Deployment contract
Production rollout checklist
Move from a runnable model to an observable, scalable, reversible production service with explicit boundaries.
Workload boundary
Decide whether it fits your production constraints
Good fit
- Preparing a launch or scale-up
- Capacity, readiness, and security boundaries are required
- Auditable release and rollback are required
Not a fit
- Using one demo result as a capacity test
- Omitting gateway and access control
- Comparing performance without version, model, and hardware records
| Models | All supported deployment models |
|---|---|
| Hardware | cpu / nvidia-gpu / desktop-edge-gpu / kubernetes |
| Operating systems | Linux / macOS / Windows / Kubernetes |
| Interfaces | HTTP / WebSocket / CLI / C++ |
Run path
From installation to known-audio verification
These commands come from the current verification registry. Pin dependencies, model, and hardware before rollout.
Install
git clone https://github.com/modelscope/FunASR.gitLaunch
Run the selected runtime with pinned versions and an immutable model cacheHealth check
Verify readiness, model identity, and one known-audio transcriptSmoke test
Minimum verification before promotion
Record version, model digest, hardware, warmup, audio duration, concurrency, latency, and output before promotionOperations and capacity
Move from runnable to operable
Operational checks
- Separate liveness, readiness, and business smoke tests
- Include warmup, queueing, and audio distribution in capacity tests
- Keep model, configuration, and previous-version rollback points
Capacity variables
- Audio duration, language, channels, and VAD segment distribution
- Concurrency, queue time, warmup, and model-cache state
- Exact hardware, driver, runtime, and thread configuration
Troubleshooting
- Confirm exact model and runtime versions first
- Use known audio to separate infrastructure and quality regressions
- Roll back before experimenting during an incident
Security boundary
Treat production ingress as untrusted
- Default-deny network access and grant minimum privileges
- Define retention for uploads, logs, caches, and temporary files
- Keep credentials, audio, and presigned URLs out of normal logs
Known limitation
A checklist does not replace workload, accuracy, failure, and security exercises for the target product.
Public benchmarks are reproduction starting points, not substitutes for target-workload testing.
Evidence and feedback