* Large number of extensions (built-in and user-contributed), including Coqui TTS for realistic voice outputs, Whisper STT for voice inputs, translation, [multimodal pipelines](https://github.com/oobabooga/text-generation-webui/tree/main/extensions/multimodal), vector databases, Stable Diffusion integration, and a lot more. See [the wiki](https://github.com/oobabooga/text-generation-webui/wiki/07-%E2%80%90-Extensions) and [the extensions directory](https://github.com/oobabooga/text-generation-webui-extensions) for details.
* LoRA: train new LoRAs with your own data, load/unload LoRAs on the fly for generation.
* Transformers library integration: load models in 4-bit or 8-bit precision through bitsandbytes, use llama.cpp with transformers samplers (`llamacpp_HF` loader), CPU inference in 32-bit precision using PyTorch.
* OpenAI-compatible API server with Chat and Completions endpoints -- see the [examples](https://github.com/oobabooga/text-generation-webui/wiki/12-%E2%80%90-OpenAI-API#examples).
To restart the web UI in the future, just run the `start_` script again. This script creates an `installer_files` folder where it sets up the project's requirements. In case you need to reinstall the requirements, you can simply delete that folder and start the web UI again.
If you ever need to install something manually in the `installer_files` environment, you can launch an interactive shell using the cmd script: `cmd_linux.sh`, `cmd_windows.bat`, `cmd_macos.sh`, or `cmd_wsl.bat`.
* To install the requirements for extensions, you can use the `extensions_reqs` script for your OS. At the end, this script will install the main requirements for the project to make sure that they take precedence in case of version conflicts.
* For automated installation, you can use the `GPU_CHOICE`, `USE_CUDA118`, `LAUNCH_AFTER_INSTALL`, and `INSTALL_EXTENSIONS` environment variables. For instance: `GPU_CHOICE=A USE_CUDA118=FALSE LAUNCH_AFTER_INSTALL=FALSE INSTALL_EXTENSIONS=TRUE ./start_linux.sh`.
2) Manually install llama-cpp-python using the appropriate command for your hardware: [Installation from PyPI](https://github.com/abetlen/llama-cpp-python#installation-with-hardware-acceleration).
* Use the `LLAMA_HIPBLAS=on` toggle.
* Note the [Windows remarks](https://github.com/abetlen/llama-cpp-python#windows-remarks).
The `requirements*.txt` above contain various wheels precompiled through GitHub Actions. If you wish to compile things manually, or if you need to because no suitable wheels are available for your hardware, you can use `requirements_nowheels.txt` and then install your desired loaders manually.
* You need to have Docker Compose v2.17 or higher installed. See [this guide](https://github.com/oobabooga/text-generation-webui/wiki/09-%E2%80%90-Docker) for instructions.
--multi-user Multi-user mode. Chat histories are not saved or automatically loaded. Warning: this is likely not safe for sharing publicly.
--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.
--settings SETTINGS 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.
--chat-buttons Show buttons on the chat tab instead of a hover menu.
--idle-timeout IDLE_TIMEOUT Unload model after this many minutes of inactivity. It will be automatically reloaded when you try to use it again.
Model loader:
--loader LOADER Choose the model loader manually, otherwise, it will get autodetected. Valid options: Transformers, llama.cpp, llamacpp_HF, ExLlamav2_HF, ExLlamav2,
AutoGPTQ, AutoAWQ, GPTQ-for-LLaMa, QuIP#.
Transformers/Accelerate:
--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 VRAM usage slightly, but it comes at a performance cost.
--trust-remote-code Set trust_remote_code=True while loading the model. Necessary for some models.
--force-safetensors Set use_safetensors=True while loading the model. This prevents arbitrary code execution.
--no_use_fast Set use_fast=False while loading the tokenizer (it's True by default). Use this if you have any problems related to use_fast.
--use_flash_attention_2 Set use_flash_attention_2=True while loading the model.
bitsandbytes 4-bit:
--load-in-4bit Load the model with 4-bit precision (using bitsandbytes).
--quant_type QUANT_TYPE quant_type for 4-bit. Valid options: nf4, fp4.
llama.cpp:
--flash-attn Use flash-attention.
--tensorcores Use llama-cpp-python compiled with tensor cores support. This increases performance on RTX cards. NVIDIA only.
--n_ctx N_CTX Size of the prompt context.
--threads THREADS Number of threads to use.
--threads-batch THREADS_BATCH Number of threads to use for batches/prompt processing.
--no_mul_mat_q Disable the mulmat kernels.
--n_batch 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.
--n-gpu-layers N_GPU_LAYERS Number of layers to offload to the GPU.
--tensor_split TENSOR_SPLIT Split the model across multiple GPUs. Comma-separated list of proportions. Example: 18,17.
--numa Activate NUMA task allocation for llama.cpp.
--logits_all Needs to be set for perplexity evaluation to work. Otherwise, ignore it, as it makes prompt processing slower.
--no_offload_kqv Do not offload the K, Q, V to the GPU. This saves VRAM but reduces the performance.
--cache-capacity CACHE_CAPACITY Maximum cache capacity (llama-cpp-python). Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed.
--row_split Split the model by rows across GPUs. This may improve multi-gpu performance.
--streaming-llm Activate StreamingLLM to avoid re-evaluating the entire prompt when old messages are removed.
--attention-sink-size ATTENTION_SINK_SIZE StreamingLLM: number of sink tokens. Only used if the trimmed prompt does not share a prefix with the old prompt.
ExLlamaV2:
--gpu-split GPU_SPLIT Comma-separated list of VRAM (in GB) to use per GPU device for model layers. Example: 20,7,7.
--autosplit Autosplit the model tensors across the available GPUs. This causes --gpu-split to be ignored.
--max_seq_len MAX_SEQ_LEN Maximum sequence length.
--cfg-cache ExLlamav2_HF: Create an additional cache for CFG negative prompts. Necessary to use CFG with that loader.
--no_flash_attn Force flash-attention to not be used.
--cache_8bit Use 8-bit cache to save VRAM.
--cache_4bit Use Q4 cache to save VRAM.
--num_experts_per_token NUM_EXPERTS_PER_TOKEN Number of experts to use for generation. Applies to MoE models like Mixtral.
AutoGPTQ:
--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 do not have a quantize_config.json, this parameter is used to define whether to set desc_act or not in BaseQuantizeConfig.
--disable_exllama Disable ExLlama kernel, which can improve inference speed on some systems.
--disable_exllamav2 Disable ExLlamav2 kernel.
GPTQ-for-LLaMa:
--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.
HQQ:
--hqq-backend HQQ_BACKEND Backend for the HQQ loader. Valid options: PYTORCH, PYTORCH_COMPILE, ATEN.
DeepSpeed:
--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.
RoPE:
--alpha_value ALPHA_VALUE Positional embeddings alpha factor for NTK RoPE scaling. Use either this or compress_pos_emb, not both.
--rope_freq_base ROPE_FREQ_BASE If greater than 0, will be used instead of alpha_value. Those two are related by rope_freq_base = 10000 * alpha_value ^ (64 / 63).
--compress_pos_emb COMPRESS_POS_EMB Positional embeddings compression factor. Should be set to (context length) / (model's original context length). Equal to 1/rope_freq_scale.
Gradio:
--listen Make the web UI reachable from your local network.
--listen-port LISTEN_PORT The listening port that the server will use.
--listen-host LISTEN_HOST The hostname 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 GRADIO_AUTH Set Gradio authentication password in the format "username:password". Multiple credentials can also be supplied with "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 the same format as above.
--ssl-keyfile SSL_KEYFILE The path to the SSL certificate key file.
--ssl-certfile SSL_CERTFILE The path to the SSL certificate cert file.
API:
--api Enable the API extension.
--public-api Create a public URL for the API using Cloudfare.
--public-api-id PUBLIC_API_ID Tunnel ID for named Cloudflare Tunnel. Use together with public-api option.
--api-port API_PORT The listening port for the API.
--api-key API_KEY API authentication key.
--admin-key ADMIN_KEY API authentication key for admin tasks like loading and unloading models. If not set, will be the same as --api-key.
--nowebui Do not launch the Gradio UI. Useful for launching the API in standalone mode.
Multimodal:
--multimodal-pipeline MULTIMODAL_PIPELINE The multimodal pipeline to use. Examples: llava-7b, llava-13b.
Models should be placed in the folder `text-generation-webui/models`. They are usually downloaded from [Hugging Face](https://huggingface.co/models?pipeline_tag=text-generation&sort=downloads).
In both cases, you can use the "Model" tab of the UI to download the model from Hugging Face automatically. It is also possible to download it via the command-line with
If you would like to contribute to the project, check out the [Contributing guidelines](https://github.com/oobabooga/text-generation-webui/wiki/Contributing-guidelines).
In August 2023, [Andreessen Horowitz](https://a16z.com/) (a16z) provided a generous grant to encourage and support my independent work on this project. I am **extremely** grateful for their trust and recognition.