GPU Acceleration

note

At this time, only NVIDIA GPU acceleration is supported. Please open an issue if you'd like to see support for other GPU vendors.

NVIDIA

AIKit supports GPU accelerated inferencing with NVIDIA Container Toolkit. You must also have NVIDIA Drivers installed on your host machine.

For Kubernetes, NVIDIA GPU Operator provides a streamlined way to install the NVIDIA drivers and container toolkit to configure your cluster to use GPUs.

To get started with GPU-accelerated inferencing, make sure to set the following in your aikitfile and build your model.

runtime: cuda         # use NVIDIA CUDA runtime

For llama backend, set the following in your config:

f16: true             # use float16 precision
gpu_layers: 35        # number of layers to offload to GPU
low_vram: true        # for devices with low VRAM

tip

Make sure to customize these values based on your model and GPU specs.

note

For exllama and exllama2 backends, GPU acceleration is enabled by default and cannot be disabled.

After building the model, you can run it with --gpus all flag to enable GPU support:

# for pre-made models, replace "my-model" with the image name
docker run --rm --gpus all -p 8080:8080 my-model

If GPU acceleration is working, you'll see output that is similar to following in the debug logs:

32AM DBG GRPC(llama-2-7b-chat.Q4_K_M.gguf-127.0.0.1:43735): stderr ggml_init_cublas: found 1 CUDA devices:
32AM DBG GRPC(llama-2-7b-chat.Q4_K_M.gguf-127.0.0.1:43735): stderr   Device 0: Tesla T4, compute capability 7.5
...
32AM DBG GRPC(llama-2-7b-chat.Q4_K_M.gguf-127.0.0.1:43735): stderr llm_load_tensors: using CUDA for GPU acceleration
32AM DBG GRPC(llama-2-7b-chat.Q4_K_M.gguf-127.0.0.1:43735): stderr llm_load_tensors: mem required  =   70.41 MB (+ 2048.00 MB per state)
32AM DBG GRPC(llama-2-7b-chat.Q4_K_M.gguf-127.0.0.1:43735): stderr llm_load_tensors: offloading 32 repeating layers to GPU
32AM DBG GRPC(llama-2-7b-chat.Q4_K_M.gguf-127.0.0.1:43735): stderr llm_load_tensors: offloading non-repeating layers to GPU
32AM DBG GRPC(llama-2-7b-chat.Q4_K_M.gguf-127.0.0.1:43735): stderr llm_load_tensors: offloading v cache to GPU
32AM DBG GRPC(llama-2-7b-chat.Q4_K_M.gguf-127.0.0.1:43735): stderr llm_load_tensors: offloading k cache to GPU
32AM DBG GRPC(llama-2-7b-chat.Q4_K_M.gguf-127.0.0.1:43735): stderr llm_load_tensors: offloaded 35/35 layers to GPU
32AM DBG GRPC(llama-2-7b-chat.Q4_K_M.gguf-127.0.0.1:43735): stderr llm_load_tensors: VRAM used: 5869 MB

NVIDIA​

NVIDIA