A unified, delightful Python concurrency library that makes parallel and distributed computing feel like writing sequential code. Built on the actor model, concurry provides workers, pools, rate limiting, retries, and seamless integration with Ray for distributed execution.
Calling LLMs in a loop is painfully slow. With concurry, get 50x faster batch processing with just 3 lines of code change:
from pydantic import BaseModel
from concurry import Worker
from tqdm import tqdm
import litellm
# Define your LLM worker
class LLM(Worker, BaseModel):
temperature: float
top_p: float
model: str
def call_llm(self, prompt: str) -> str:
response = litellm.completion(
model=self.model,
messages=[{"role": "user", "content": prompt}],
temperature=self.temperature,
top_p=self.top_p,
)
return response
# Load prompts (e.g., evaluating AI-generated responses for harmfulness)
prompts = [...] # 1000 prompts
# ❌ Sequential: ~775 seconds
llm = LLM(temperature=0.1, top_p=0.9, model="meta-llama/llama-3.1-8b-instruct")
responses = [llm.call_llm(prompt) for prompt in tqdm(prompts)]
# ✅ Concurrent with concurry: ~16 seconds (50x faster!)
llm = LLM.options(
mode='thread',
max_workers=100
).init(temperature=0.1, top_p=0.9, model="meta-llama/llama-3.1-8b-instruct")
futures = [llm.call_llm(prompt) for prompt in tqdm(prompts, desc="Submitting")]
responses = [f.result() for f in tqdm(futures, desc="Collecting results")]What changed? Just added .options(mode='thread', max_workers=100).init(...) and called .result() on futures. That's it.
Python's concurrency landscape is fragmented. Threading, multiprocessing, asyncio, and Ray all have different APIs, behaviors, and gotchas. Concurry unifies them with a consistent, elegant interface that works the same way everywhere.
# Different APIs for different backends
from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor
import asyncio
import ray
# Thread pool - one API
with ThreadPoolExecutor() as executor:
future = executor.submit(task, arg)
result = future.result()
# Process pool - same API, different behavior
with ProcessPoolExecutor() as executor:
future = executor.submit(task, arg)
result = future.result()
# Asyncio - completely different API
async def main():
result = await asyncio.create_task(async_task(arg))
# Ray - yet another API
@ray.remote
def ray_task(arg):
return result
future = ray_task.remote(arg)
result = ray.get(future)from concurry import Worker
class DataProcessor(Worker):
def __init__(self, multiplier: int):
self.multiplier = multiplier
def process(self, value: int) -> int:
return value * self.multiplier
# Same code, different backends - just change one parameter!
worker = DataProcessor.options(mode="thread").init(10) # Thread
# worker = DataProcessor.options(mode="process").init(10) # Process
# worker = DataProcessor.options(mode="asyncio").init(10) # Asyncio
# worker = DataProcessor.options(mode="ray").init(10) # Ray (distributed!)
result = worker.process(42).result() # 420
worker.stop()One interface. Five execution modes. Zero headaches.
Stateful workers that run across sync, thread, process, asyncio, and Ray backends with a unified API.
class Counter(Worker):
def __init__(self):
self.count = 0
def increment(self) -> int:
self.count += 1
return self.count
# State is isolated per worker
counter = Counter.options(mode="thread").init()
print(counter.increment().result()) # 1
print(counter.increment().result()) # 2Distribute work across multiple workers with pluggable strategies (round-robin, least-active, random).
# Pool of 10 workers with round-robin load balancing
pool = DataProcessor.options(
mode="thread",
max_workers=10,
load_balancing="round_robin"
).init()
# Work automatically distributed across all workers
futures = [pool.process(i) for i in range(1000)]
results = [f.result() for f in futures]Token bucket, leaky bucket, sliding window algorithms. Enforce rate limits across workers with atomic multi-resource acquisition.
from concurry import RateLimit, CallLimit
# Limit to 100 API calls and 10k tokens per minute
pool = APIWorker.options(
mode="thread",
max_workers=20,
limits=[
CallLimit(window_seconds=60, capacity=100),
RateLimit(key="tokens", window_seconds=60, capacity=10_000)
]
).init()
# Limits automatically enforced across all 20 workersExponential backoff, exception filtering, output validation, and automatic resource release between retries.
# Retry on transient errors with exponential backoff
worker = APIWorker.options(
mode="thread",
num_retries=5,
retry_algorithm="exponential",
retry_on=[ConnectionError, TimeoutError],
retry_until=lambda result: result.get("status") == "ok"
).init()
# Automatically retries up to 5 times on failurePass futures between workers seamlessly. Concurry automatically unwraps them - even with zero-copy optimization for Ray.
# Producer creates futures
producer = DataSource.options(mode="thread").init()
data_future = producer.get_data()
# Consumer automatically unwraps the future
consumer = DataProcessor.options(mode="process").init()
result = consumer.process(data_future).result() # Auto-unwrapped!Beautiful progress bars with state indicators, automatic style detection, and rich customization.
from concurry.utils.progress import ProgressBar
for item in ProgressBar(items, desc="Processing"):
process(item)
# Shows: Processing: 100%|██████████| 1000/1000 [00:05<00:00] ✓ CompleteFull validation support with both model inheritance and decorators (Ray-compatible @validate decorator included).
from morphic import validate
class ValidatedWorker(Worker):
@validate
def __init__(self, multiplier: int):
self.multiplier = multiplier
@validate
def process(self, x: int) -> int:
return x * self.multiplier
# Automatic type coercion and validation
worker = ValidatedWorker.options(mode="ray").init(multiplier="5") # str→intAsyncIO workers route async methods to an event loop and sync methods to a dedicated thread for optimal performance (10-50x speedup for I/O).
class AsyncAPIWorker(Worker):
async def fetch(self, url: str) -> dict:
async with aiohttp.ClientSession() as session:
async with session.get(url) as resp:
return await resp.json()
worker = AsyncAPIWorker.options(mode="asyncio").init()
# 100 concurrent requests instead of sequential!
result = worker.fetch_many(urls).result()# Basic installation
pip install concurry
# With Ray support for distributed computing
pip install concurry[ray]
# Development installation with all extras
pip install concurry[all]Requirements: Python 3.10+
from concurry import Worker
class DataProcessor(Worker):
def process(self, x: int) -> int:
return x ** 2
# Context manager automatically cleans up all workers
with DataProcessor.options(mode="thread", max_workers=5).init() as pool:
futures = [pool.process(i) for i in range(100)]
results = [f.result() for f in futures]
# All workers automatically stopped herefrom concurry import TaskWorker
worker = TaskWorker.options(mode="process").init()
# Submit any function
future = worker.submit(lambda x: x ** 2, 42)
print(future.result()) # 1764
# Use map() for batch processing
results = list(worker.map(lambda x: x * 2, range(10)))
print(results) # [0, 2, 4, 6, 8, 10, 12, 14, 16, 18]
worker.stop()import ray
from concurry import Worker
ray.init()
class DistributedProcessor(Worker):
def __init__(self, model_name: str):
self.model = load_large_model(model_name)
def predict(self, data: list) -> list:
return self.model.predict(data)
# 50 Ray actors across your cluster
pool = DistributedProcessor.options(
mode="ray",
max_workers=50,
num_cpus=2,
num_gpus=0.5
).init(model_name="bert-large")
# Distribute work across entire cluster
batches = [data[i:i+32] for i in range(0, len(data), 32)]
futures = [pool.predict(batch) for batch in batches]
results = [f.result() for f in futures]
pool.stop()
ray.shutdown()from concurry import Worker, RateLimit, CallLimit
from morphic import validate
import openai
class LLMWorker(Worker):
@validate
def __init__(self, model: str = "gpt-4", temperature: float = 0.7):
self.model = model
self.temperature = temperature
self.client = openai.OpenAI()
@validate
def generate(self, prompt: str, max_tokens: int = 500) -> dict:
# Rate limits automatically enforced
with self.limits.acquire(requested={"tokens": max_tokens}) as acq:
response = self.client.chat.completions.create(
model=self.model,
messages=[{"role": "user", "content": prompt}],
max_tokens=max_tokens,
temperature=self.temperature
)
result = {
"text": response.choices[0].message.content,
"tokens": response.usage.total_tokens,
}
# Report actual usage for accurate rate limiting
acq.update(usage={"tokens": result["tokens"]})
return result
# Pool of 10 workers with shared rate limits and automatic retries
pool = LLMWorker.options(
mode="thread",
max_workers=10,
# Shared rate limits across all workers
limits=[
RateLimit(key="tokens", window_seconds=60, capacity=10_000),
CallLimit(window_seconds=60, capacity=100)
],
# Automatic retry with exponential backoff
num_retries=3,
retry_algorithm="exponential",
retry_on=[openai.RateLimitError, openai.APIConnectionError],
retry_until=lambda r: len(r.get("text", "")) > 50
).init(model="gpt-4")
# Process 100 prompts with automatic rate limiting and retries
prompts = [f"Summarize topic {i}" for i in range(100)]
futures = [pool.generate(prompt, max_tokens=200) for prompt in prompts]
results = [f.result() for f in futures]
print(f"Processed {len(results)} prompts")
print(f"Total tokens: {sum(r['tokens'] for r in results)}")
pool.stop()from concurry import Worker
import aiohttp
import asyncio
class AsyncAPIWorker(Worker):
def __init__(self, base_url: str):
self.base_url = base_url
async def fetch(self, endpoint: str) -> dict:
"""Async method - runs in event loop."""
async with aiohttp.ClientSession() as session:
async with session.get(f"{self.base_url}/{endpoint}") as resp:
return await resp.json()
async def fetch_many(self, endpoints: list) -> list:
"""Fetch multiple URLs concurrently."""
tasks = [self.fetch(ep) for ep in endpoints]
return await asyncio.gather(*tasks)
worker = AsyncAPIWorker.options(mode="asyncio").init("https://api.example.com")
# All 100 requests execute concurrently (10-50x faster)!
result = worker.fetch_many([f"data/{i}" for i in range(100)]).result()
worker.stop()- User Guide - Comprehensive tutorials and examples
- API Reference - Detailed API documentation
- Examples - Real-world usage patterns
- Contributing - How to contribute
- Unified API: One interface for all concurrency paradigms
- Actor Model: Stateful workers with isolated state
- Production-Ready: Rate limiting, retries, validation, monitoring
- Performance: Zero-copy optimizations where possible
- Developer Experience: Intuitive API, rich documentation, great error messages
Contributions are welcome! Please see CONTRIBUTING.md for guidelines.
This project is licensed under the Apache 2.0 License - see the LICENSE file for details.
- Built on top of morphic for validation
- Inspired by Ray, Pydantic, and the actor model
- Progress bars powered by tqdm
Made with ❤️ by the Amazon Science team