- Allow 4bit loras and use of the --autograd implementation.
- Use GPT-J 4-bits (GPTQv1/v2)
- GPT-NeoXT 4-bits (GPTQv1/v2)
- 8 bit threshold slider, default 1.5 (pre compute 7.0)
- load 4-bit lora from web ui
- V1 Models work in --autograd (declare with --v1)
- V2 Models work in both.
- Offloading works in autograd with --gpu-memory but doesn't 100% hodl while generating
- Offloading for llama GPTQ works
- Offloading for all other GPTQ models getting worked on
- Probably not as fast as the old version (still figuring this out)
- 4bit loras only work in autograd
- Only load one 4bit lora at a time and apply with no loras before switch.
- Train 4-bit loras
- AutoGPTQ (https://github.com/PanQiWei/AutoGPTQ) < Finally Merged
https://github.com/Ph0rk0z/GPTQ-Merged (dual module branch)
https://github.com/sterlind/peft (now auto patches)
5/17/23
Update submodules, supporting a new method of splitting that makes 65b possible over 2,
even janky cards at higher speed. No more OOM on 65b at full context.
5/8/23
I think autograd problem is fixed.. equal or faster than GPTQ
Update the submodules git submodule update --recursive --remote
4/22/23
New --mlp-attn, slightly faster on some contexts but no lora support added yet.
both --xformers and --sdp-attention prevent the 30b from going OOM at full context.
4/18/23
Using the patch for PEFT and no longer depends on PEFT fork.
Makes it easier to run main branch side by side.
Rewrote the GPTQ loader as well to be more compact.
You may have to update tokenizers agian and install colorama from pip.
4/11/23
Update to new PEFT version
https://github.com/sterlind/peft
4/10/23
pip install deepspeed -U
pip install xmformers
Xformers install will upgrade torch to 2.0
YOU WILL HAVE TO RECOMPILE YOUR CUDA KERNELS!!
4/8/23 - Update transformers!
pip install tokenizers==0.13.1
pip install protobuf==3.20.0
pip install git+https://github.com/huggingface/transformers
Repos are linked as submodules.. you may have to update them: https://stackoverflow.com/a/1032653
git submodule update --remote
- 13b and 30b llama response times for me become usable with a lora or not.
- Changes aren't so clean to be accepted as a p/r
- Clone and re-use your oobabooga/text-generation-webui conda environment.
- Build GPTQ kernel with python setup.py install after cloing into repositories/
- Also build and install patched PEFT.
- I don't know, can't use it. Try WSL
python server.py --model llama-30b --chat --autograd --wbits 4
python server.py --model opt-13b --chat --autograd --wbits 4 --lora opt-13b-lora-1.0ep
python server.py --model oasst-sft-1-pythia-12b --chat --autograd --wbits 4 --model_type gptneox
python server.py --model oasst-sft-1-pythia-12b --chat --autograd --wbits 4 --model_type gptneox --v1
python server.py --model llama-7b-4bit-128g --chat --groupsize 128 --wbits 4 --model_type llama
python server.py --model llama-30b-4bit-128g --chat --autograd --groupsize 128 --wbits 4 --model_type llama
- Dropdown menu for switching between models
- Notebook mode that resembles OpenAI's playground
- Chat mode for conversation and role-playing
- Instruct mode compatible with various formats, including Alpaca, Vicuna, Open Assistant, Dolly, Koala, ChatGLM, MOSS, RWKV-Raven, Galactica, StableLM, WizardLM, Baize, Ziya, Chinese-Vicuna, MPT, INCITE, Wizard Mega, KoAlpaca, Vigogne, Bactrian, h2o, and OpenBuddy
- Multimodal pipelines, including LLaVA and MiniGPT-4
- Markdown output for GALACTICA, including LaTeX rendering
- Nice HTML output for GPT-4chan
- Custom chat characters
- Advanced chat features (send images, get audio responses with TTS)
- Very efficient text streaming
- Parameter presets
- LLaMA model
- 4-bit GPTQ mode
- LoRA (loading and training)
- llama.cpp
- RWKV model
- 8-bit mode
- Layers splitting across GPU(s), CPU, and disk
- CPU mode
- FlexGen
- DeepSpeed ZeRO-3
- API with streaming and without streaming
- Extensions - see the user extensions list
Windows | Linux | macOS |
---|---|---|
oobabooga-windows.zip | oobabooga-linux.zip | oobabooga-macos.zip |
Just download the zip above, extract it, and double-click on "start". The web UI and all its dependencies will be installed in the same folder.
- The source codes are here: https://github.com/oobabooga/one-click-installers
- There is no need to run the installers as admin.
- AMD doesn't work on Windows.
- Huge thanks to @jllllll, @ClayShoaf, and @xNul for their contributions to these installers.
Recommended if you have some experience with the command line.
On Windows, I additionally recommend carrying out the installation on WSL instead of the base system: WSL installation guide.
https://docs.conda.io/en/latest/miniconda.html
On Linux or WSL, it can be automatically installed with these two commands:
curl -sL "https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh" > "Miniconda3.sh"
bash Miniconda3.sh
Source: https://educe-ubc.github.io/conda.html
41
42
43
[[Try it on Google Colab]](https://colab.research.google.com/github/oobabooga/AI-Notebooks/blob/main/Colab-TextGen-GPU.ipynb)
44
conda create -n textgen python=3.10.9
conda activate textgen
System | GPU | Command |
---|---|---|
Linux/WSL | NVIDIA | pip3 install torch torchvision torchaudio |
Linux | AMD | pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2 |
MacOS + MPS (untested) | Any | pip3 install torch torchvision torchaudio |
The up-to-date commands can be found here: https://pytorch.org/get-started/locally/.
- MacOS users: oobabooga#393
- AMD users: https://rentry.org/eq3hg
git clone https://github.com/oobabooga/text-generation-webui
cd text-generation-webui
pip install -r requirements.txt
The base installation covers transformers models (AutoModelForCausalLM
and AutoModelForSeq2SeqLM
specifically) and llama.cpp (GGML) models.
To use 4-bit GPU models, the additional installation steps below are necessary:
As an alternative to the recommended WSL method, you can install the web UI natively on Windows using this guide. It will be a lot harder and the performance may be slower: Windows installation guide.
ln -s docker/{Dockerfile,docker-compose.yml,.dockerignore} .
cp docker/.env.example .env
# Edit .env and set TORCH_CUDA_ARCH_LIST based on your GPU model
docker compose up --build
You need to have docker compose v2.17 or higher installed in your system. To see how to install docker compose itself, see the guide in here.
Contributed by @loeken in #633
From time to time, the requirements.txt
changes. To update, use this command:
conda activate textgen
cd text-generation-webui
pip install -r requirements.txt --upgrade
Models should be placed inside the models/
folder.
Hugging Face is the main place to download models. These are some examples:
You can automatically download a model from HF using the script download-model.py
:
python download-model.py organization/model
For example:
python download-model.py facebook/opt-1.3b
If you want to download a model manually, note that all you need are the json, txt, and pytorch*.bin (or model*.safetensors) files. The remaining files are not necessary.
You can drop these directly into the models/
folder, making sure that the file name contains ggml
somewhere and ends in .bin
.
GPT-4chan has been shut down from Hugging Face, so you need to download it elsewhere. You have two options:
The 32-bit version is only relevant if you intend to run the model in CPU mode. Otherwise, you should use the 16-bit version.
After downloading the model, follow these steps:
- Place the files under
models/gpt4chan_model_float16
ormodels/gpt4chan_model
. - Place GPT-J 6B's config.json file in that same folder: config.json.
- Download GPT-J 6B's tokenizer files (they will be automatically detected when you attempt to load GPT-4chan):
python download-model.py EleutherAI/gpt-j-6B --text-only
conda activate textgen
cd text-generation-webui
python server.py
Then browse to
http://localhost:7860/?__theme=dark
Optionally, you can use the following command-line flags:
Flag | Description |
---|---|
-h , --help |
Show this help message and exit. |
--notebook |
Launch the web UI in notebook mode, where the output is written to the same text box as the input. |
--chat |
Launch the web UI in chat mode. |
--character CHARACTER |
The name of the character to load in chat mode by default. |
--model MODEL |
Name of the model to load by default. |
--lora LORA [LORA ...] |
The list of LoRAs to load. If you want to load more than one LoRA, write the names separated by spaces. |
--model-dir MODEL_DIR |
Path to directory with all the models. |
--lora-dir LORA_DIR |
Path to directory with all the loras. |
--model-menu |
Show a model menu in the terminal when the web UI is first launched. |
--no-stream |
Don't stream the text output in real time. |
--settings SETTINGS_FILE |
Load the default interface settings from this json file. See settings-template.json for an example. If you create a file called settings.json , this file will be loaded by default without the need to use the --settings flag. |
--extensions EXTENSIONS [EXTENSIONS ...] |
The list of extensions to load. If you want to load more than one extension, write the names separated by spaces. |
--verbose |
Print the prompts to the terminal. |
Flag | Description |
---|---|
--cpu |
Use the CPU to generate text. Warning: Training on CPU is extremely slow. |
--auto-devices |
Automatically split the model across the available GPU(s) and CPU. |
--gpu-memory GPU_MEMORY [GPU_MEMORY ...] |
Maxmimum GPU memory in GiB to be allocated per GPU. Example: --gpu-memory 10 for a single GPU, --gpu-memory 10 5 for two GPUs. You can also set values in MiB like --gpu-memory 3500MiB . |
--cpu-memory CPU_MEMORY |
Maximum CPU memory in GiB to allocate for offloaded weights. Same as above. |
--disk |
If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk. |
--disk-cache-dir DISK_CACHE_DIR |
Directory to save the disk cache to. Defaults to cache/ . |
--load-in-8bit |
Load the model with 8-bit precision. |
--threshold |
Threshold for 8bit precision for older cards. It will use more memory while performing infrerence so watch out. NaN == too high. OOM == too low. |
--bf16 |
Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU. |
--no-cache |
Set use_cache to False while generating text. This reduces the VRAM usage a bit with a performance cost. |
--xformers |
Use xformer's memory efficient attention. This should increase your tokens/s. |
--sdp-attention |
Use torch 2.0's sdp attention. |
--trust-remote-code |
Set trust_remote_code=True while loading a model. Necessary for ChatGLM. |
Flag | Description |
---|---|
--threads |
Number of threads to use. |
--n_batch |
Maximum number of prompt tokens to batch together when calling llama_eval. |
--no-mmap |
Prevent mmap from being used. |
--mlock |
Force the system to keep the model in RAM. |
--cache-capacity CACHE_CAPACITY |
Maximum cache capacity. Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed. |
--n-gpu-layers N_GPU_LAYERS |
Number of layers to offload to the GPU. Only works if llama-cpp-python was compiled with BLAS. Set this to 1000000000 to offload all layers to the GPU. |
Flag | Description |
---|---|
--wbits WBITS |
GPTQ: Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported. |
--model_type MODEL_TYPE |
GPTQ: Model type of pre-quantized model. Currently LLaMA, OPT, and GPT-J are supported. |
--groupsize GROUPSIZE |
GPTQ: Group size. |
--pre_layer PRE_LAYER [PRE_LAYER ...] |
The number of layers to allocate to the GPU. Setting this parameter enables CPU offloading for 4-bit models. For multi-gpu, write the numbers separated by spaces, eg --pre_layer 30 60 . |
--checkpoint CHECKPOINT |
The path to the quantized checkpoint file. If not specified, it will be automatically detected. |
--autograd |
GPTQ: Autograd implementation to use 4bit lora and run multiple models |
--v1 |
GPTQ: Explicitly declare a GPTQv1 model to load into autograd. |
---mlp_attn |
MLP attention hijack. Slightly faster inference. |
--quant_attn |
(triton) Enable quant attention. |
--warmup_autotune |
(triton) Enable warmup autotune. |
--fused_mlp |
(triton) Enable fused mlp. |
--autogptq |
Load with autogptq. Look in shared.py for more options like triton or using act order w/ groupsize kernel |
Flag | Description |
---|---|
--flexgen |
Enable the use of FlexGen offloading. |
--percent PERCENT [PERCENT ...] |
FlexGen: allocation percentages. Must be 6 numbers separated by spaces (default: 0, 100, 100, 0, 100, 0). |
--compress-weight |
FlexGen: Whether to compress weight (default: False). |
--pin-weight [PIN_WEIGHT] |
FlexGen: whether to pin weights (setting this to False reduces CPU memory by 20%). |
Flag | Description |
---|---|
--deepspeed |
Enable the use of DeepSpeed ZeRO-3 for inference via the Transformers integration. |
--nvme-offload-dir NVME_OFFLOAD_DIR |
DeepSpeed: Directory to use for ZeRO-3 NVME offloading. |
--local_rank LOCAL_RANK |
DeepSpeed: Optional argument for distributed setups. |
Flag | Description |
---|---|
--rwkv-strategy RWKV_STRATEGY |
RWKV: The strategy to use while loading the model. Examples: "cpu fp32", "cuda fp16", "cuda fp16i8". |
--rwkv-cuda-on |
RWKV: Compile the CUDA kernel for better performance. |
Flag | Description |
---|---|
--listen |
Make the web UI reachable from your local network. |
--listen-host LISTEN_HOST |
The hostname that the server will use. |
--listen-port LISTEN_PORT |
The listening port that the server will use. |
--share |
Create a public URL. This is useful for running the web UI on Google Colab or similar. |
--auto-launch |
Open the web UI in the default browser upon launch. |
--gradio-auth-path GRADIO_AUTH_PATH |
Set the gradio authentication file path. The file should contain one or more user:password pairs in this format: "u1:p1,u2:p2,u3:p3" |
Flag | Description |
---|---|
--api |
Enable the API extension. |
--public-api |
Create a public URL for the API using Cloudfare. |
Flag | Description |
---|---|
--multimodal-pipeline PIPELINE |
The multimodal pipeline to use. Examples: llava-7b , llava-13b . |
Out of memory errors? Check the low VRAM guide.
Inference settings presets can be created under presets/
as text files. These files are detected automatically at startup.
By default, 10 presets by NovelAI and KoboldAI are included. These were selected out of a sample of 43 presets after applying a K-Means clustering algorithm and selecting the elements closest to the average of each cluster.
Make sure to check out the documentation for an in-depth guide on how to use the web UI.
https://github.com/oobabooga/text-generation-webui/tree/main/docs
Pull requests, suggestions, and issue reports are welcome.
You are also welcome to review open pull requests.
Before reporting a bug, make sure that you have:
- Created a conda environment and installed the dependencies exactly as in the Installation section above.
- Searched to see if an issue already exists for the issue you encountered.
- Gradio dropdown menu refresh button, code for reloading the interface: https://github.com/AUTOMATIC1111/stable-diffusion-webui
- Verbose preset: Anonymous 4chan user.
- NovelAI and KoboldAI presets: https://github.com/KoboldAI/KoboldAI-Client/wiki/Settings-Presets
- Code for early stopping in chat mode, code for some of the sliders: https://github.com/PygmalionAI/gradio-ui/