Skip to content

GitLab

"...ssh:/git@developer.sourcefind.cn:2222/OpenDAS/dynamo.git" did not exist on "9ab148dcb9daa96db3a836e7ad3218895d6abbe1"
Unverified Commit 5c5cec3d authored by ishandhanani's avatar ishandhanani Committed by GitHub
Browse files

feat(sglang): add dockerfile/pyproject toml entry + steps to run dsr1 disagg (#1193) 


Signed-off-by: default avatarishandhanani <82981111+ishandhanani@users.noreply.github.com>
Co-authored-by: default avatarNeelay Shah <neelays@nvidia.com>
parent 088f7eeb
Hide whitespace changes
Inline Side-by-side
Showing with 626 additions and 10 deletions
container/Dockerfile.sglang 0 → 100644
View file @ 5c5cec3d
# SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
ARG BASE_IMAGE="nvcr.io/nvidia/cuda-dl-base"
# FIXME: NCCL will hang with 25.03, so use 25.01 for now
# Please check https://github.com/ai-dynamo/dynamo/pull/1065
# for details and reproducer to manually test if the image
# can be updated to later versions.
ARG BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"
ARG RELEASE_BUILD
ARG RUNTIME_IMAGE="nvcr.io/nvidia/cuda"
ARG RUNTIME_IMAGE_TAG="12.8.1-runtime-ubuntu24.04"
# TODO: Move to published pypi tags
ARG GENAI_PERF_TAG="e67e853413a07a778dd78a55e299be7fba9c9c24"
# Define general architecture ARGs for supporting both x86 and aarch64 builds.
# ARCH: Used for package suffixes (e.g., amd64, arm64)
# ARCH_ALT: Used for Rust targets, manylinux suffix (e.g., x86_64, aarch64)
#
# Default values are for x86/amd64:
# --build-arg ARCH=amd64 --build-arg ARCH_ALT=x86_64
#
# For arm64/aarch64, build with:
# --build-arg ARCH=arm64 --build-arg ARCH_ALT=aarch64
#
# NOTE: There isn't an easy way to define one of these values based on the other value
# without adding if statements everywhere, so just define both as ARGs for now.
ARG ARCH=amd64
ARG ARCH_ALT=x86_64
FROM ${BASE_IMAGE}:${BASE_IMAGE_TAG} AS nixl_base
# Redeclare ARCH and ARCH_ALT so they're available in this stage
ARG ARCH
ARG ARCH_ALT
WORKDIR /opt/nixl
# Add a cache hint that only changes when the nixl commit changes
ARG NIXL_COMMIT
# This line acts as a cache key - it only changes when NIXL_COMMIT changes
RUN echo "NIXL commit: ${NIXL_COMMIT}" > /opt/nixl/commit.txt
# Copy the nixl source
COPY --from=nixl . .
##################################
########## Base Image ############
##################################
FROM ${BASE_IMAGE}:${BASE_IMAGE_TAG} AS base
# Redeclare ARCH and ARCH_ALT so they're available in this stage
ARG ARCH
ARG ARCH_ALT
USER root
ARG PYTHON_VERSION=3.12
RUN apt-get update -y && \
apt-get install -y \
# NIXL build dependencies
cmake \
meson \
ninja-build \
pybind11-dev \
# Rust build dependencies
clang \
libclang-dev \
git \
# Install utilities
nvtop \
tmux \
vim \
autoconf \
libtool
WORKDIR /workspace
### TODO: Bring back UCX EFA setup once we are confident it works with IB devices
ENV CPATH=/usr/include:$CPATH
ENV PATH=/usr/bin:$PATH
ENV PKG_CONFIG_PATH=/usr/lib/pkgconfig:$PKG_CONFIG_PATH
SHELL ["/bin/bash", "-c"]
WORKDIR /workspace
### NIXL SETUP ###
# Copy nixl source, and use commit hash as cache hint
COPY --from=nixl_base /opt/nixl /opt/nixl
COPY --from=nixl_base /opt/nixl/commit.txt /opt/nixl/commit.txt
RUN if [ "$ARCH" = "arm64" ]; then \
cd /opt/nixl && \
mkdir build && \
meson setup build/ --prefix=/usr/local/nixl -Dgds_path=/usr/local/cuda/targets/sbsa-linux && \
cd build/ && \
ninja && \
ninja install; \
else \
cd /opt/nixl && \
mkdir build && \
meson setup build/ --prefix=/usr/local/nixl && \
cd build/ && \
ninja && \
ninja install; \
fi
### NATS & ETCD SETUP ###
# nats
RUN wget --tries=3 --waitretry=5 https://github.com/nats-io/nats-server/releases/download/v2.10.24/nats-server-v2.10.24-${ARCH}.deb && \
dpkg -i nats-server-v2.10.24-${ARCH}.deb && rm nats-server-v2.10.24-${ARCH}.deb
# etcd
ENV ETCD_VERSION="v3.5.18"
RUN wget --tries=3 --waitretry=5 https://github.com/etcd-io/etcd/releases/download/$ETCD_VERSION/etcd-$ETCD_VERSION-linux-${ARCH}.tar.gz -O /tmp/etcd.tar.gz && \
mkdir -p /usr/local/bin/etcd && \
tar -xvf /tmp/etcd.tar.gz -C /usr/local/bin/etcd --strip-components=1 && \
rm /tmp/etcd.tar.gz
ENV PATH=/usr/local/bin/etcd/:$PATH
### VIRTUAL ENVIRONMENT SETUP ###
# Install uv and create virtualenv
COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/
RUN mkdir /opt/dynamo && \
uv venv /opt/dynamo/venv --python 3.12
# Activate virtual environment
ENV VIRTUAL_ENV=/opt/dynamo/venv
ENV PATH="${VIRTUAL_ENV}/bin:${PATH}"
# Install NIXL Python module
# TODO: Move gds_path selection based on arch into NIXL build
RUN if [ "$ARCH" = "arm64" ]; then \
cd /opt/nixl && uv pip install . --config-settings=setup-args="-Dgds_path=/usr/local/cuda/targets/sbsa-linux"; \
else \
cd /opt/nixl && uv pip install . ; \
fi
# Install sglang
# TODO: NIXL transfer is currently broken as of https://github.com/sgl-project/sglang/commit/7513558074adc4c4015b68e2ae7cf719d3401d5d
# Once this is fixed we will have to install from that commit until a new post is released
ARG SGLANG_COMMIT="4d643f6c7a291c86de64a9e52eca526b2d99775d"
RUN --mount=type=cache,target=/root/.cache/uv \
git clone https://github.com/sgl-project/sglang.git && \
cd sglang && \
git checkout ${SGLANG_COMMIT} && \
uv pip install -e "python[all]" && \
cd .. && \
rm -rf sglang
# Common dependencies
RUN --mount=type=bind,source=./container/deps/requirements.txt,target=/tmp/requirements.txt \
uv pip install --requirement /tmp/requirements.txt
# Install test dependencies
RUN --mount=type=bind,source=./container/deps/requirements.test.txt,target=/tmp/requirements.txt \
uv pip install --requirement /tmp/requirements.txt
# ### MISC UTILITY SETUP ###
# Finish pyright install
RUN pyright --help > /dev/null 2>&1
# Enable Git operations in the /workspace directory
RUN printf "[safe]\n directory=/workspace\n" > /root/.gitconfig
RUN ln -sf /bin/bash /bin/sh
### BUILDS ###
# Rust build/dev dependencies
RUN apt update -y && \
apt install --no-install-recommends -y \
build-essential \
protobuf-compiler \
cmake \
libssl-dev \
pkg-config
ENV RUSTUP_HOME=/usr/local/rustup \
CARGO_HOME=/usr/local/cargo \
PATH=/usr/local/cargo/bin:$PATH \
RUST_VERSION=1.87.0
# Define Rust target based on ARCH_ALT ARG
ARG RUSTARCH=${ARCH_ALT}-unknown-linux-gnu
# Install Rust using RUSTARCH derived from ARCH_ALT
RUN wget --tries=3 --waitretry=5 "https://static.rust-lang.org/rustup/archive/1.28.1/${RUSTARCH}/rustup-init" && \
# TODO: Add SHA check back based on RUSTARCH
chmod +x rustup-init && \
./rustup-init -y --no-modify-path --profile minimal --default-toolchain $RUST_VERSION --default-host ${RUSTARCH} && \
rm rustup-init && \
chmod -R a+w $RUSTUP_HOME $CARGO_HOME
ARG CARGO_BUILD_JOBS
# Set CARGO_BUILD_JOBS to 16 if not provided
# This is to prevent cargo from building $(nproc) jobs in parallel,
# which might exceed the number of opened files limit.
ENV CARGO_BUILD_JOBS=${CARGO_BUILD_JOBS:-16}
#######################################
########## Local Development ##########
#######################################
FROM base AS local-dev
# https://code.visualstudio.com/remote/advancedcontainers/add-nonroot-user
# Will use the default ubuntu user, but give sudo access
# Needed so files permissions aren't set to root ownership when writing from inside container
# Don't want ubuntu to be editable, just change uid and gid. User ubuntu is hardcoded in .devcontainer
ENV USERNAME=ubuntu
ARG USER_UID=1000
ARG USER_GID=1000
RUN apt-get update && apt-get install -y sudo gnupg2 gnupg1 \
&& echo "$USERNAME ALL=(root) NOPASSWD:ALL" > /etc/sudoers.d/$USERNAME \
&& chmod 0440 /etc/sudoers.d/$USERNAME \
&& mkdir -p /home/$USERNAME \
&& chown -R $USERNAME:$USERNAME /home/$USERNAME \
&& rm -rf /var/lib/apt/lists/* \
&& chsh -s /bin/bash $USERNAME
# This is a slow operation (~40s on my cpu)
# Much better than chown -R $USERNAME:$USERNAME /opt/dynamo/venv (~10min on my cpu)
COPY --from=base --chown=$USER_UID:$USER_GID /opt/dynamo/venv/ /opt/dynamo/venv/
RUN chown $USERNAME:$USERNAME /opt/dynamo/venv
COPY --from=base --chown=$USERNAME:$USERNAME /usr/local/bin /usr/local/bin
USER $USERNAME
ENV HOME=/home/$USERNAME
ENV PYTHONPATH=$HOME/dynamo/deploy/sdk/src:$PYTHONPATH:$HOME/dynamo/components/planner/src:$PYTHONPATH
WORKDIR $HOME
# https://code.visualstudio.com/remote/advancedcontainers/persist-bash-history
RUN SNIPPET="export PROMPT_COMMAND='history -a' && export HISTFILE=$HOME/.commandhistory/.bash_history" \
&& mkdir -p $HOME/.commandhistory \
&& touch $HOME/.commandhistory/.bash_history \
&& echo "$SNIPPET" >> "$HOME/.bashrc"
RUN mkdir -p /home/$USERNAME/.cache/
ENV VLLM_KV_CAPI_PATH=$HOME/dynamo/.build/target/debug/libdynamo_llm_capi.so
ENTRYPOINT ["/opt/nvidia/nvidia_entrypoint.sh"]
##################################
##### Wheel Build Image ##########
##################################
# Redeclare ARCH_ALT ARG so it's available for interpolation in the FROM instruction
ARG ARCH_ALT
FROM quay.io/pypa/manylinux_2_28_${ARCH_ALT} AS wheel_builder
ARG CARGO_BUILD_JOBS
# Set CARGO_BUILD_JOBS to 16 if not provided
# This is to prevent cargo from building $(nproc) jobs in parallel,
# which might exceed the number of opened files limit.
ENV CARGO_BUILD_JOBS=${CARGO_BUILD_JOBS:-16}
# Use build arg RELEASE_BUILD = true to generate wheels for Python 3.10, 3.11 and 3.12.
ARG RELEASE_BUILD
WORKDIR /workspace
RUN yum update -y \
&& yum install -y llvm-toolset \
&& yum install -y python3.12-devel \
&& yum install -y protobuf-compiler \
&& yum clean all \
&& rm -rf /var/cache/yum
ENV RUSTUP_HOME=/usr/local/rustup \
CARGO_HOME=/usr/local/cargo \
CARGO_TARGET_DIR=/workspace/target \
VIRTUAL_ENV=/opt/dynamo/venv
COPY --from=base $RUSTUP_HOME $RUSTUP_HOME
COPY --from=base $CARGO_HOME $CARGO_HOME
COPY --from=base /usr/local/nixl /opt/nvidia/nvda_nixl
COPY --from=base /workspace /workspace
COPY --from=base $VIRTUAL_ENV $VIRTUAL_ENV
ENV PATH=$CARGO_HOME/bin:$VIRTUAL_ENV/bin:$PATH
# Copy configuration files
COPY pyproject.toml /workspace/
COPY README.md /workspace/
COPY LICENSE /workspace/
COPY Cargo.toml /workspace/
COPY Cargo.lock /workspace/
COPY rust-toolchain.toml /workspace/
COPY hatch_build.py /workspace/
# Copy source code
COPY lib/ /workspace/lib/
COPY components /workspace/components
COPY launch /workspace/launch
COPY deploy/sdk /workspace/deploy/sdk
# Build Rust crate binaries packaged with the wheel
RUN cargo build --release --locked --features mistralrs,python,dynamo-llm/block-manager \
-p dynamo-run \
-p llmctl \
# Multiple http named crates are present in dependencies, need to specify the path
-p file://$PWD/components/http \
-p metrics
# Build dynamo wheel
RUN uv build --wheel --out-dir /workspace/dist && \
cd /workspace/lib/bindings/python && \
uv pip install maturin[patchelf] && \
maturin build --release --features block-manager --out /workspace/dist && \
if [ "$RELEASE_BUILD" = "true" ]; then \
uv run --python 3.11 maturin build --release --features block-manager --out /workspace/dist && \
uv run --python 3.10 maturin build --release --features block-manager --out /workspace/dist; \
fi
#######################################
########## CI Minimum Image ###########
#######################################
FROM base AS ci_minimum
ENV DYNAMO_HOME=/workspace
ENV CARGO_TARGET_DIR=/workspace/target
WORKDIR /workspace
COPY --from=wheel_builder /workspace/dist/ /workspace/dist/
COPY --from=wheel_builder /workspace/target/ /workspace/target/
COPY --from=wheel_builder /opt/nvidia/nvda_nixl /opt/nvidia/nvda_nixl
# Copy Cargo cache to avoid re-downloading dependencies
COPY --from=wheel_builder $CARGO_HOME $CARGO_HOME
COPY . /workspace
# Build rest of the crates
# Need to figure out rust caching to avoid rebuilding and remove exclude flags
RUN cargo build --release --locked --features block-manager --workspace \
--exclude dynamo-run \
--exclude llmctl \
--exclude file://$PWD/components/http \
--exclude metrics
# Package the bindings
RUN mkdir -p /opt/dynamo/bindings/wheels && \
mkdir /opt/dynamo/bindings/lib && \
cp dist/ai_dynamo*cp312*.whl /opt/dynamo/bindings/wheels/. && \
cp target/release/libdynamo_llm_capi.so /opt/dynamo/bindings/lib/. && \
cp -r lib/bindings/c/include /opt/dynamo/bindings/. && \
cp target/release/dynamo-run /usr/local/bin && \
cp target/release/http /usr/local/bin && \
cp target/release/llmctl /usr/local/bin && \
cp target/release/metrics /usr/local/bin && \
cp target/release/mock_worker /usr/local/bin
RUN uv pip install /workspace/dist/ai_dynamo_runtime*cp312*.whl && \
uv pip install /workspace/dist/ai_dynamo*any.whl
# Copy launch banner
RUN --mount=type=bind,source=./container/launch_message.txt,target=/workspace/launch_message.txt \
sed '/^#\s/d' /workspace/launch_message.txt > ~/.launch_screen && \
echo "cat ~/.launch_screen" >> ~/.bashrc
# Tell vllm to use the Dynamo LLM C API for KV Cache Routing
ENV VLLM_KV_CAPI_PATH=/opt/dynamo/bindings/lib/libdynamo_llm_capi.so
ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/nvidia/nvda_nixl/lib/x86_64-linux-gnu/
##########################################
########## Perf Analyzer Image ###########
##########################################
FROM ${BASE_IMAGE}:${BASE_IMAGE_TAG} AS perf_analyzer
ARG GENAI_PERF_TAG
WORKDIR /workspace
# Build and install Perf Analyzer for benchmarking
RUN apt-get update -y && apt-get -y install cmake g++ libssl-dev python3 rapidjson-dev zlib1g-dev
RUN git clone https://github.com/triton-inference-server/perf_analyzer.git
RUN git -C perf_analyzer checkout ${GENAI_PERF_TAG}
RUN mkdir perf_analyzer/build
RUN cmake -B perf_analyzer/build -S perf_analyzer -D TRITON_ENABLE_PERF_ANALYZER_OPENAI=ON
RUN cmake --build perf_analyzer/build -- -j8
RUN mkdir bin && \
cp -r perf_analyzer/build/perf_analyzer/src/perf-analyzer-build /workspace/bin/
########################################
########## Development Image ###########
########################################
FROM ci_minimum AS dev
ARG GENAI_PERF_TAG
COPY --from=perf_analyzer /workspace/bin/perf-analyzer-build/ /perf/bin
COPY --from=perf_analyzer /workspace/perf_analyzer /perf_analyzer
ENV PATH="/perf/bin:${PATH}"
# Install genai-perf for benchmarking
RUN uv pip install "git+https://github.com/triton-inference-server/perf_analyzer.git@${GENAI_PERF_TAG}#subdirectory=genai-perf"
RUN uv pip uninstall tritonclient
COPY . /workspace
ENTRYPOINT ["/opt/nvidia/nvidia_entrypoint.sh"]
CMD []
####################################
########## Runtime Image ###########
####################################
FROM ${RUNTIME_IMAGE}:${RUNTIME_IMAGE_TAG} AS runtime
WORKDIR /workspace
ENV DYNAMO_HOME=/workspace
ENV VIRTUAL_ENV=/opt/dynamo/venv
ENV PATH="${VIRTUAL_ENV}/bin:${PATH}"
# Setup the python environment
COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/
RUN apt-get update && \
DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends python3-dev && \
rm -rf /var/lib/apt/lists/* && \
uv venv $VIRTUAL_ENV --python 3.12 && \
echo "source $VIRTUAL_ENV/bin/activate" >> ~/.bashrc
# Install the wheels and symlink executables to /usr/local/bin so dynamo components can use them
# Dynamo components currently do not have the VIRTUAL_ENV in their PATH, so we need to symlink the executables
COPY --from=wheel_builder /workspace/dist/*.whl wheelhouse/
RUN uv pip install ai-dynamo[vllm] --find-links wheelhouse && \
ln -sf $VIRTUAL_ENV/bin/* /usr/local/bin/ && \
rm -r wheelhouse
# Tell vllm to use the Dynamo LLM C API for KV Cache Routing
ENV VLLM_KV_CAPI_PATH="/opt/dynamo/bindings/lib/libdynamo_llm_capi.so"
# Copy launch banner
RUN --mount=type=bind,source=./container/launch_message.txt,target=/workspace/launch_message.txt \
sed '/^#\s/d' /workspace/launch_message.txt > ~/.launch_screen && \
echo "cat ~/.launch_screen" >> ~/.bashrc
# Copy examples
COPY ./examples examples/
ENTRYPOINT [ "/usr/bin/bash" ]
CMD []
container/build.sh
View file @ 5c5cec3d
......@@ -49,7 +49,7 @@ PYTHON_PACKAGE_VERSION=${current_tag:-$latest_tag.dev+$commit_id}
# dependencies are specified in the /container/deps folder and
# installed within framework specific sections of the Dockerfile.
declare -A FRAMEWORKS=(["VLLM"]=1 ["TENSORRTLLM"]=2 ["NONE"]=3)
declare -A FRAMEWORKS=(["VLLM"]=1 ["TENSORRTLLM"]=2 ["NONE"]=3, ["SGLANG"]=4)
DEFAULT_FRAMEWORK=VLLM
SOURCE_DIR=$(dirname "$(readlink -f "$0")")
......@@ -106,6 +106,9 @@ VLLM_BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"
NONE_BASE_IMAGE="ubuntu"
NONE_BASE_IMAGE_TAG="24.04"
SGLANG_BASE_IMAGE="nvcr.io/nvidia/cuda-dl-base"
SGLANG_BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"
NIXL_COMMIT=78695c2900cd7fff506764377386592dfc98e87e
NIXL_REPO=ai-dynamo/nixl.git
......@@ -370,6 +373,8 @@ elif [[ $FRAMEWORK == "TENSORRTLLM" ]]; then
DOCKERFILE=${SOURCE_DIR}/Dockerfile.tensorrt_llm
elif [[ $FRAMEWORK == "NONE" ]]; then
DOCKERFILE=${SOURCE_DIR}/Dockerfile.none
elif [[ $FRAMEWORK == "SGLANG" ]]; then
DOCKERFILE=${SOURCE_DIR}/Dockerfile.sglang
fi
NIXL_DIR="/tmp/nixl/nixl_src"
......@@ -514,4 +519,4 @@ if [ -z "$RUN_PREFIX" ]; then
set -x
fi
{ set +x; } 2>/dev/null
{ set +x; } 2>/dev/null
\ No newline at end of file
container/run.sh
View file @ 5c5cec3d
......@@ -24,7 +24,7 @@ RUN_PREFIX=
# dependencies are specified in the /container/deps folder and
# installed within framework specific sections of the Dockerfile.
declare -A FRAMEWORKS=(["VLLM"]=1 ["TENSORRTLLM"]=2)
declare -A FRAMEWORKS=(["VLLM"]=1 ["TENSORRTLLM"]=2 ["SGLANG"]=3)
DEFAULT_FRAMEWORK=VLLM
SOURCE_DIR=$(dirname "$(readlink -f "$0")")
......
examples/sglang/README.md
View file @ 5c5cec3d
......@@ -19,12 +19,9 @@ limitations under the License.
This directory contains examples and reference implementations for deploying Large Language Models (LLMs) in various configurations using SGLang. SGLang internally uses ZMQ to communicate between the ingress and the engine processes. For Dynamo, we leverage the runtime to communicate directly with the engine processes and handle ingress and pre/post processing on our end.
> [!IMPORTANT]
> In order to run these examples, you will need to install sglang using `uv pip install "sglang[all]>=0.4.6.post4"`. Additionally, SGLang currently does not have pre-built wheels for ARM. If you are on an ARM machine - you will need to install SGLang from source.
## Deployment Architectures
See [deployment architectures](../llm/README.md#deployment-architectures) to learn about the general idea of the architecture. SGLang currently support only aggregated serving but routing and disaggregation support are coming very soon!
See [deployment architectures](../llm/README.md#deployment-architectures) to learn about the general idea of the architecture. SGLang currently supports aggregated and disaggregated serving. KV routing support is coming soon!
## Getting Started
......@@ -44,13 +41,13 @@ docker compose -f deploy/docker-compose.yml up -d
```bash
# On an x86 machine - sglang does not support ARM yet
./container/build.sh
./container/build.sh --framework sglang
```
### Run container
```bash
./container/run.sh -it
./container/run.sh -it --framework sglang
```
### Example architectures
......
examples/sglang/multinode-examples.md 0 → 100644
View file @ 5c5cec3d
# Multinode Examples
## Multi-node sized models
SGLang allows you to deploy multi-node sized models by adding in the `dist-init-addr`, `nnodes`, and `node-rank` arguments. Below we demonstrate and example of deploying DeepSeek R1 for disaggregated serving across 4 nodes. This example requires
4 nodes of 8xH100 GPUs.
**Step 1**: Start NATS/ETCD on your head node. Ensure you have the correct firewall rules to allow communication between the nodes as you will need the NATS/ETCD endpoints to be accessible by all other nodes.
```bash
# node 1
docker compose -f lib/runtime/docker-compose.yml up -d
```
**Step 2**: Ensure that your configuration file has the required arguments. Here's an example configuration that runs prefill and the model in TP16:
Node 1: Run HTTP ingress, processor, and 8 shards of the prefill worker
```yaml
# configs/prefill-1.yaml
Frontend:
served_model_name: deepseek-ai/DeepSeek-R1
endpoint: dynamo.SGLangWorker.generate
port: 8000
SGLangWorker:
model-path: deepseek-ai/DeepSeek-R1
served-model-name: deepseek-ai/DeepSeek-R1
tp: 16
trust-remote-code: true
skip-tokenizer-init: true
dist-init-addr: <node-1-ip>:29500
disaggregation-bootstrap-port: 30001
disaggregation-mode: prefill
disaggregation-transfer-backend: nixl
nnodes: 2
node-rank: 0
mem-fraction-static: 0.82
ServiceArgs:
workers: 1
resources:
gpu: 8
```
Run this with:
```bash
cd examples/sglang
dynamo serve graphs.agg:Frontend -f configs/prefill-1.yaml
```
Node 2: Run the remaining 8 shards of the prefill worker and the decode worker
```yaml
# configs/prefill-2.yaml
SGLangWorker:
model-path: deepseek-ai/DeepSeek-R1
served-model-name: deepseek-ai/DeepSeek-R1
tp: 16
trust-remote-code: true
skip-tokenizer-init: true
mem-fraction-static: 0.82
dist-init-addr: <node-1-ip>:29500
disaggregation-bootstrap-port: 30001
disaggregation-mode: prefill
disaggregation-transfer-backend: nixl
nnodes: 2
node-rank: 1
ServiceArgs:
workers: 1
resources:
gpu: 8
```
On all other nodes, we need to export the NATS and ETCD endpoints. Run this with with:
```bash
export NATS_SERVER="nats://<node-1-ip>"
export ETCD_ENDPOINTS="<node-1-ip>:2379"
cd examples/sglang
dynamo serve graphs.disagg:Frontend -f configs/prefill-2.yaml --service-name SGLangWorker
```
Node 3: Run the first 8 shards of the decode worker
```yaml
# configs/decode-1.yaml
SGLangDecodeWorker:
model-path: deepseek-ai/DeepSeek-R1
served-model-name: deepseek-ai/DeepSeek-R1
tp: 16
trust-remote-code: true
skip-tokenizer-init: true
mem-fraction-static: 0.80
dist-init-addr: 2:29500
disaggregation-mode: decode
disaggregation-transfer-backend: nixl
disaggregation-bootstrap-port: 30001
nnodes: 2
node-rank: 0
ServiceArgs:
workers: 1
resources:
gpu: 8
```
Run this with:
```bash
export NATS_SERVER="nats://<node-1-ip>"
export ETCD_ENDPOINTS="<node-1-ip>:2379"
cd examples/sglang
dynamo serve graphs.disagg:Frontend -f configs/decode-1.yaml --service-name SGLangDecodeWorker
```
Node 4: Run the remaining 8 shards of the decode worker
```yaml
# configs/decode-2.yaml
SGLangDecodeWorker:
model-path: deepseek-ai/DeepSeek-R1
served-model-name: deepseek-ai/DeepSeek-R1
tp: 16
trust-remote-code: true
skip-tokenizer-init: true
mem-fraction-static: 0.80
dist-init-addr: 2:29500
disaggregation-mode: decode
disaggregation-transfer-backend: nixl
disaggregation-bootstrap-port: 30001
disable-cuda-graph: true
nnodes: 2
node-rank: 1
ServiceArgs:
workers: 1
resources:
gpu: 8
```
Run this with:
```bash
export NATS_SERVER="nats://<node-1-ip>"
export ETCD_ENDPOINTS="<node-1-ip>:2379"
cd examples/sglang
dynamo serve graphs.disagg:Frontend -f configs/decode-2.yaml --service-name SGLangDecodeWorker
```
**Step 3**: Run inference
SGLang typically requires a warmup period to ensure the DeepGEMM kernels are loaded. We recommend running a few warmup requests and ensuring that the DeepGEMM kernels load in.
```bash
curl <node-1-ip>:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
"messages": [
{
"role": "user",
"content": "In the heart of the tennis world, where champions rise and fall with each Grand Slam, lies the legend of the Golden Racket of Wimbledon. Once wielded by the greatest players of antiquity, this mythical racket is said to bestow unparalleled precision, grace, and longevity upon its rightful owner. For centuries, it remained hidden, its location lost to all but the most dedicated scholars of the sport. You are Roger Federer, the Swiss maestro whose elegant play and sportsmanship have already cemented your place among the legends, but whose quest for perfection remains unquenched even as time marches on. Recent dreams have brought you visions of this ancient artifact, along with fragments of a map that seems to lead to its resting place. Your journey will take you through the hallowed grounds of tennis history, from the clay courts of Roland Garros to the hidden training grounds of forgotten champions, and finally to a secret chamber beneath Centre Court itself. Character Background: Develop a detailed background for Roger Federer in this quest. Describe his motivations for seeking the Golden Racket, his tennis skills and personal weaknesses, and any connections to the legends of the sport that came before him. Is he driven by a desire to extend his career, to secure his legacy as the greatest of all time, or perhaps by something more personal? What price might he be willing to pay to claim this artifact, and what challenges from rivals past and present might stand in his way?"
}
],
"stream":false,
"max_tokens": 30
}'
```
pyproject.toml
View file @ 5c5cec3d
......@@ -66,6 +66,10 @@ vllm = [
"ai-dynamo-vllm~=0.8.4"
]
sglang = [
"sglang[all]@git+https://github.com/sgl-project/sglang@4d643f6c7a291c86de64a9e52eca526b2d99775d#subdirectory=python"
]
[project.scripts]
dynamo = "dynamo.sdk.cli.cli:cli"
dynamo-run = "dynamo.sdk.cli.run_executable:dynamo_run"
......@@ -87,6 +91,9 @@ packages = ["deploy/sdk/src/dynamo", "components/planner/src/dynamo"]
[tool.hatch.build.hooks.custom]
path = "hatch_build.py"
[tool.hatch.metadata]
allow-direct-references = true
[tool.codespell]
# note: pre-commit passes explicit lists of files here, which this skip file list doesn't override -
# this is only to allow you to run codespell interactively
......@@ -205,4 +212,4 @@ ignore_missing_imports = true
# '''
# <script type="text/javascript">if (typeof _satellite !== "undefined") {_satellite.pageBottom();}</script>
# ''',
#]
\ No newline at end of file
#]
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment