TGI support
TGI support spans three layers: open-source community help, Hugging Face enterprise offerings, and your own platform team’s runbooks. Know which layer owns your incident before you page someone.
Support paths
| Path | Best for | Response model |
|---|---|---|
| GitHub Issues | OSS bugs, feature requests, community questions | Community + HF maintainers |
| HF Enterprise | Production SLAs, security reviews, dedicated support | Contracted |
| Inference Endpoints | Managed TGI without running K8s | HF ops team |
| Your platform team | Internal deploys, ingress, guardrails | Internal on-call |
Hugging Face Inference Endpoints
Inference Endpoints is managed TGI on AWS, GCP, and Azure. You pick a model, instance type, and region — HF runs the server, upgrades, and scaling.
| Factor | Self-hosted TGI | Inference Endpoints |
|---|---|---|
| Ops burden | You own K8s, GPUs, upgrades | HF owns infrastructure |
| Cost model | GPU-hours + engineer time | Per-hour instance pricing |
| Customization | Full flag control | Dashboard-configured options |
| Compliance | Your cluster, your rules | HF-managed environment |
Meter managed endpoints like cloud APIs. Self-hosted TGI meters GPU-hours. Compare both on cost per resolved outcome, not sticker price.
Enterprise support
HF Enterprise adds:
- Dedicated support channels and SLAs
- Security and compliance review for model serving
- Priority access to TGI releases and roadmap input
Engage enterprise support when TGI is on your production critical path and GitHub Issues response times are too slow for your SLOs.
HF token management
Gated models require a Hugging Face token on every environment that pulls weights:
export HF_TOKEN=hf_...| Environment | Pattern |
|---|---|
| Local Docker | -e HF_TOKEN on docker run |
| Kubernetes | Secret mounted as env var — see Kubernetes |
| CI | Short-lived token in CI secrets; rotate quarterly |
Failure mode: gated model pulls fail silently in CI when tokens expire. Add a smoke test that verifies model access on every deploy.
Version pinning
Pin TGI container images in production — latest breaks API paths and flag names without warning:
image: ghcr.io/huggingface/text-generation-inference:2.4.0Check the release notes before upgrading. Test chat completions compatibility on your TGI host in staging.
When to escalate
| Symptom | First check | Escalate to |
|---|---|---|
| OOM on all replicas | Server flags — quantize or shrink input | Platform team |
| API 404 after upgrade | API version differences | Pin image, file GitHub Issue |
| Hub download failures | Token expiry, rate limits | Rotate HF_TOKEN |
| Quality regression | Quantization or batch config | Re-benchmark vs vLLM |
Community resources
- TGI GitHub — source, issues, releases
- TGI documentation — official reference
- HF Discord — community help
Related
- Overview — when TGI fits your workload
- Kubernetes — self-hosted production deploys
- Observability — metrics before you open a support ticket