simple lora with only q_proj and v_proj.

Usage:

from model_attn_mlp_patch import make_quant_attn, make_fused_mlp, inject_lora_layers
make_quant_attn(model)
make_fused_mlp(model)

# Lora
inject_lora_layers(model, lora_path)

tiro2000/alpaca_lora_4bit