Stateful workers
One ML model per worker, loaded once
`.class_task(MyClass, MyClass.method, args)` instantiates `MyClass` once per worker. The instance persists across items — perfect for ML models, DB connections, accumulators.
python
from olympipe import Pipeline
from transformers import pipeline as hf_pipeline
class EmbeddingWorker:
def __init__(self, model_name: str):
# Chargé une fois au démarrage du worker
self.model = hf_pipeline("feature-extraction", model=model_name)
def embed(self, text: str) -> list:
return self.model(text)[0][0] # CLS token
# 2 workers GPU, chacun charge le modèle une fois
embeddings = (
Pipeline(texts) # 50 000 textes
.batch(32)
.class_task(
EmbeddingWorker,
EmbeddingWorker.embed,
["sentence-transformers/all-MiniLM-L6-v2"],
count=2,
)
.wait_for_result()
)⚡ Performance
50 000 texts, HuggingFace embeddings
Sequential 2.0min
1 GPU worker 40.0s
2 GPU workers 21.0s
4 GPU workers 11.0s
🚀 10.9× faster
How it works
When the worker process starts, the constructor is called with `args`. All subsequent invocations reuse the same instance — model weights stay in RAM (or VRAM) without reloading.
Related examples
ML embedding generation
Generate embeddings for 50 000 texts with sentence-transformers.
MLNLP
LLM inference server
Serve a HuggingFace model as an HTTP API with Olympipe.
MLServerLLM
Parallel audio processing
Transcribe audio files with Whisper in parallel.
MLAudio