README_WINDOWS.md

Windows 10/11

If using GPU on Windows 10/11 Pro 64-bit, we recommend using Windows installers.

For newer builds of windows versions of 10/11.

Installation

• Download Visual Studio 2022: Download Link
  • Run Installer, click ok to run, click Continue
  • Click on Individual Components
  • Search for these in the search bar and click on them:
    • Windows 11 SDK (e.g. 10.0.22000.0)
    • C++ Universal Windows Platform support (e.g. for v143 build tools)
    • MSVC VS 2022 C++ x64/x86 build tools (latest)
    • C++ CMake tools for Windows
    • 
  • Click Install, and follow through installation, and do not need to launch VS 2022 at end.
• Download the MinGW installer: MiniGW
  • Run Installer, Click Install, Continue, Install/Run to launch installation manager.
  • Select packages to install:
    • minigw32-base
    • mingw32-gcc-g++
    • 
  • Go to installation tab, then apply changes.
• Download and install Miniconda
• Run Miniconda shell (not powershell!) as Administrator
• Run: set path=%path%;c:\MinGW\msys\1.0\bin\ to get C++ in path. In some cases it may be instead correct to use set path=%path%;c:\MinGW\bin\
• Download latest nvidia driver for windows if one has old drivers before CUDA 11.7 supported
• Confirm can run nvidia-smi and see driver version
• Setup Conda Environment:
  • 

   conda create -n h2ogpt -y
   conda activate h2ogpt
   conda install python=3.10 -c conda-forge -y
   python --version  # should say python 3.10.xx
   python -c "import os, sys ; print('hello world')"  # should print "hello world"

• GPU Only: Install CUDA

   conda install cudatoolkit=11.7 -c conda-forge -y
   set CUDA_HOME=$CONDA_PREFIX

• Install Git:

   conda install -c conda-forge git

• Install h2oGPT:

   git clone https://github.com/h2oai/h2ogpt.git
   cd h2ogpt

• Install primary dependencies.
  • Remove any bad dependencies that existed (required for new transformers it seems):

    pip uninstall flash-attn

  • For CPU Only:

    pip install -r requirements.txt --extra-index https://download.pytorch.org/whl/cpu

  • For GPU:

    pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu117

    In some cases this may lead to the message No GPU and in which case you can run next something like:

    pip install torch==2.0.0+cu117 torchvision==0.15.1+cu117 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu117

• Optional: for bitsandbytes 4-bit and 8-bit:

  pip uninstall bitsandbytes -y
  pip install https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.41.1-py3-none-win_amd64.whl

• Install document question-answer dependencies:

   # Required for Doc Q/A: LangChain:
   pip install -r reqs_optional/requirements_optional_langchain.txt --extra-index-url https://download.pytorch.org/whl/cu117
   # Required for CPU: LLaMa/GPT4All:
   pip install -r reqs_optional/requirements_optional_gpt4all.txt --extra-index-url https://download.pytorch.org/whl/cu117
   # Optional: PyMuPDF/ArXiv:
   pip install -r reqs_optional/requirements_optional_langchain.gpllike.txt --extra-index-url https://download.pytorch.org/whl/cu117
   # Optional: Selenium/PlayWright:
   pip install -r reqs_optional/requirements_optional_langchain.urls.txt --extra-index-url https://download.pytorch.org/whl/cu117
   # Optional: for supporting unstructured package
   python -m nltk.downloader all
   # Optional but required for PlayWright
   playwright install --with-deps
   # Note: for Selenium, we match versions of playwright so above installer will add chrome version needed

• GPU Optional: For optional AutoGPTQ support:

   pip uninstall -y auto-gptq
   pip install https://github.com/PanQiWei/AutoGPTQ/releases/download/v0.4.2/auto_gptq-0.4.2+cu118-cp310-cp310-win_amd64.whl

• GPU Optional: For optional AutoAWQ support:

   pip uninstall -y autoawq
   pip install autoawq

• GPU Optional: For optional exllama support:

  pip uninstall -y exllama
  pip install https://github.com/jllllll/exllama/releases/download/0.0.13/exllama-0.0.13+cu118-cp310-cp310-win_amd64.whl --no-cache-dir

• GPU Optional: Support LLaMa.cpp with CUDA via llama-cpp-python:
  • Download/Install CUDA llama-cpp-python wheel, or choose link and run pip directly. E.g.:

      pip uninstall -y llama-cpp-python llama_cpp_python_cuda
      # GGMLv3 ONLY:
      pip install https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda-0.1.73+cu117-cp310-cp310-win_amd64.whl
      # GGUF ONLY for GPU:
      pip install https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda-0.1.83+cu117-cp310-cp310-win_amd64.whl
      # GGUF ONLY for CPU for AVX2:
      pip install https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/cpu/llama_cpp_python-0.1.83+cpuavx2-cp310-cp310-win_amd64.whl

    See https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases for other releases, try to stick to same version.
  • If any issues, then must compile llama-cpp-python with CUDA support:

    pip uninstall -y llama-cpp-python
    set LLAMA_CUBLAS=1
    set CMAKE_ARGS=-DLLAMA_CUBLAS=on
    set FORCE_CMAKE=1
    pip install llama-cpp-python==0.1.68 --no-cache-dir --verbose

  • By default, we set n_gpu_layers to large value, so llama.cpp offloads all layers for maximum GPU performance. You can control this by passing --llamacpp_dict="{'n_gpu_layers':20}" for value 20, or setting in UI. For highest performance, offload all layers. That is, one gets maximum performance if one sees in startup of h2oGPT all layers offloaded:

    llama_model_load_internal: offloaded 35/35 layers to GPU
    

  but this requires sufficient GPU memory. Reduce if you have low memory GPU, say 15.
  • Pass to generate.py the option --max_seq_len=2048 or some other number if you want model have controlled smaller context, else default (relatively large) value is used that will be slower on CPU.
  • If one sees /usr/bin/nvcc mentioned in errors, that file needs to be removed as would likely conflict with version installed for conda.
  • Note that once llama-cpp-python is compiled to support CUDA, it no longer works for CPU mode, so one would have to reinstall it without the above options to recovers CPU mode or have a separate h2oGPT env for CPU mode.
• For supporting Word and Excel documents, if you don't have Word/Excel already, then download and install libreoffice: https://www.libreoffice.org/download/download-libreoffice/ .
• To support OCR, download and install tesseract, see also: Tesseract Documentation. Please add the installation directories to your PATH.

---

Run

• For document Q/A with UI using LLaMa.cpp-based model on CPU or GPU:

  • Click Download LLaMa2 Model and place file in h2oGPT repo directory. Any other TheBloke GGML v3 model can be used by changing value of --model_path_llama to path previously downloaded or URL.

    python generate.py --base_model='llama' --prompt_type=llama2 --score_model=None --langchain_mode='UserData' --user_path=user_path --model_path_llama=https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGML/resolve/main/llama-2-7b-chat.ggmlv3.q8_0.bin --max_seq_len=4096

    Choose some other model_path_llama from TheBloke if desired, e.g. 13B. If no model passed, the 7B LLaMa-2 GGML is used. For an absolute windows path, change to --user_path=C:\Users\YourUsername\h2ogpt or something similar for some user YourUsername. If llama-cpp-python was compiled with CUDA support, you should see in the output:

    Starting get_model: llama
    ggml_init_cublas: found 2 CUDA devices:
      Device 0: NVIDIA GeForce RTX 3090 Ti
      Device 1: NVIDIA GeForce RTX 2080
    llama.cpp: loading model from llama-2-7b-chat.ggmlv3.q8_0.bin
    llama_model_load_internal: format     = ggjt v3 (latest)
    llama_model_load_internal: n_vocab    = 32001
    llama_model_load_internal: n_ctx      = 1792
    llama_model_load_internal: n_embd     = 4096
    llama_model_load_internal: n_mult     = 256
    llama_model_load_internal: n_head     = 32
    llama_model_load_internal: n_layer    = 32
    llama_model_load_internal: n_rot      = 128
    llama_model_load_internal: ftype      = 7 (mostly Q8_0)
    llama_model_load_internal: n_ff       = 11008
    llama_model_load_internal: model size = 7B
    llama_model_load_internal: ggml ctx size =    0.08 MB
    llama_model_load_internal: using CUDA for GPU acceleration
    ggml_cuda_set_main_device: using device 0 (NVIDIA GeForce RTX 3090 Ti) as main device
    llama_model_load_internal: mem required  = 4518.85 MB (+ 1026.00 MB per state)
    llama_model_load_internal: allocating batch_size x (512 kB + n_ctx x 128 B) = 368 MB VRAM for the scratch buffer
    llama_model_load_internal: offloading 20 repeating layers to GPU
    llama_model_load_internal: offloaded 20/35 layers to GPU
    llama_model_load_internal: total VRAM used: 4470 MB
    llama_new_context_with_model: kv self size  =  896.00 MB
    AVX = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | VSX = 0 |
    Model {'base_model': 'llama', 'tokenizer_base_model': '', 'lora_weights': '', 'inference_server': '', 'prompt_type': 'llama2', 'prompt_dict': {'promptA': 'Below is an instruction that describes a task. Write a response that appropriately completes the request.', 'promptB': 'Below is an instruction that describes a task. Write a response that appropriately completes the request.', 'PreInstruct': '\n### Instruction:\n', 'PreInput': None, 'PreResponse': '\n### Response:\n', 'terminate_response': ['\n### Response:\n'], 'chat_sep': '\n', 'chat_turn_sep': '\n', 'humanstr': '\n### Instruction:\n', 'botstr': '\n### Response:\n', 'generates_leading_space': False}}
    Running on local URL:  http://0.0.0.0:7860
    
    To create a public link, set `share=True` in `launch()`.
    

  • Go to http://127.0.0.1:7860 (ignore message above). Add --share=True to get sharable secure link.

  • To just chat with LLM, click Resources and click LLM in Collections, or start without --langchain_mode=UserData.

  • In nvidia-smi or some other GPU monitor program you should see python.exe using GPUs in C (Compute) mode and using GPU resources.

  • If you have multiple GPUs, best to specify to use the fasted GPU by doing (e.g. if device 0 is fastest and largest memory GPU):

    set CUDA_VISIBLE_DEVICES=0

  • On an i9 with 3090Ti, one gets about 5 tokens/second.

  • 

  • For LLaMa2 70B model, launch as

    python generate.py --base_model=llama --model_path_llama=llama-2-70b-chat.ggmlv3.q8_0.bin n_gqa=8

    where one should have downloaded the zip and extracted from here. See LLaMa.cpp Instructions for more details.

• To use Hugging Face type models (faster on GPU than LLaMa.cpp if one has a powerful GPU with enough memory):

  python generate.py --base_model=h2oai/h2ogpt-gm-oasst1-en-2048-falcon-7b-v3 --langchain_mode=UserData --score_model=None

  • On an i9 with 3090Ti, one gets about 9 tokens/second.

• To use Hugging Face type models in 8-bit do:

  python generate.py --base_model=h2oai/h2ogpt-gm-oasst1-en-2048-falcon-7b-v3 --langchain_mode=UserData --score_model=None --load_8bit=True

  When running windows on GPUs with bitsandbytes in 8-bit you should see something like the below in output:

  bin C:\Users\pseud\.conda\envs\h2ogpt\lib\site-packages\bitsandbytes\libbitsandbytes_cuda117.dll

  • On an i9 with 3090Ti, one gets about 5 tokens/second, so about half 16-bit speed.
  • You can confirm GPU use via nvidia-smi showing GPU memory consumed is less than 16-bit, at about 9.2GB when in use. Also try 13B models in 8-bit for similar memory usage.
  • Note 8-bit inference is about twice slower than 16-bit inference, and the only use of 8-bit is to keep memory profile low.
  • Bitsandbytes can be uninstalled (pip uninstall bitsandbytes) and still h2oGPT can be used if one does not pass --load_8bit=True.

• To use Hugging Face type models in 4-bit do:

  python generate.py --base_model=h2oai/h2ogpt-gm-oasst1-en-2048-falcon-7b-v3 --langchain_mode=UserData --score_model=None --load_4bit=True

  • On an i9 with 3090Ti, one gets about 4 tokens/second, so still about half 16-bit speed. Memory use is about 6.6GB.

See CPU and GPU for some other general aspects about using h2oGPT on CPU or GPU, such as which models to try, quantization, etc.

Issues

• SSL Certification failure when connecting to Hugging Face.
  • Your org may be blocking HF
  • Try: https://stackoverflow.com/a/75111104
  • Or try: huggingface/transformers#17611 (comment)
  • Try using proxy.
• If you see import problems, then try setting PYTHONPATH in a .bat file:

  SET PYTHONPATH=.:src:$PYTHONPATH
  python generate.py ...

  for some options ...
• For easier handling of command line operations, consider using bash in windows with coreutils.