LLM Deployment using TensorRT-LLM#

This directory contains examples and reference implementations for deploying Large Language Models (LLMs) in various configurations using TensorRT-LLM.

Use the Latest Release#

We recommend using the latest stable release of dynamo to avoid breaking changes:

You can find the latest release here and check out the corresponding branch with:

git checkout $(git describe --tags $(git rev-list --tags --max-count=1))

Feature Support Matrix#

Core Dynamo Features#

Feature	TensorRT-LLM	Notes
Disaggregated Serving	✅
Conditional Disaggregation	🚧	Not supported yet
KV-Aware Routing	✅
SLA-Based Planner	🚧	Planned
Load Based Planner	🚧	Planned
KVBM	🚧	Planned

Large Scale P/D and WideEP Features#

Feature	TensorRT-LLM	Notes
WideEP	✅
Attention DP	✅
GB200 Support	✅

Quick Start#

Below we provide a guide that lets you run all of our the common deployment patterns on a single node.

Start NATS and ETCD in the background#

Start using Docker Compose

docker compose -f deploy/docker-compose.yml up -d

Build container#

# TensorRT-LLM uses git-lfs, which needs to be installed in advance.
apt-get update && apt-get -y install git git-lfs

# On an x86 machine:
./container/build.sh --framework tensorrtllm

# On an ARM machine:
./container/build.sh --framework tensorrtllm --platform linux/arm64

# Build the container with the default experimental TensorRT-LLM commit
# WARNING: This is for experimental feature testing only.
# The container should not be used in a production environment.
./container/build.sh --framework tensorrtllm --use-default-experimental-tensorrtllm-commit

Run container#

./container/run.sh --framework tensorrtllm -it

Single Node Examples#

Important

Below we provide some simple shell scripts that run the components for each configuration. Each shell script is simply running the python3 -m dynamo.frontend <args> to start up the ingress and using python3 -m dynamo.trtllm <args> to start up the workers. You can easily take each command and run them in separate terminals.

This figure shows an overview of the major components to deploy:

+------+      +-----------+      +------------------+             +---------------+
| HTTP |----->| processor |----->|      Worker1     |------------>|    Worker2    |
|      |<-----|           |<-----|                  |<------------|               |
+------+      +-----------+      +------------------+             +---------------+
                  |    ^                  |
       query best |    | return           | publish kv events
           worker |    | worker_id        v
                  |    |         +------------------+
                  |    +---------|     kv-router    |
                  +------------->|                  |
                                 +------------------+

Note: The diagram above shows all possible components in a deployment. Depending on the chosen disaggregation strategy, you can configure whether Worker1 handles prefill and Worker2 handles decode, or vice versa. For more information on how to select and configure these strategies, see the Disaggregation Strategy section below.

Aggregated#

cd $DYNAMO_HOME/components/backends/trtllm
./launch/agg.sh

Aggregated with KV Routing#

cd $DYNAMO_HOME/components/backends/trtllm
./launch/agg_router.sh

Disaggregated#

Important

Disaggregated serving supports two strategies for request flow: "prefill_first" and "decode_first". By default, the script below uses the "decode_first" strategy, which can reduce response latency by minimizing extra hops in the return path. You can switch strategies by setting the DISAGGREGATION_STRATEGY environment variable.

cd $DYNAMO_HOME/components/backends/trtllm
./launch/disagg.sh

Disaggregated with KV Routing#

Important

Disaggregated serving with KV routing uses a “prefill first” workflow by default. Currently, Dynamo supports KV routing to only one endpoint per model. In disaggregated workflow, it is generally more effective to route requests to the prefill worker. If you wish to use a “decode first” workflow instead, you can simply set the DISAGGREGATION_STRATEGY environment variable accordingly.

cd $DYNAMO_HOME/components/backends/trtllm
./launch/disagg_router.sh

Aggregated with Multi-Token Prediction (MTP) and DeepSeek R1#

cd $DYNAMO_HOME/components/backends/trtllm

export AGG_ENGINE_ARGS=./engine_configs/deepseek_r1/mtp/mtp_agg.yaml
export SERVED_MODEL_NAME="nvidia/DeepSeek-R1-FP4"
# nvidia/DeepSeek-R1-FP4 is a large model
export MODEL_PATH="nvidia/DeepSeek-R1-FP4"
./launch/agg.sh

Notes:

MTP is only available within the container built with the experimental TensorRT-LLM commit. Please add –use-default-experimental-tensorrtllm-commit to the arguments of the build.sh script.

Example: ./container/build.sh --framework tensorrtllm --use-default-experimental-tensorrtllm-commit
There is a noticeable latency for the first two inference requests. Please send warm-up requests before starting the benchmark.
MTP performance may vary depending on the acceptance rate of predicted tokens, which is dependent on the dataset or queries used while benchmarking. Additionally, ignore_eos should generally be omitted or set to false when using MTP to avoid speculating garbage outputs and getting unrealistic acceptance rates.

Advanced Examples#

Below we provide a selected list of advanced examples. Please open up an issue if you’d like to see a specific example!

Multinode Deployment#

For comprehensive instructions on multinode serving, see the multinode-examples.md guide. It provides step-by-step deployment examples and configuration tips for running Dynamo with TensorRT-LLM across multiple nodes. While the walkthrough uses DeepSeek-R1 as the model, you can easily adapt the process for any supported model by updating the relevant configuration files. You can see Llama4+eagle guide to learn how to use these scripts when a single worker fits on the single node.

Speculative Decoding#

Llama 4 Maverick Instruct + Eagle Speculative Decoding

Kubernetes Deployment#

For complete Kubernetes deployment instructions, configurations, and troubleshooting, see TensorRT-LLM Kubernetes Deployment Guide

Client#

To send a request to a multi-node deployment, target the node which is running python3 -m dynamo.frontend <args>.

Benchmarking#

To benchmark your deployment with GenAI-Perf, see this utility script, configuring the model name and host based on your deployment:

{REPO_ROOT}/benchmarks/llm/perf.sh

Disaggregation Strategy#

The disaggregation strategy controls how requests are distributed between the prefill and decode workers in a disaggregated deployment.

By default, Dynamo uses a decode first strategy: incoming requests are initially routed to the decode worker, which then forwards them to the prefill worker in round-robin fashion. The prefill worker processes the request and returns results to the decode worker for any remaining decode operations.

When using KV routing, however, Dynamo switches to a prefill first strategy. In this mode, requests are routed directly to the prefill worker, which can help maximize KV cache reuse and improve overall efficiency for certain workloads. Choosing the appropriate strategy can have a significant impact on performance, depending on your use case.

The disaggregation strategy can be set using the DISAGGREGATION_STRATEGY environment variable. You can set the strategy before launching your deployment, for example:

DISAGGREGATION_STRATEGY="prefill_first" ./launch/disagg.sh

KV Cache Transfer in Disaggregated Serving#

Dynamo with TensorRT-LLM supports two methods for transferring KV cache in disaggregated serving: UCX (default) and NIXL (experimental). For detailed information and configuration instructions for each method, see the KV cache transfer guide.

Request Migration#

In a Distributed System, a request may fail due to connectivity issues between the Frontend and the Backend.

The Frontend will automatically track which Backends are having connectivity issues with it and avoid routing new requests to the Backends with known connectivity issues.

For ongoing requests, there is a --migration-limit flag which can be set on the Backend that tells the Frontend how many times a request can be migrated to another Backend should there be a loss of connectivity to the current Backend.

For example,

python3 -m dynamo.trtllm ... --migration-limit=3

indicates a request to this model may be migrated up to 3 times to another Backend, before failing the request, should the Frontend detects a connectivity issue to the current Backend.

The migrated request will continue responding to the original request, allowing for a seamless transition between Backends, and a reduced overall request failure rate at the Frontend for enhanced user experience.

Client#

NOTE: To send a request to a multi-node deployment, target the node which is running python3 -m dynamo.frontend <args>.