Additional ResourcesBackend DetailsSGLang

Enable SGLang Hierarchical Cache (HiCache)

View as Markdown

This guide shows how to enable SGLang’s Hierarchical Cache (HiCache) inside Dynamo.

1) Start the SGLang worker with HiCache enabled

$python -m dynamo.sglang \
>  --model-path Qwen/Qwen3-0.6B \
>  --host 0.0.0.0 --port 8000 \
>  --page-size 64 \
>  --enable-hierarchical-cache \
>  --hicache-ratio 2 \
>  --hicache-write-policy write_through \
>  --hicache-storage-backend nixl \
>  --log-level debug \
>  --skip-tokenizer-init
  • —enable-hierarchical-cache: Enables hierarchical KV cache/offload
  • —hicache-ratio: The ratio of the size of host KV cache memory pool to the size of device pool. Lower this number if your machine has less CPU memory.
  • —hicache-write-policy: Write policy (e.g., write_through for synchronous host writes)
  • —hicache-storage-backend: Host storage backend for HiCache (e.g., nixl). NIXL selects the concrete store automatically; see PR #8488

Then, start the frontend:

$python -m dynamo.frontend --http-port 8000

2) Send a single request

$curl localhost:8000/v1/chat/completions \
>  -H "Content-Type: application/json" \
>  -d '{
>    "model": "Qwen/Qwen3-0.6B",
>    "messages": [
>      {
>        "role": "user",
>        "content": "Explain why Roger Federer is considered one of the greatest tennis players of all time"
>      }
>    ],
>    "stream": false,
>    "max_tokens": 30
>  }'

3) (Optional) Benchmarking

Run the perf script:

$bash -x $DYNAMO_ROOT/benchmarks/llm/perf.sh \
>  --model Qwen/Qwen3-0.6B \
>  --tensor-parallelism 1 \
>  --data-parallelism 1 \
>  --concurrency "2,4,8" \
>  --input-sequence-length 2048 \
>  --output-sequence-length 256