README_GPU.md

GPU Details

Hugging Face type models and LLaMa.cpp models are supported via CUDA on Linux and via MPS on macOS.

To run in ChatBot mode using bitsandbytes in 8-bit, run the following command:

python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6_9b --load_8bit=True

Then point your browser at http://0.0.0.0:7860 (linux) or http://localhost:7860 (windows/mac) or the public live URL printed by the server (disable shared link with --share=False). Note that for 4-bit or 8-bit support, older GPUs may require older bitsandbytes installed as pip uninstall bitsandbytes -y ; pip install bitsandbytes==0.38.1. For production uses, we recommend at least the 12B model, ran as:

python generate.py --base_model=HuggingFaceH4/zephyr-7b-beta --load_8bit=True

and one can use --h2ocolors=False to get soft blue-gray colors instead of H2O.ai colors. Here is a list of environment variables that can control some things in generate.py.

Note that if you download the model yourself and point --base_model to that location, you'll also need to specify the prompt_type by running:

python generate.py --base_model=<user path> --load_8bit=True --prompt_type=human_bot

for some user path <user path>. The prompt_type must match the model or a new version created in prompter.py or added in the UI/CLI via prompt_dict.

For quickly using a private document collection for Q/A, place documents (PDFs, text, etc.) into a folder called user_path and run the following command:

python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6_9b  --load_8bit=True --langchain_mode=UserData --user_path=user_path

For more details about document Q/A, see the LangChain Readme.

For 4-bit support when running generate.py, pass --load_4bit=True, which is only supported for certain architectures like GPT-NeoX-20B, GPT-J, LLaMa, etc.

Any other instruct-tuned base models can be used, including non-h2oGPT ones. Note that larger models require more GPU memory.

AutoGPTQ

Important: When running the following commands, if you encounter the message CUDA extension not installed during the loading of the model, you need to recompile. If you don't recompile, the generation will be significantly slower, even when using GPU.

An example with AutoGPTQ is:

python generate.py --base_model=TheBloke/Nous-Hermes-13B-GPTQ --score_model=None --load_gptq=model --use_safetensors=True --prompt_type=instruct --langchain_mode=UserData

This will use about 9800MB. You can also add --hf_embedding_model=sentence-transformers/all-MiniLM-L6-v2 to save some memory on embedding to reach 9340MB.

For LLaMa2 70B model quantized in 4-bit AutoGPTQ, you can run:

CUDA_VISIBLE_DEVICES=0 python generate.py --base_model=Llama-2-70B-chat-GPTQ --load_gptq="gptq_model-4bit--1g" --use_safetensors=True --prompt_type=llama2 --save_dir='save`

which gives about 12 tokens/sec. For 7b run:

python generate.py --base_model=TheBloke/Llama-2-7b-Chat-GPTQ --load_gptq="model" --use_safetensors=True --prompt_type=llama2 --save_dir='save`

For full 16-bit with 16k context across all GPUs:

pip install transformers==4.31.0  # breaks load_in_8bit=True in some cases (https://github.com/huggingface/transformers/issues/25026)
python generate.py --base_model=meta-llama/Llama-2-70b-chat-hf --prompt_type=llama2 --rope_scaling="{'type': 'linear', 'factor': 4}" --use_gpu_id=False --save_dir=savemeta70b

and running on 4xA6000 gives about 4tokens/sec consuming about 35GB per GPU of 4 GPUs when idle. Or for GPTQ with RoPE:

pip install transformers==4.31.0  # breaks load_in_8bit=True in some cases (https://github.com/huggingface/transformers/issues/25026)
python generate.py --base_model=TheBloke/Llama-2-7b-Chat-GPTQ --load_gptq="model" --use_safetensors=True --prompt_type=llama2 --score_model=None --save_dir='7bgptqrope4` --rope_scaling="{'type':'dynamic', 'factor':4}"
--max_max_new_tokens=15000 --max_new_tokens=15000 --max_time=12000

for which the GPU only uses 5.5GB. One can add (e.g.) --min_new_tokens=4096 to force generation to continue beyond model's training norms, although this may give lower quality responses. Currently, Hugging Face transformers does not support GPTQ directly except in text-generation-inference (TGI) server, but TGI does not support RoPE scaling. Also, vLLM supports LLaMa2 and AutoGPTQ but not RoPE scaling. Only exllama supports AutoGPTQ with RoPE scaling.

AutoAWQ

For 13B on 1 24GB board using about 14GB:

CUDA_VISIBLE_DEVICES=0 python generate.py --base_model=TheBloke/Llama-2-13B-chat-AWQ --score_model=None --load_awq=model --use_safetensors=True --prompt_type=llama2

or for 70B on 1 48GB board using about 39GB:

CUDA_VISIBLE_DEVICES=0 python generate.py --base_model=TheBloke/Llama-2-70B-chat-AWQ --score_model=None --load_awq=model --use_safetensors=True --prompt_type=llama2

or for 70B on 2 24GB boards:

CUDA_VISIBLE_DEVICES=2,3 python generate.py --base_model=TheBloke/Llama-2-70B-chat-AWQ --score_model=None --load_awq=model --use_safetensors=True --prompt_type=llama2

See for more details.

To run vLLM with 70B on 2 A100's using h2oGPT, follow the vLLM install instructions and then do:

python -m vllm.entrypoints.openai.api_server \
        --port=5000 \
        --host=0.0.0.0 \
        --model=h2oai/h2ogpt-4096-llama2-70b-chat-4bit \
        --tensor-parallel-size=2 \
        --seed 1234 \
        --trust-remote-code \
	    --max-num-batched-tokens 8192 \
	    --quantization awq \
        --download-dir=/$HOME/.cache/huggingface/hub

for choice of port, IP, model, some number of GPUs matching tensor-parallel-size, etc. Or with docker with built-in vLLM:

mkdir -p $HOME/.cache/huggingface/hub
mkdir -p $HOME/.cache/huggingface/modules/
mkdir -p $HOME/.triton/cache/
mkdir -p $HOME/.config/vllm
docker run -d \
    --runtime=nvidia \
    --gpus '"device=0,1"' \
    --shm-size=10.24gb \
    -p 5000:5000 \
    -e NCCL_IGNORE_DISABLED_P2P=1 \
    -e HUGGING_FACE_HUB_TOKEN=$HUGGING_FACE_HUB_TOKEN \
    -e VLLM_NO_USAGE_STATS=1 \
    -e VLLM_NCCL_SO_PATH=/usr/local/lib/python3.10/dist-packages/nvidia/nccl/lib/libnccl.so.2 \
    -e DO_NOT_TRACK=1 \
    -e NUMBA_CACHE_DIR=/tmp/ \
    -v /etc/passwd:/etc/passwd:ro \
    -v /etc/group:/etc/group:ro \
    -u `id -u`:`id -g` \
    -v "${HOME}"/.cache:$HOME/.cache/ -v "${HOME}"/.config:$HOME/.config/   -v "${HOME}"/.triton:$HOME/.triton/  \
    --network host \
    vllm/vllm-openai:latest \
        --port=5000 \
        --host=0.0.0.0 \
        --model=h2oai/h2ogpt-4096-llama2-70b-chat-4bit \
        --tensor-parallel-size=2 \
        --seed 1234 \
        --trust-remote-code \
	      --max-num-batched-tokens 8192 \
	      --quantization awq \
        --download-dir=/workspace/.cache/huggingface/hub &>> logs.vllm_server.70b_awq.txt

Can run same thing with 4 GPUs (to be safe) on 4*A10G like more available on AWS.

exllama

Currently, only exllama supports AutoGPTQ with RoPE scaling. To run RoPE scaling the LLaMa-2 7B model for 16k context:

python generate.py --base_model=TheBloke/Llama-2-7b-Chat-GPTQ --load_gptq="model" --use_safetensors=True --prompt_type=llama2 --save_dir='save' --load_exllama=True --revision=gptq-4bit-32g-actorder_True --rope_scaling="{'alpha_value':4}"

which shows how to control alpha_value and the revision for a given model on TheBloke/Llama-2-7b-Chat-GPTQ. Be careful as setting alpha_value higher consumes substantially more GPU memory. Also, some models have incorrect config values for max_position_embeddings or max_sequence_length, and we try to fix those for LLaMa2 if llama-2 appears in the lower-case version of the model name. Another type of model is

python generate.py --base_model=TheBloke/Nous-Hermes-Llama2-GPTQ --load_gptq="model" --use_safetensors=True --prompt_type=llama2 --save_dir='save' --load_exllama=True --revision=gptq-4bit-32g-actorder_True --rope_scaling="{'alpha_value':4}"

and note the different prompt_type. For LLaMa2 70B run:

python generate.py --base_model=TheBloke/Llama-2-70B-chat-GPTQ --load_gptq=gptq_model-4bit-128g --use_safetensors=True --prompt_type=llama2 --load_exllama=True --revision=main

which uses about 48GB of memory on 1 GPU and runs at about 12 tokens/second on an A6000, which is about half the speed of 16-bit if run that on 2*A100 GPUs.

With exllama, ensure --concurrency_count=1 else the model will share states and mix-up concurrent requests.

You can set other exllama options by passing --exllama_dict. For example, for LLaMa-2-70B on 2 GPUs each using 20GB, you can run the following command:

python generate.py --base_model=TheBloke/Llama-2-70B-chat-GPTQ --load_exllama=True --use_safetensors=True --use_gpu_id=False --load_gptq=main --prompt_type=llama2 --exllama_dict="{'set_auto_map':'20,20'}"

For LLaMa.cpp on GPU run:

python generate.py --base_model=HuggingFaceH4/zephyr-7b-beta --prompt_type=zephyr --score_model=None --user_path=user_path

and ensure that the output shows that one or more GPUs is in use by looking at the logs.

• 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.