openai-http 0.4.2

OpenAI v1 API-compatible HTTP service backed by HuggingFace Transformers

OpenAI HTTP API Service

OpenAI v1 API compatible HTTP service with pluggable inference backends. Built with FastAPI and supports streaming responses, request queueing, and OpenTelemetry metrics.

Features

• ✅ OpenAI v1 API compatibility (Models, Chat Completions, Completions, Embeddings)
• ✅ Server-Sent Events (SSE) streaming support
• ✅ Request queue with GPU serialization (prevents concurrent GPU access)
• ✅ TOML + environment variable configuration
• ✅ Structured JSON logging with request IDs
• ✅ OpenTelemetry metrics (Prometheus endpoint)
• ✅ Mock backend for testing (1536-dim embeddings, streaming chat, tool calls)
• ✅ Transformers backend example (examples/transformers-backend/ — uses Qwen3.5-0.8B)
• ✅ Custom backend SDK (BackendBase ABC, run_server() entry point)
• 🚧 Audio (speech, transcriptions, translations) endpoints
• 🚧 Image generation / editing endpoints
• ✅ Tool calling (function definitions in chat) — supported in transformers backend

Quick Start

Prerequisites

• Python 3.12+ (see .python-version)
• uv package manager

Installation

# Clone and enter directory
git clone <repo>
cd openai-http

# Create venv and install dependencies
uv sync --all-extras

Run the Server

# Start with mock backend (no GPU required)
uv run -m openai_http

# Server runs on http://0.0.0.0:8000

Test with curl

# List models
curl http://localhost:8000/v1/models

# Chat completion (non-streaming)
curl -X POST http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mock-gpt",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

# Chat completion (streaming)
curl -X POST http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mock-gpt",
    "messages": [{"role": "user", "content": "Hello"}],
    "stream": true
  }'

# Health check
curl http://localhost:8000/health

Run with a real Transformers backend

# Install heavy deps (not included in default install)
uv pip install torch transformers accelerate

# Start with Qwen3.5-0.8B (downloads ~1.6 GB on first run)
uv run python examples/transformers-backend/transformers_backend.py

# Customize model, temperature, or enable thinking mode
uv run python examples/transformers-backend/transformers_backend.py \
  --model Qwen/Qwen2.5-0.5B-Instruct \
  --temperature 0.7 \
  --thinking \
  --port 8000

See examples/transformers-backend/README.md for details.

Configuration

Configuration uses config.toml with environment variable overrides.

config.toml Example

[server]
host = "0.0.0.0"
port = 8000

[auth]
enabled = false
api_keys = []  # ["sk-..."] when enabled

[queue]
depth = 32  # max queued requests

[observability]
log_level = "info"
log_format = "json"  # "json" or "text"
metrics_enabled = true
metrics_port = 9464

Environment Variables

Override any config with OPENAI_HTTP__ prefix (double underscore for nesting):

OPENAI_HTTP__SERVER__PORT=9000
OPENAI_HTTP__AUTH__ENABLED=true
OPENAI_HTTP__AUTH__API_KEYS=["sk-test-key"]

API Endpoints

Implemented

• GET /v1/models — List available models
• GET /v1/models/{model_id} — Get model info
• POST /v1/chat/completions — Chat completion (streaming + non-streaming)
• POST /v1/completions — Text completion
• POST /v1/embeddings — Text embeddings
• GET /health — Health check

Stub endpoints (registered but return 501)

• POST /v1/audio/speech, /v1/audio/transcriptions, /v1/audio/translations
• POST /v1/images/generations, /v1/images/edits, /v1/images/variations

Planned

• Audio / image generation (currently return 501)
• POST /v1/moderations
• POST /v1/files, file management
• POST /v1/fine_tuning/jobs
• Batch API

Development

Run Tests

# All tests
uv run pytest tests/ -v

# SDK tests only (auto-starts server)
uv run pytest tests/sdk/ -v

# Unit tests only
uv run pytest tests/unit/ -v

# Single test
uv run pytest tests/sdk/test_models.py::TestModelsAPI::test_models_list -v

Note: SDK tests (tests/sdk/) automatically start a mock server on port 8000 in a background thread. Don't start the server manually before running tests.

Linting & Type Checking

uv run ruff check openai_http/ tests/
uv run mypy openai_http/

Project Structure

openai_http/
├── __init__.py         # Public API: BackendBase, run_server, setup_logging
├── _server.py          # run_server() entry point (library mode)
├── app.py              # FastAPI factory, lifespan, middleware
├── config.py           # pydantic-settings configuration
├── errors.py           # OpenAI-format error handlers
├── queue.py            # Request queue (asyncio.Semaphore)
├── routers/            # API endpoints
│   ├── chat.py        # /v1/chat/completions
│   ├── completions.py # /v1/completions
│   ├── embeddings.py  # /v1/embeddings
│   ├── models.py      # /v1/models
│   ├── audio.py       # /v1/audio/* (stub)
│   ├── images.py      # /v1/images/* (stub)
│   └── health.py      # /health
├── schemas/            # Pydantic v2 models
├── backends/           # Inference backends
│   ├── base.py        # BackendBase ABC
│   └── mock_backend.py # Mock implementation
└── observability/      # Logging + metrics

examples/
└── transformers-backend/  # Real Transformers backend using Qwen2.5-0.5B

tests/
├── conftest.py         # Async httpx fixtures
├── unit/              # Unit tests
└── sdk/               # OpenAI SDK compatibility tests
    └── conftest.py    # Sync OpenAI client fixtures

config.toml            # Configuration
specs/                 # Feature specs & plans

Architecture

Request Flow

1. Request arrives at FastAPI router
2. RequestIDMiddleware assigns unique request ID
3. RequestLoggingMiddleware logs request metadata
4. Router calls queue.acquire() (waits if GPU busy)
5. Backend executes inference (mock or transformers)
6. Response formatted to OpenAI v1 schema
7. Error handlers catch exceptions and format {"error": {...}}

Custom backends

Create your own backend by subclassing BackendBase and plugging it into run_server():

import openai_http

class MyBackend(openai_http.BackendBase):
    async def generate(self, prompt, **kwargs):
        return {"generated_text": "...", "usage": {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}}

    async def generate_stream(self, prompt, **kwargs):
        yield "Hello"
        yield " world"

    async def list_models(self):
        return [{"id": "my-model", "object": "model", "created": 0, "owned_by": "me"}]

    async def get_model(self, model_id):
        return {"id": model_id, "object": "model", "created": 0, "owned_by": "me"} if model_id == "my-model" else None

openai_http.run_server(MyBackend(), port=8000)

See examples/transformers-backend/ for a full working implementation.

Built-in Backend Protocol

All backends subclass BackendBase (defined in backends/base.py):

Method	Required	Purpose
generate()	✅	Non-streaming completion
generate_stream()	✅	Streaming completion (yields str chunks)
list_models()	✅	List available models
get_model()	✅	Get a model by ID
embed()	❌	Embeddings (default → HTTP 501)
generate_tool_calls()	❌	Tool/function calling (default → HTTP 501)
setup() / teardown()	❌	Lifecycle hooks (default no-op)

Request Queue

GPU inference is serialized using asyncio.Semaphore(1) to prevent out-of-memory errors. When the queue reaches queue.depth, new requests receive HTTP 429 (Too Many Requests).

Testing Strategy

• Unit tests (tests/unit/): Test individual functions/classes in isolation
• Integration tests (tests/integration/): Test endpoint behavior with httpx
• Contract tests (tests/contract/): Verify OpenAI API contract compliance
• SDK tests (tests/sdk/): Use official openai Python SDK against running server

Unimplemented endpoints use pytest.skip() via _call_or_skip() helper and auto-activate when routes are added.

License

MIT

Contributing

See specs/001-openai-http-api/plan.md for implementation roadmap and AGENTS.md for development guidelines.