Status: draft cheapgpu.endpoint.v1-draft.4. Registration and inference services
are not available. The workflow below defines the intended contract.
You provide an HTTPS inference endpoint that implements a declared API profile. CheapGPU sends requests to that endpoint. You can reuse an existing server if it meets the profile; otherwise, adjust your server or provide your own compatibility layer. CheapGPU does not build supplier-specific adapters or fund accounts in your upstream billing system.
All suppliers follow the same rules. Support for machines without a public endpoint is deferred; no supplier agent is part of this phase. Shared rules apply per profile. Suppliers choose the profiles they implement; the model catalog does not require every task to use Chat Completions. Native video creation/status/result operations need a separate profile and remain undefined in the current configuration schema.
Registration workflow
Section titled “Registration workflow”- Register the full inference URL, a credential binding, and model mappings. Use the configuration example. The control-plane HTTP routes and secure credential provisioning flow are not defined yet; this example is a stored configuration, not a POST request.
- Declare one profile and its capabilities for each model mapping. Map a
platform model and immutable revision to
supplier_model_id. Your server must accept that ID in requests and return the same ID in responses. - Start a conformance run after accepting its resource limits. Inspect each result and fix failures. No manual approval queue is used.
- Once implemented, backend makes only a current, passing scope eligible for matching. An active offer and the platform’s admission checks are also required. A conformance pass does not create an offer or authorize dispatch.
For both defined HTTP profiles, inference_url is the exact POST target. CheapGPU does not append /v1 or another
path. This draft permits public DNS hostnames over HTTPS on port 443, without
userinfo, a query, or a fragment. See Endpoint security.
Configuration ownership
Section titled “Configuration ownership”Backend assigns endpoint_id, supplier_id, and configuration_revision.
You supply the URL and model bindings through the future management interface.
credential_binding_id refers to a platform-managed credential record, not a
token value or an arbitrary secret-store path. Both defined profiles use bearer
authentication. The configuration cannot contain custom headers or credentials.
Each revision is immutable. A URL, credential rotation, model mapping, returned model ID, profile, or capability change creates a new revision and invalidates the previous revision’s eligibility for new traffic. An old check finishing later cannot activate the new revision. Price changes use separate offer versions and do not rewrite an admitted trade.
Requests and failures
Section titled “Requests and failures”Gateway obtains authoritative backend admission and a one-shot dispatch grant before sending inference. Backend freezes the selected offer, endpoint revision, model revision, profile, budget, and original financial deadlines.
This draft does not require a CheapGPU wrapper around the native JSON body or supplier-side processing of custom CheapGPU headers. Gateway records the request, attempt, dispatch claim, endpoint revision, and upstream response identity. This binding does not provide supplier-side durable deduplication or exactly-once execution.
Gateway must disable automatic inference retries and redirects. A lost response, timeout, process restart, or disconnected stream does not authorize another attempt. Closing an upstream connection requests cancellation but cannot prove the server stopped generating. Separate cancellation acknowledgments and a durable supplier deduplication/status extension remain open work.
Accounting boundary
Section titled “Accounting boundary”Usage fields are evidence, not permission to debit a buyer or pay a supplier. Backend determines charges and supplier shares. A share requires confirmed, collected buyer revenue for that transaction; CheapGPU does not guarantee income.
If execution or billable usage remains unknown, bounded recovery closes the transaction at its original deadline with zero buyer charge, zero supplier share, and zero commission, releasing the reservation. Late evidence cannot reopen that closure. Successful collection, accepted usage evidence, partial output, and buyer-cancellation policies still need production definitions.