sglang-worker-minimax-build

MiniMax-H3 Ref2VA on RunPod Serverless

Queue-based RunPod worker for MiniMax-H3 reference-to-video-and-audio generation. It starts a private SGLang server only after resolving a pre-cached Hugging Face snapshot, stages reference media with SSRF checks, generates one video at a time, validates the MP4, uploads it to S3-compatible storage, and returns a presigned URL.

Performance profile

The production default is resident BF16/FP32 Ref2VA with Ulysses-8 on 8×H200. It deliberately disables Torch compile and warms 1344x768, matching the validated lossless path. Breakable CUDA graphs remain opt-in because the published recipe reports no measured speedup for the captured Ref2VA signature.

The linked SGLang selector's 8×B200 FL2VA/T2VA profile is a different workload. It has much lower warmed latency, but it cannot accept reference conditions and should be deployed as a separate endpoint.

Required endpoint configuration

Non-secret variables:

MODEL_ID=<private-ref2va-only-repo>
MODEL_REVISION=<immutable-commit>
MODEL_VARIANT=ref2va
HF_CACHE_ROOT=/runpod-volume/huggingface-cache/hub
SGLANG_PORT=30010
NUM_GPUS=8
ULYSSES_DEGREE=8
WARMUP_RESOLUTIONS=1344x768
RUNPOD_INIT_TIMEOUT=1200
GENERATION_TIMEOUT_SECONDS=840
POLL_INTERVAL_SECONDS=2
MAX_REFERENCE_BYTES=536870912
MAX_PROMPT_CHARS=6000
ALLOWED_REFERENCE_HOSTS=<comma-separated-storage-hosts>
S3_REGION=auto
S3_BUCKET=<bucket>
S3_OUTPUT_PREFIX=generations
OUTPUT_URL_TTL_SECONDS=3600

Secrets:

HF_TOKEN=<read-only-token-for-cached-model>
S3_ENDPOINT_URL=<s3-or-r2-endpoint>
S3_ACCESS_KEY_ID=<least-privilege-key>
S3_SECRET_ACCESS_KEY=<least-privilege-secret>

The RunPod key is used by the caller and deployment tooling only. It is never passed to the worker.

Local verification

python3 -m venv .venv
.venv/bin/pip install -r requirements-dev.txt
.venv/bin/pytest -q
.venv/bin/ruff check worker tests

Build

docker buildx build --platform linux/amd64 \
  -t <registry>/sglang-worker-minimax:<version> \
  --push .

The image contains SGLang and worker dependencies, but no model weights or credentials.

Request

Submit examples/ref2va-job.json to POST https://api.runpod.ai/v2/<endpoint-id>/run. The top-level policy.executionTimeout and policy.ttl values use milliseconds, per RunPod's queue API.

Deployment settings

• Queue endpoint, H200 only, 8 GPUs per worker
• Active workers 0, max workers 1, handler concurrency 1
• Queue-delay autoscaling at 2 seconds
• FlashBoot enabled, 600-second idle timeout
• 900-second execution timeout and 3600-second job TTL
• No network volume; attach the private Ref2VA-only repository as the cached model