edgequake-litellm 0.1.2

Drop-in LiteLLM replacement backed by Rust — same API, 10× lower latency

edgequake-litellm

Drop-in LiteLLM replacement backed by Rust — same API, lower overhead.

edgequake-litellm wraps the edgequake-llm Rust core via PyO3, providing a high-performance drop-in for LiteLLM. Swap the import — the rest of your code stays unchanged.

# Before
import litellm

# After — same API, Rust-backed
import edgequake_litellm as litellm

Features

• LiteLLM-compatible API — completion(), acompletion(), stream(), embedding(), same call signatures, same response shape (resp.choices[0].message.content).
• Multi-provider routing — OpenAI, Anthropic, Gemini, Mistral, OpenRouter, xAI, Ollama, LM Studio, HuggingFace, and more, via provider/model strings.
• Async-native — built on Tokio; sync and async Python both supported.
• Single wheel per platform — uses PyO3's abi3-py39 stable ABI, one .whl covers Python 3.9–3.13+.
• Zero Python runtime dependencies — the Rust extension is self-contained.
• Full type annotations — ships with py.typed and .pyi stubs.
• max_completion_tokens support — works for all OpenAI model families including o1, o3-mini, o4-mini, gpt-4.1, gpt-4.1-nano that require this field.
• Cache hit tokens — resp.cache_hit_tokens exposes OpenAI prompt cache hits and Anthropic cache reads.
• Reasoning tokens — resp.thinking_tokens surfaces o-series reasoning and Claude extended thinking token counts.

What's New in 0.1.1

• max_completion_tokens fixed for OpenAI o-series and gpt-4.1 model families (previously returned 400 Bad Request).
• resp.cache_hit_tokens — new property returning tokens served from provider cache (None if not applicable).
• resp.thinking_tokens — new property returning reasoning/thinking token count for o-series and Claude models.
• Both new properties are included in resp.to_dict().

See CHANGELOG.md for the full history.

Installation

pip install edgequake-litellm

Quick Start

import edgequake_litellm as litellm   # drop-in import alias

# ── Synchronous chat ────────────────────────────────────────────────────────
resp = litellm.completion(
    "openai/gpt-4o-mini",
    [{"role": "user", "content": "Hello, world!"}],
)
# litellm-compatible access
print(resp.choices[0].message.content)
# convenience shortcut
print(resp.content)

# ── Asynchronous chat ───────────────────────────────────────────────────────
import asyncio

async def main():
    resp = await litellm.acompletion(
        "anthropic/claude-3-5-haiku-20241022",
        [{"role": "user", "content": "Tell me a joke."}],
        max_tokens=128,
        temperature=0.8,
    )
    print(resp.choices[0].message.content)

asyncio.run(main())

# ── Streaming (async generator) ─────────────────────────────────────────────
async def stream_example():
    messages = [{"role": "user", "content": "Count to five."}]
    async for chunk in litellm.acompletion("openai/gpt-4o", messages, stream=True):
        print(chunk.choices[0].delta.content or "", end="", flush=True)

# ── Embeddings ──────────────────────────────────────────────────────────────
result = litellm.embedding(
    "openai/text-embedding-3-small",
    ["Hello world", "Rust is fast"],
)
# litellm-compatible access
print(result.data[0].embedding[:3])
# legacy list access still works
print(len(result), len(result[0]))  # 2 1536

Provider Routing

Pass provider/model as the first argument — the prefix selects the provider:

Provider	Example model string
OpenAI	openai/gpt-4o
Anthropic	anthropic/claude-3-5-sonnet-20241022
Google Gemini	gemini/gemini-2.0-flash
Mistral	mistral/mistral-large-latest
OpenRouter	openrouter/meta-llama/llama-3.1-70b-instruct
xAI	xai/grok-3-beta
Ollama	ollama/llama3.2
LM Studio	lmstudio/local-model
HuggingFace	huggingface/mistralai/Mixtral-8x7B-Instruct-v0.1
Mock (tests)	mock/any-name

API Reference

completion(model, messages, **kwargs) → ModelResponseCompat

Synchronous chat completion. Blocks but releases the GIL during Rust I/O so other Python threads keep running.

resp = litellm.completion(
    "openai/gpt-4o",
    messages,
    max_tokens=256,
    temperature=0.7,
    system="You are a helpful assistant.",
    max_completion_tokens=256,  # alias for max_tokens; required for o1/o3/gpt-4.1 models
    seed=42,
    response_format={"type": "json_object"},  # or "text" / "json_object"
)

# All of these access the same content:
resp.choices[0].message.content   # litellm path
resp.content                       # shortcut
resp["choices"][0]["message"]["content"]  # dict-style

resp.usage.total_tokens
resp.model
resp.response_ms                  # latency in milliseconds
resp.to_dict()                    # plain dict

# New in 0.1.1 — cache and reasoning token metadata
resp.cache_hit_tokens             # int | None — tokens served from provider cache
resp.thinking_tokens              # int | None — reasoning tokens (o-series, Claude)
resp.thinking_content             # str | None — visible thinking text (Claude)

# The same data via usage object:
resp.usage.cache_read_input_tokens  # same as resp.cache_hit_tokens
resp.usage.reasoning_tokens         # same as resp.thinking_tokens

acompletion(model, messages, stream=False, **kwargs)

Async chat completion. Returns ModelResponseCompat or (if stream=True) AsyncGenerator[StreamChunkCompat, None].

# Non-streaming
resp = await litellm.acompletion("openai/gpt-4o", messages)

# Streaming
async for chunk in await litellm.acompletion("openai/gpt-4o", messages, stream=True):
    print(chunk.choices[0].delta.content or "", end="")

stream(model, messages, **kwargs) → AsyncGenerator[StreamChunk, None]

Low-level streaming. Raw StreamChunk objects:

async for chunk in litellm.stream("openai/gpt-4o", messages):
    if chunk.content:
        print(chunk.content, end="")
    elif chunk.is_finished:
        print(f"\n[stop: {chunk.finish_reason}]")

embedding(model, input, **kwargs) → EmbeddingResponseCompat

Synchronous embeddings. Returns an EmbeddingResponseCompat that supports both litellm-style and legacy list-style access:

result = litellm.embedding("openai/text-embedding-3-small", ["foo", "bar"])

# litellm path
result.data[0].embedding

# backwards-compatible list access
for vec in result:          # iterates List[float]
    print(len(vec))
result[0]                   # List[float]
len(result)                 # number of vectors

aembedding(model, input, **kwargs) → EmbeddingResponseCompat

Async embeddings — same return type as embedding().

stream_chunk_builder(chunks, messages=None) → ModelResponseCompat

Reconstruct a full ModelResponseCompat from a collected list of streaming chunks:

from edgequake_litellm import stream_chunk_builder

chunks = []
async for chunk in litellm.stream("openai/gpt-4o", messages):
    chunks.append(chunk)

full = stream_chunk_builder(chunks, messages=messages)
print(full.content)

Configuration

Module-level globals mirror litellm:

import edgequake_litellm as litellm

litellm.set_verbose = True      # enable debug logging
litellm.drop_params = True      # drop unknown params (always True)

# Set default provider / model
litellm.set_default_provider("anthropic")
litellm.set_default_model("claude-3-5-haiku-20241022")

# Now the provider prefix can be omitted:
resp = litellm.completion("claude-3-5-haiku-20241022", messages)

Exception Hierarchy

Exceptions mirror LiteLLM for painless migration:

import edgequake_litellm as litellm

try:
    resp = litellm.completion("openai/gpt-4o", messages)
except litellm.AuthenticationError as e:
    print(f"Check your API key: {e}")
except litellm.RateLimitError:
    time.sleep(5)
except litellm.ContextWindowExceededError:
    # trim messages and retry
    pass
except litellm.NotFoundError:      # alias for ModelNotFoundError
    pass
except litellm.APIConnectionError:
    pass

All exceptions (AuthenticationError, RateLimitError, ContextWindowExceededError, ModelNotFoundError, Timeout, APIConnectionError, APIError) are also available from edgequake_litellm.exceptions.

Environment Variables

Provider credentials follow the standard naming convention:

Provider	Environment variable
OpenAI	OPENAI_API_KEY
Anthropic	ANTHROPIC_API_KEY
Gemini	GEMINI_API_KEY
Mistral	MISTRAL_API_KEY
OpenRouter	OPENROUTER_API_KEY
xAI	XAI_API_KEY
HuggingFace	HF_TOKEN
Ollama	OLLAMA_HOST (default: http://localhost:11434)
LM Studio	LMSTUDIO_HOST (default: http://localhost:1234)

Defaults can also be set via LITELLM_EDGE_PROVIDER / LITELLM_EDGE_MODEL.

Development

Prerequisites

• Rust ≥ 1.83 (rustup toolchain install stable)
• Python ≥ 3.9
• pip install maturin

Build from source

git clone https://github.com/raphaelmansuy/edgequake-llm.git
cd edgequake-llm/edgequake-litellm

# Create a virtual environment
python -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate

pip install maturin pytest pytest-asyncio ruff mypy

# Build & install in dev mode (incremental Rust + Python)
maturin develop --release

# Run unit tests (mock provider — no API keys needed)
pytest tests/ -k "not e2e" -v

Running E2E tests

export OPENAI_API_KEY=sk-...
pytest tests/test_e2e_openai.py -v

Publishing

# Bump version in pyproject.toml AND Cargo.toml (must match), then:
git tag py-v0.2.0
git push --tags
# GitHub Actions builds and publishes to PyPI automatically.

License

Apache-2.0 — see LICENSE-APACHE.