phoenix-oss/llama-stack
1
0
Fork 0
forked from phoenix-oss/llama-stack-mirror
Code Pull requests Releases Packages 2 Activity Actions
llama-stack/docs/source/distributions/self_hosted_distro/remote-vllm.md
AlexHe99 983f6feeb8
docs: Update remote-vllm.md with AMD GPU vLLM server supported. (#1858) 
Add the content to use AMD GPU as the vLLM server. Split the original
part to two sub chapters,
1. AMD vLLM server
2. NVIDIA vLLM server (orignal)

# What does this PR do?
[Provide a short summary of what this PR does and why. Link to relevant
issues if applicable.]

[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])

## Test Plan
[Describe the tests you ran to verify your changes with result
summaries. *Provide clear instructions so the plan can be easily
re-executed.*]

[//]: # (## Documentation)

---------

Signed-off-by: Alex He <alehe@amd.com>
2025-04-08 21:35:32 -07:00

8.3 KiB
Raw Blame History

orphan
true

Remote vLLM Distribution

:maxdepth: 2
:hidden:

self

The llamastack/distribution-remote-vllm distribution consists of the following provider configurations:

API Provider(s)
agents inline::meta-reference
datasetio remote::huggingface, inline::localfs
eval inline::meta-reference
inference remote::vllm, inline::sentence-transformers
safety inline::llama-guard
scoring inline::basic, inline::llm-as-judge, inline::braintrust
telemetry inline::meta-reference
tool_runtime remote::brave-search, remote::tavily-search, inline::code-interpreter, inline::rag-runtime, remote::model-context-protocol, remote::wolfram-alpha
vector_io inline::faiss, remote::chromadb, remote::pgvector

You can use this distribution if you have GPUs and want to run an independent vLLM server container for running inference.

Environment Variables

The following environment variables can be configured:

  • LLAMA_STACK_PORT: Port for the Llama Stack distribution server (default: 8321)
  • INFERENCE_MODEL: Inference model loaded into the vLLM server (default: meta-llama/Llama-3.2-3B-Instruct)
  • VLLM_URL: URL of the vLLM server with the main inference model (default: http://host.docker.internal:5100/v1)
  • MAX_TOKENS: Maximum number of tokens for generation (default: 4096)
  • SAFETY_VLLM_URL: URL of the vLLM server with the safety model (default: http://host.docker.internal:5101/v1)
  • SAFETY_MODEL: Name of the safety (Llama-Guard) model to use (default: meta-llama/Llama-Guard-3-1B)

Setting up vLLM server

Both AMD and NVIDIA GPUs can serve as accelerators for the vLLM server, which acts as both the LLM inference provider and the safety provider.

Setting up vLLM server on AMD GPU

AMD provides two main vLLM container options:

  • rocm/vllm: Production-ready container
  • rocm/vllm-dev: Development container with the latest vLLM features

Please check the Blog about ROCm vLLM Usage to get more details.

Here is a sample script to start a ROCm vLLM server locally via Docker:

export INFERENCE_PORT=8000
export INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct
export CUDA_VISIBLE_DEVICES=0
export VLLM_DIMG="rocm/vllm-dev:main"

docker run \
    --pull always \
    --ipc=host \
    --privileged \
    --shm-size 16g \
    --device=/dev/kfd \
    --device=/dev/dri \
    --group-add video \
    --cap-add=SYS_PTRACE \
    --cap-add=CAP_SYS_ADMIN \
    --security-opt seccomp=unconfined \
    --security-opt apparmor=unconfined \
    --env "HUGGING_FACE_HUB_TOKEN=$HF_TOKEN" \
    --env "HIP_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" \
    -p $INFERENCE_PORT:$INFERENCE_PORT \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    $VLLM_DIMG \
    python -m vllm.entrypoints.openai.api_server \
    --model $INFERENCE_MODEL \
    --port $INFERENCE_PORT

Note that you'll also need to set --enable-auto-tool-choice and --tool-call-parser to enable tool calling in vLLM.

If you are using Llama Stack Safety / Shield APIs, then you will need to also run another instance of a vLLM with a corresponding safety model like meta-llama/Llama-Guard-3-1B using a script like:

export SAFETY_PORT=8081
export SAFETY_MODEL=meta-llama/Llama-Guard-3-1B
export CUDA_VISIBLE_DEVICES=1
export VLLM_DIMG="rocm/vllm-dev:main"

docker run \
    --pull always \
    --ipc=host \
    --privileged \
    --shm-size 16g \
    --device=/dev/kfd \
    --device=/dev/dri \
    --group-add video \
    --cap-add=SYS_PTRACE \
    --cap-add=CAP_SYS_ADMIN \
    --security-opt seccomp=unconfined \
    --security-opt apparmor=unconfined \
    --env "HUGGING_FACE_HUB_TOKEN=$HF_TOKEN" \
    --env "HIP_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" \
    -p $SAFETY_PORT:$SAFETY_PORT \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    $VLLM_DIMG \
    python -m vllm.entrypoints.openai.api_server \
    --model $SAFETY_MODEL \
    --port $SAFETY_PORT

Setting up vLLM server on NVIDIA GPU

Please check the vLLM Documentation to get a vLLM endpoint. Here is a sample script to start a vLLM server locally via Docker:

export INFERENCE_PORT=8000
export INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct
export CUDA_VISIBLE_DEVICES=0

docker run \
    --pull always \
    --runtime nvidia \
    --gpus $CUDA_VISIBLE_DEVICES \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HUGGING_FACE_HUB_TOKEN=$HF_TOKEN" \
    -p $INFERENCE_PORT:$INFERENCE_PORT \
    --ipc=host \
    vllm/vllm-openai:latest \
    --gpu-memory-utilization 0.7 \
    --model $INFERENCE_MODEL \
    --port $INFERENCE_PORT

Note that you'll also need to set --enable-auto-tool-choice and --tool-call-parser to enable tool calling in vLLM.

If you are using Llama Stack Safety / Shield APIs, then you will need to also run another instance of a vLLM with a corresponding safety model like meta-llama/Llama-Guard-3-1B using a script like:

export SAFETY_PORT=8081
export SAFETY_MODEL=meta-llama/Llama-Guard-3-1B
export CUDA_VISIBLE_DEVICES=1

docker run \
    --pull always \
    --runtime nvidia \
    --gpus $CUDA_VISIBLE_DEVICES \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HUGGING_FACE_HUB_TOKEN=$HF_TOKEN" \
    -p $SAFETY_PORT:$SAFETY_PORT \
    --ipc=host \
    vllm/vllm-openai:latest \
    --gpu-memory-utilization 0.7 \
    --model $SAFETY_MODEL \
    --port $SAFETY_PORT

Running Llama Stack

Now you are ready to run Llama Stack with vLLM as the inference provider. You can do this via Conda (build code) or Docker which has a pre-built image.

Via Docker

This method allows you to get started quickly without having to build the distribution code.

export INFERENCE_PORT=8000
export INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct
export LLAMA_STACK_PORT=8321

# You need a local checkout of llama-stack to run this, get it using
# git clone https://github.com/meta-llama/llama-stack.git
cd /path/to/llama-stack

docker run \
  --pull always \
  -p $LLAMA_STACK_PORT:$LLAMA_STACK_PORT \
  -v ./llama_stack/templates/remote-vllm/run.yaml:/root/my-run.yaml \
  llamastack/distribution-remote-vllm \
  --yaml-config /root/my-run.yaml \
  --port $LLAMA_STACK_PORT \
  --env INFERENCE_MODEL=$INFERENCE_MODEL \
  --env VLLM_URL=http://host.docker.internal:$INFERENCE_PORT/v1

If you are using Llama Stack Safety / Shield APIs, use:

export SAFETY_PORT=8081
export SAFETY_MODEL=meta-llama/Llama-Guard-3-1B

# You need a local checkout of llama-stack to run this, get it using
# git clone https://github.com/meta-llama/llama-stack.git
cd /path/to/llama-stack

docker run \
  --pull always \
  -p $LLAMA_STACK_PORT:$LLAMA_STACK_PORT \
  -v ~/.llama:/root/.llama \
  -v ./llama_stack/templates/remote-vllm/run-with-safety.yaml:/root/my-run.yaml \
  llamastack/distribution-remote-vllm \
  --yaml-config /root/my-run.yaml \
  --port $LLAMA_STACK_PORT \
  --env INFERENCE_MODEL=$INFERENCE_MODEL \
  --env VLLM_URL=http://host.docker.internal:$INFERENCE_PORT/v1 \
  --env SAFETY_MODEL=$SAFETY_MODEL \
  --env SAFETY_VLLM_URL=http://host.docker.internal:$SAFETY_PORT/v1

Via Conda

Make sure you have done uv pip install llama-stack and have the Llama Stack CLI available.

export INFERENCE_PORT=8000
export INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct
export LLAMA_STACK_PORT=8321

cd distributions/remote-vllm
llama stack build --template remote-vllm --image-type conda

llama stack run ./run.yaml \
  --port $LLAMA_STACK_PORT \
  --env INFERENCE_MODEL=$INFERENCE_MODEL \
  --env VLLM_URL=http://localhost:$INFERENCE_PORT/v1

If you are using Llama Stack Safety / Shield APIs, use:

export SAFETY_PORT=8081
export SAFETY_MODEL=meta-llama/Llama-Guard-3-1B

llama stack run ./run-with-safety.yaml \
  --port $LLAMA_STACK_PORT \
  --env INFERENCE_MODEL=$INFERENCE_MODEL \
  --env VLLM_URL=http://localhost:$INFERENCE_PORT/v1 \
  --env SAFETY_MODEL=$SAFETY_MODEL \
  --env SAFETY_VLLM_URL=http://localhost:$SAFETY_PORT/v1