trainsight 0.1.0

AI Training Intelligence Dashboard

TrainSight

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

Architecture

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

Core modules:

• trainsight/core/event.py
• trainsight/core/bus.py
• trainsight/core/dispatcher.py
• trainsight/collectors/gpu_collector.py
• trainsight/analyzers/regression_anomaly.py
• trainsight/predictors/oom_predictor.py
• trainsight/plugins/loader.py

Installation

pip install .

CLI Usage

trainsight

HuggingFace Integration

from trainsight import Dashboard
from trainsight.integrations.huggingface import TrainSightCallback

See examples/huggingface_example.py.

Plugin Example

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

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


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

Config Example

Default config: trainsight/config/default.yaml

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

Official Build Verification

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

Enable strict mode with either:

export TRAINSIGHT_OFFICIAL_ONLY=1

or in config:

strict_official_build: true

Prometheus Exporter

Enable in config:

enable_prometheus: true
prometheus_port: 9108

Metrics endpoint:

http://127.0.0.1:9108/metrics

Test Framework

Install dev deps and run tests:

pip install -e .[dev]
pytest -q

Tests are hardware-independent and use pure logic paths.

Debug / Simulation / Replay

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

Textual inspector:

TEXTUAL_DEVTOOLS=1 trainsight --debug