/text-generation-webui-macos

A macOS version of the oobabooga gradio web UI for running Large Language Models like LLaMA, llama.cpp, GPT-J, Pythia, OPT, and GALACTICA.

Primary LanguagePythonGNU Affero General Public License v3.0AGPL-3.0

OLD VERSION - 1.3.1 Patched for macOS and Apple Silicon

Patched and working with macOS and Apple Silicon M1/M2 GPU now.

GGML support is in this release, and has not been extensively tested. From the look of upstream commits, there are some changes which must be made before this will work with Llama2 models.

If you want the most recent version, from the oobabooga reposiotry, go here: oobabooga/text-generation-webgui

Otherwise, use these instructions I have on putting together the macOS Python environment. These instructions are not only useful for setting up oobabooga, but also for anyone working in data analytics, machine learning, deep learning, scientific computing, and other areas that can benefit from an optimized Python GPU environment on Apple Silicon.

I will be updating this README file with new information specifically regarding macOS and Apple Silicon.

I would like to work closely with the oobaboogs team and try to implement simkilar solutions so the web UI can have a similar look and feel.

Maintaining and improving support for macOS and Apple Silicon in this project has required significant research, debugging, and development effort. If you find my contributions helpful and want to show your appreciation, you can Buy Me a Coffee, sponsor this project, or consider me for job opportunities.

While the focus of this branch is to enhance macOS and Apple Silicon support, I aim to maintain compatibility with Linux and POSIX operating systems. Contributions and feedback related to Linux compatibility are always welcome.

Anyone who would like to assist with supporting Apple Silicon, let me know. There is much to do and I can only do so much by myself.

Features

  • 3 interface modes: default, notebook, and chat
  • Multiple model backends: tranformers, llama.cpp, AutoGPTQ, GPTQ-for-LLaMa, ExLlama, RWKV, FlexGen
  • Dropdown menu for quickly switching between different models
  • LoRA: load and unload LoRAs on the fly, load multiple LoRAs at the same time, train a new LoRA
  • Precise instruction templates for chat mode, 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
  • 8-bit and 4-bit inference through bitsandbytes CPU only mode for macOS, bitsandbytes does not support Apple Silicon M1/M2 processors
  • CPU mode for transformers models
  • DeepSpeed ZeRO-3 inference
  • Extensions
  • Custom chat characters
  • Very efficient text streaming
  • Markdown output with LaTeX rendering, to use for instance with GALACTICA
  • Nice HTML output for GPT-4chan
  • API, including endpoints for websocket streaming (see the examples)

To learn how to use the various features, check out the Documentation: https://github.com/unixwzrd/text-generation-webui/tree/main/docs

Installation

Currently only manual installation until an installer can be put together.

Downloading models

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

To download a protected model, set env vars HF_USER and HF_PASS to your Hugging Face username and password (or User Access Token). The model's terms must first be accepted on the HF website.

NOTE: You may want to create symbolic links in the model directory since sometimes there are multiple models downloaded with different quantizations, or on external storage due to space on th edisk where you insatlled this. If you have external storge, like I do, you can create a link by going to the models directory and creating the link to your model.

GGML models

You can drop these directly into the models/ folder, making sure that the file name contains ggml somewhere and ends in .bin.

GPT-4chan

Instructions

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:

  1. Place the files under models/gpt4chan_model_float16 or models/gpt4chan_model.
  2. Place GPT-J 6B's config.json file in that same folder: config.json.
  3. 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

When you load this model in default or notebook modes, the "HTML" tab will show the generated text in 4chan format.

Starting the web UI

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:

Basic settings

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.
--multi-user Multi-user mode. Chat histories are not saved or automatically loaded. WARNING: this is highly experimental.
--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 yaml file. See settings-template.yaml for an example. If you create a file called settings.yaml, 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.

Model loader

Flag Description
--loader LOADER Choose the model loader manually, otherwise, it will get autodetected. Valid options: transformers, autogptq, gptq-for-llama, exllama, exllama_hf, llamacpp, rwkv, flexgen

Accelerate/transformers

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 ...] Maximum 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 (using bitsandbytes).
--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 and Falcon.

Accelerate 4-bit

⚠️ Requires minimum compute of 7.0 on Windows at the moment.

Flag Description
--load-in-4bit Load the model with 4-bit precision (using bitsandbytes).
--compute_dtype COMPUTE_DTYPE compute dtype for 4-bit. Valid options: bfloat16, float16, float32.
--quant_type QUANT_TYPE quant_type for 4-bit. Valid options: nf4, fp4.
--use_double_quant use_double_quant for 4-bit.

llama.cpp

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.
--n_ctx N_CTX Size of the prompt context.
--llama_cpp_seed SEED Seed for llama-cpp models. Default 0 (random).
--n_gqa N_GQA grouped-query attention. Must be 8 for llama2 70b.
--rms_norm_eps RMS_NORM_EPS Must be 1e-5 for llama2 70b.

AutoGPTQ

Flag Description
--triton Use triton.
--no_inject_fused_attention Disable the use of fused attention, which will use less VRAM at the cost of slower inference.
--no_inject_fused_mlp Triton mode only: disable the use of fused MLP, which will use less VRAM at the cost of slower inference.
--no_use_cuda_fp16 This can make models faster on some systems.
--desc_act For models that don't have a quantize_config.json, this parameter is used to define whether to set desc_act or not in BaseQuantizeConfig.

ExLlama

Flag Description
--gpu-split Comma-separated list of VRAM (in GB) to use per GPU device for model layers, e.g. 20,7,7
--max_seq_len MAX_SEQ_LEN Maximum sequence length.
--compress_pos_emb COMPRESS_POS_EMB Positional embeddings compression factor. Should typically be set to max_seq_len / 2048.
--alpha_value ALPHA_VALUE Positional embeddings alpha factor for NTK RoPE scaling. Same as above. Use either this or compress_pos_emb, not both. `

GPTQ-for-LLaMa

Flag Description
--wbits WBITS Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported.
--model_type MODEL_TYPE Model type of pre-quantized model. Currently LLaMA, OPT, and GPT-J are supported.
--groupsize GROUPSIZE 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.
--monkey-patch Apply the monkey patch for using LoRAs with quantized models.
--quant_attn (triton) Enable quant attention.
--warmup_autotune (triton) Enable warmup autotune.
--fused_mlp (triton) Enable fused mlp.

FlexGen

Flag Description
--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%).

DeepSpeed

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.

RWKV

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.

Gradio

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 USER:PWD set gradio authentication like "username:password"; or comma-delimit multiple like "u1:p1,u2:p2,u3:p3"
--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"

API

Flag Description
--api Enable the API extension.
--public-api Create a public URL for the API using Cloudfare.
--api-blocking-port BLOCKING_PORT The listening port for the blocking API.
--api-streaming-port STREAMING_PORT The listening port for the streaming API.

Multimodal

Flag Description
--multimodal-pipeline PIPELINE The multimodal pipeline to use. Examples: llava-7b, llava-13b.

Out of memory errors? Check the low VRAM guide.

Presets

Inference settings presets can be created under presets/ as yaml files. These files are detected automatically at startup.

The presets that are included by default are the result of a contest that received 7215 votes. More details can be found here.

Contributing

  • Pull requests, suggestions, and issue reports are welcome.
  • Make sure to carefully search existing issues before starting a new one.
  • If you have some experience with git, testing an open pull request and leaving a comment on whether it works as expected or not is immensely helpful.
  • A simple way to contribute, even if you are not a programmer, is to leave a 👍 on an issue or pull request that you find relevant.

Community

I will be checking in on the oobabooga Discord server at the #mac-setup channel.

Credits