Deploy frontier AI fashions on AWS Trainium with SGLang-Neuron

In the present day, Amazon introduced that SGLang-Neuron is publicly out there and could be accessed by an AWS Deep Studying Container, launched by RadixArk in collaboration with Annapurna Labs. With SGLang-Neuron, you possibly can serve fashions and generate reinforcement studying (RL) rollouts with SGLang on AWS Trainium utilizing the identical SGLang interface you employ in the present day. Trainium is a portfolio of AWS AI chips that delivers sturdy economics for AI coaching and inference at scale. Now, operating SGLang on Trainium follows the SGLang workflow you already know: you launch fashions by SGLang’s server interface and invoke the endpoint utilizing the identical technology APIs. Beneath that interface, SGLang-Neuron runs on native PyTorch. With its out-of-tree plugin design, Trainium assist and SGLang core develop independently and iterate sooner.

How SGLang runs on Trainium

SGLang-Neuron is an out-of-tree plugin that connects SGLang with AWS Trainium. Putting in the plugin registers the neuron platform by SGLang’s plugin entry factors, and setting SGLANG_PLATFORM=neuron selects it when the server begins. SGLang core continues to supply the API server, request lifecycle, RadixAttention, and prefix-cache bookkeeping. The plugin implements the Trainium-specific fashions, consideration, combination of consultants (MoE), and the graph runner.

Architecture of SGLang-Neuron on AWS Trainium, showing SGLang core and the Trainium plugin components

Determine 1: SGLang-Neuron on AWS Trainium

TorchNeuron registers Trainium as a local PyTorch machine. SGLang-Neuron makes use of customary PyTorch operations for the mannequin execution path and incorporates kernels written with the Neuron Kernel Interface (NKI) for efficiency. It compiles these operations and kernels into executable packages utilizing torch.compile, which the graph runner prepares and executes throughout serving.

SGLang-Neuron additionally helps PD disaggregation. On this configuration, a prefill server computes the immediate KV cache and transfers it to a decode server utilizing NIXL over Elastic Cloth Adapter (EFA). The decode server then generates the remaining tokens. Every server nonetheless makes use of SGLang’s scheduler and request lifecycle, whereas the plugin manages the Trainium execution and KV-cache switch path.

Serving fashions on AWS Trainium utilizing SGLang-Neuron

To comply with this walkthrough, you employ a Trainium3 Amazon Elastic Compute Cloud (Amazon EC2) occasion with a pytorch-inference-sglang-neuronx container. Listed below are the steps:

Begin the container.

docker run --rm -it 
--privileged --ipc=host 
-v /path/to/work:/work 
public.ecr.aws/neuron/pytorch-inference-sglang-neuron:v0.1.0-neuron-py314-sdk-2.33.0-ubuntu26.04 
/bin/bash

Launch the SGLang mannequin server contained in the container.

SGLANG_PLATFORM=neuron 
python -m sglang.launch_server 
--model-path openai/gpt-oss-120b 
--tp-size 8 
--ep-size 8 
--kv-cache-dtype fp8_e4m3 
--context-length 10240 
--chunked-prefill-size 8192 
--max-running-requests 64 
--max-total-tokens 655360 
--mem-fraction-static 0.8 
--host 127.0.0.1 
--port 30000

Ship a request to the mannequin server.

curl -s http://127.0.0.1:30000/v1/chat/completions 
-H 'Content material-Sort: software/json' 
-d '{
"mannequin": "openai/gpt-oss-120b",
"messages": [
{"role": "user", "content": "Explain why the sky appears blue."}
],
"temperature": 0,
"max_tokens": 128
}'

For added data, please see the SGLang-Neuron getting began information for setup directions, and discover the repository documentation for supported fashions, options, and validated configurations.

Conclusion

SGLang-Neuron brings SGLang’s acquainted serving workflow to AWS Trainium by native PyTorch. You will get began by launching the SGLang-Neuron container and following the quickstart to serve a supported mannequin. Strive it together with your workloads, contribute by the venture repository, and share suggestions to assist form future mannequin and have assist.