trainsight 0.2.0

AI Training Intelligence Dashboard - Live GPU telemetry, anomaly detection, and training diagnostics

TrainSight

AI Training Intelligence Dashboard for live GPU telemetry, anomaly detection, and training-oriented diagnostics.

Features

• Live GPU Monitoring: Real-time GPU utilization, memory, temperature, and power metrics
• Anomaly Detection: Automatic detection of thermal throttling, memory leaks, and training anomalies
• OOM Prediction: Predict Out-of-Memory errors before they crash your training
• Framework Integrations: Native support for PyTorch, Lightning, HuggingFace, DeepSpeed, Accelerate, Ray Train
• Experiment Tracking: Bridge to MLflow and Weights & Biases
• Production Ready: Prometheus exporter, Kubernetes GPU pod monitoring, cloud cost estimation
• Batch Size Optimization: Automatic batch size recommendation based on GPU memory

Installation

From PyPI (Recommended)

pip install trainsight

From GitHub

pip install git+https://github.com/modalgrasp/trainsight.git

Optional Dependencies

Install only what you need:

# PyTorch integration
pip install "trainsight[pytorch]"

# PyTorch Lightning integration
pip install "trainsight[lightning]"

# DeepSpeed integration
pip install "trainsight[deepspeed]"

# HuggingFace Accelerate integration
pip install "trainsight[accelerate]"

# Ray Train integration
pip install "trainsight[ray]"

# Experiment tracking
pip install "trainsight[mlflow]"
pip install "trainsight[wandb]"

# Kubernetes monitoring
pip install "trainsight[kubernetes]"

# Install everything
pip install "trainsight[all]"

Quick Start

CLI Dashboard

trainsight

Programmatic Usage

from trainsight import Dashboard
from trainsight.core.bus import EventBus
from trainsight.collectors.gpu_collector import GPUCollector

# Create event bus and collector
bus = EventBus()
collector = GPUCollector()

# Subscribe to GPU events
def on_gpu_stats(event):
    print(f"GPU Util: {event.payload['utilization']}%")

bus.subscribe("gpu.stats", on_gpu_stats)

# Start dashboard
# dashboard = Dashboard(bus, collector)
# dashboard.run()

PyTorch Integration

import torch
from trainsight.integrations.pytorch import TrainSightHook

model = torch.nn.Linear(100, 10)
hook = TrainSightHook(model)

# Hook automatically monitors gradients and activations
output = model(torch.randn(32, 100))
output.backward()

PyTorch Lightning Integration

import pytorch_lightning as pl
from trainsight.integrations.lightning import TrainSightCallback

trainer = pl.Trainer(
    callbacks=[TrainSightCallback()],
    max_epochs=10,
)
trainer.fit(model)

HuggingFace Transformers Integration

from transformers import Trainer, TrainingArguments
from trainsight.integrations.huggingface import TrainSightCallback

training_args = TrainingArguments(
    output_dir="./output",
    callbacks=[TrainSightCallback],
)
trainer = Trainer(model=model, args=training_args)

Architecture

Collectors -> EventBus -> Analyzers -> Predictors -> Dashboard / CLI / Logger

Core modules:

• trainsight/core/event.py - Event types and payloads
• trainsight/core/bus.py - Synchronous event bus
• trainsight/core/async_bus.py - Non-blocking async event bus
• trainsight/core/dispatcher.py - Collector orchestration
• trainsight/collectors/gpu_collector.py - NVIDIA GPU metrics
• trainsight/analyzers/ - Anomaly and bottleneck detection
• trainsight/predictors/oom_predictor.py - OOM prediction
• trainsight/integrations/ - Framework integrations

Configuration

Default config: trainsight/config/default.yaml

mode: full
enable_behavior_learning: true
oom_model: statistical
thermal_limit: 85
refresh_rate: 30

Prometheus Exporter

Enable in config:

enable_prometheus: true
prometheus_port: 9108

Metrics endpoint: http://127.0.0.1:9108/metrics

Official Build Verification

• Soft check (default): warns on signature/hash mismatch.
• Strict mode: refuses startup when verification fails.

Enable strict mode:

export TRAINSIGHT_OFFICIAL_ONLY=1

Or in config:

strict_official_build: true

Plugin System

Create ~/.trainsight/plugins/my_plugin.py:

def register(bus):
    bus.subscribe("gpu.stats", custom_handler)

def custom_handler(event):
    print("Custom plugin:", event.payload)

Debug / Simulation / Replay

trainsight --debug
trainsight --simulate
trainsight --replay gpu_usage_log.csv

Textual inspector:

TEXTUAL_DEVTOOLS=1 trainsight --debug

Testing

Install dev dependencies and run tests:

pip install -e ".[dev]"
pytest -q

Tests are hardware-independent and use pure logic paths.

License

TrainSight Community License v1.0 - See LICENSE for details.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.