mlrunx 0.1.3

MLRunX Python SDK - async, non-blocking ML experiment tracking

MLRunX Python SDK

Async, non-blocking ML experiment tracking SDK for Python.

Installation

pip install mlrunx

Quick Start

import mlrunx

# Initialize a run (works offline if server unavailable)
run = mlrunx.init(
    project="my-project",
    name="training-run-1",
    tags={"model": "resnet50", "dataset": "imagenet"},
)

# Log hyperparameters
run.log_params({
    "learning_rate": 0.001,
    "batch_size": 32,
    "optimizer": "adam",
})

# Training loop - logging is non-blocking!
for step in range(1000):
    loss, accuracy = train_step()
    run.log({"loss": loss, "accuracy": accuracy}, step=step)

# Finish (flushes all pending data)
run.finish()

Using Context Manager

import mlrunx

with mlrunx.init(project="my-project") as run:
    run.log_params({"lr": 0.001})

    for step in range(1000):
        run.log({"loss": loss}, step=step)

    # Automatically finished on exit

Features

Non-Blocking Logging

All log() calls are non-blocking. Events are queued in memory and flushed in the background, ensuring your training loop runs at full speed.

# This won't slow down your training!
for step in range(100000):
    loss = train_step()  # Your expensive computation
    run.log({"loss": loss}, step=step)  # < 1μs overhead

Adaptive Batching

Events are batched intelligently before being sent:

• Size trigger: Flush when batch reaches max items (default: 1000)
• Bytes trigger: Flush when batch reaches max bytes (default: 1MB)
• Time trigger: Flush after max age (default: 1 second)

Configure via environment variables:

export MLRUNX_BATCH_SIZE=500
export MLRUNX_BATCH_MAX_BYTES=500000
export MLRUNX_BATCH_TIMEOUT_MS=2000

Metric Coalescing

When logging the same metric multiple times at the same step, only the latest value is sent (configurable):

# Only the last value (0.3) is sent for step 0
run.log({"loss": 0.5}, step=0)
run.log({"loss": 0.4}, step=0)
run.log({"loss": 0.3}, step=0)

Disable with:

export MLRUNX_COALESCE_METRICS=false

Offline Mode & Disk Spool

If the server is unavailable, events are automatically spooled to disk and synced when the connection is restored:

# Works even if server is down!
run = mlrunx.init(project="my-project")
print(run.is_offline)  # True if server unavailable

# Events are saved to ~/.mlrunx/spool/
for step in range(1000):
    run.log({"loss": loss}, step=step)

# When server comes back online, data syncs automatically

Configure spool settings:

export MLRUNX_SPOOL_ENABLED=true
export MLRUNX_SPOOL_DIR=~/.mlrunx/spool
export MLRUNX_SPOOL_MAX_SIZE=100000000  # 100MB

Compression

Large batches are automatically compressed with gzip:

export MLRUNX_COMPRESSION=true
export MLRUNX_COMPRESSION_LEVEL=6  # 1-9
export MLRUNX_COMPRESSION_MIN_BYTES=1000  # Only compress if > 1KB

API Reference

mlrunx.init()

Initialize a new run.

run = mlrunx.init(
    project="my-project",       # Required: project name
    name="experiment-1",        # Optional: run name (auto-generated if not provided)
    tags={"key": "value"},      # Optional: initial tags
    config={"lr": 0.001},       # Optional: initial config (logged as params)
)

run.log()

Log metrics (non-blocking).

run.log(
    {"loss": 0.5, "accuracy": 0.8},  # Metrics dict
    step=100,                         # Optional: step number
    timestamp=time.time(),            # Optional: custom timestamp
)

run.log_params()

Log hyperparameters (non-blocking).

run.log_params({
    "learning_rate": 0.001,
    "batch_size": 32,
    "model": "resnet50",
})

run.log_tags()

Log or update tags (non-blocking).

run.log_tags({
    "status": "running",
    "gpu": "A100",
})

run.finish()

Finish the run and flush all pending data.

run.finish(status="finished")  # or "failed", "killed"

Examples

Simple Training Loop

import mlrunx

run = mlrunx.init(project="demo")
run.log_params({"lr": 0.001, "epochs": 10})

for epoch in range(10):
    for batch in dataloader:
        loss = train_step(batch)
        run.log({"train/loss": loss})

    val_loss = validate()
    run.log({"val/loss": val_loss}, step=epoch)

run.finish()

PyTorch Integration

See examples/pytorch_mnist.py for a complete example.

import mlrunx
import torch

with mlrunx.init(project="mnist", tags={"framework": "pytorch"}) as run:
    run.log_params({"lr": 0.01, "epochs": 10})

    for epoch in range(10):
        for batch_idx, (data, target) in enumerate(train_loader):
            loss = train_step(data, target)
            run.log({"train/loss": loss.item()}, step=epoch * len(train_loader) + batch_idx)

        val_loss, val_acc = validate()
        run.log({"val/loss": val_loss, "val/accuracy": val_acc}, step=epoch)

HuggingFace Transformers

See examples/huggingface_text_classification.py for a complete example.

import mlrunx
from transformers import Trainer, TrainerCallback

class MLRunCallback(TrainerCallback):
    def __init__(self, run):
        self.run = run

    def on_log(self, args, state, control, logs=None, **kwargs):
        if logs:
            self.run.log(logs, step=state.global_step)

with mlrunx.init(project="nlp", tags={"framework": "transformers"}) as run:
    callback = MLRunCallback(run)
    trainer = Trainer(..., callbacks=[callback])
    trainer.train()

Configuration

All settings can be configured via environment variables:

Variable	Default	Description
MLRUNX_SERVER_URL	http://localhost:3001	Server URL
MLRUNX_API_KEY	None	API key for authentication
MLRUNX_BATCH_SIZE	1000	Max events per batch
MLRUNX_BATCH_MAX_BYTES	1000000	Max batch size in bytes
MLRUNX_BATCH_TIMEOUT_MS	1000	Max time before flush (ms)
MLRUNX_COALESCE_METRICS	true	Merge same metric at same step
MLRUNX_DEDUPE_PARAMS	true	Keep only last value for params
MLRUNX_COMPRESSION	true	Enable gzip compression
MLRUNX_SPOOL_ENABLED	true	Enable disk spooling
MLRUNX_SPOOL_DIR	~/.mlrunx/spool	Spool directory
MLRUNX_OFFLINE	false	Force offline mode
MLRUNX_DEBUG	false	Enable debug logging

Development

# Clone the repository
git clone https://github.com/your-org/mlrunx.git
cd mlrunx/sdks/python

# Create virtual environment
python -m venv .venv
source .venv/bin/activate

# Install in development mode
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Run examples
python examples/simple_training.py
python examples/pytorch_mnist.py

License

MIT License - see LICENSE for details.