> ## Documentation Index > Fetch the complete documentation index at: https://docs.aireserve.com/llms.txt > Use this file to discover all available pages before exploring further. # Chat completion on a dedicated GPU deployment > OpenAI-compatible chat completions on your organization's own dedicated HuggingFace vLLM deployment, with your ordinary AI Reserve API key — the HuggingFace credential stays server-side. `deploymentId` is the deployment's id from **Admin → GPUs → How to call endpoint**; a key from another organization gets the same `404` as an unknown deployment. `model` is optional and, when sent, must name the repository actually deployed. `stream: true` answers server-sent events with usage on the final chunk. Inference is not billed per token — the deployment's GPU time is billed from the organization wallet — so wallet-balance and budget-cap gates do not apply here, while key status, IP allowlists and rate limits do. A scale-to-zero deployment that is asleep is woken by the request (the first call can take minutes); a paused, failed or provisioning deployment answers `409`, and a deployment of an adult-only model answers `403` once the organization's adult-content access is turned off. ## OpenAPI ````yaml /api-reference/openapi.json post /v1/gpu/{deploymentId}/chat/completions openapi: 3.0.3 info: title: AI Reserve Gateway API description: >- The AI Reserve API surface — inference, durable environments, workspaces, sandboxes, executions, agent runs, GPU infrastructure, and fine-tuning — all behind one endpoint and one API key. version: 1.0.0 servers: - url: https://api.aireserve.com security: - bearerAuth: [] tags: - name: Customers description: >- Per-end-customer usage, cost, and billable amounts for organizations that tag requests with their own customers. - name: Dedicated GPU endpoints description: >- OpenAI-compatible inference on your organization's own dedicated HuggingFace vLLM deployments, called with an ordinary AI Reserve API key. GPU time is billed from the wallet; requests add no per-token charge. - name: GPU Infrastructure description: >- Provision and manage provider-hidden GPU instances. The recommended Instances API accepts workload requirements and an hourly price ceiling; AI Reserve handles provider placement internally. Every mutation is asynchronous and returns an operation to poll. - name: Environments description: Immutable, versioned runtime definitions. - name: Workspaces description: Durable project storage pinned to one environment version. - name: Sandboxes description: >- Ephemeral provider-disclosed compute allocated from sealed quotes. Requests remain provider-neutral. - name: Executions description: Durable asynchronous operations and command, service, or agent executions. - name: Agent Runs description: Durable Agent SDK runs with replayable events and external artifacts. - name: Customer keys description: >- Mint, list, and revoke API keys pinned to your own end customers (and their organizations), so spend splits per customer without per-request tagging. - name: AWS CLI login description: >- Browser-authorization endpoints consumed by `aws login --profile …` (AWS CLI v2.32.0+). All three routes are unauthenticated — the caller has no credential until `/v1/token` succeeds. - name: Fine-tuning description: >- Fine-tune open models on your own data. Data lives in an S3 bucket inside your organization's AWS hub account; you name a base, LoRA or full, a price ceiling and how the result must be servable, and AI Reserve places the run on a backend internally. The result is a model in your bucket you can deploy or download. - name: 'Admin: Catalog' description: Discover the operations available to this token. - name: 'Admin: Analytics' description: Usage, spend and attribution reads. - name: 'Admin: API Keys' description: Organization and team inference keys. - name: 'Admin: Billing' description: Wallet, ledger and invoice reads. - name: 'Admin: Compute' description: GPU machines and persistent storage (quote, launch, inspect, stop). - name: 'Admin: Compute Runtime' description: >- Portable environments, workspaces, sandboxes, executions, operations and durable agent runs. - name: 'Admin: Containers' description: The organization's AWS account, ECS services, logs and secrets. - name: 'Admin: Files' description: Organization shared files. - name: 'Admin: Members' description: 'People in the organization: invite, profile, team, access, role.' - name: 'Admin: Models & Routing' description: Model availability, defaults, fallbacks and routing. - name: 'Admin: Organization' description: Retention, logging, branding, sign-in domains, classification. - name: 'Admin: Personal' description: The caller's own usage, budget, API keys and preferences. - name: 'Admin: Teams & Budgets' description: Teams, spend caps and allocations. - name: 'Admin: Training' description: Fine-tuning datasets, runs, models and deployments. paths: /v1/gpu/{deploymentId}/chat/completions: post: tags: - Dedicated GPU endpoints summary: Chat completion on a dedicated GPU deployment description: >- OpenAI-compatible chat completions on your organization's own dedicated HuggingFace vLLM deployment, with your ordinary AI Reserve API key — the HuggingFace credential stays server-side. `deploymentId` is the deployment's id from **Admin → GPUs → How to call endpoint**; a key from another organization gets the same `404` as an unknown deployment. `model` is optional and, when sent, must name the repository actually deployed. `stream: true` answers server-sent events with usage on the final chunk. Inference is not billed per token — the deployment's GPU time is billed from the organization wallet — so wallet-balance and budget-cap gates do not apply here, while key status, IP allowlists and rate limits do. A scale-to-zero deployment that is asleep is woken by the request (the first call can take minutes); a paused, failed or provisioning deployment answers `409`, and a deployment of an adult-only model answers `403` once the organization's adult-content access is turned off. parameters: - name: deploymentId in: path required: true schema: format: uuid type: string requestBody: required: true content: application/json: schema: type: object additionalProperties: true example: model: Qwen/Qwen2.5-7B-Instruct messages: - role: user content: Hello! responses: '200': description: Default Response components: securitySchemes: bearerAuth: type: http scheme: bearer description: >- Your AI Reserve API key (aireserve_api_…) from the API Keys page in the portal (https://portal.aireserve.com/keys). ```` This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.