pt-snap-cli

中文文档 | English

A command-line tool for analyzing PyTorch memory snapshots. Set a snapshot database, run built-in queries, and inspect memory usage, leaks, and timelines.

Installation

pip install -e .

Quick Start

To start from a raw PyTorch memory snapshot, import the pickle into a SnapshotDB with the built-in backend, then list the available query templates:

pt-snap import snapshot.pkl
pt-snap query --list

# Set the snapshot database and device
pt-snap focus examples/snapshot_expandable.pkl.db --device 0

# List available queries
pt-snap query --list

# Run a query (automatically uses the focused device)
pt-snap query --template-use memory_peak

# Detect potential memory leaks
pt-snap query --template-use leak_detection --params '{"min_size": 1024}'

See the full quick start guide for a walkthrough.

Commands

MCP Server

pt-snap-cli provides an MCP (Model Context Protocol) server so AI agents can interact with PyTorch memory snapshots programmatically.

# Start the MCP server
pt-snap-mcp

The server exposes the following tools:

See the MCP guide for setup and usage details.

Documentation

Query Templates

6 built-in templates across 3 categories:

• Basic: block, event, allocation
• Statistical: callstack_analysis, memory_peak
• Business: leak_detection

See Querying for details.

Project Structure

pt-snap-cli/
├── src/
│   └── pt_snap_cli/
│       ├── cli.py              # CLI entry point
│       ├── context.py          # Database context manager
│       ├── config.py           # Focus management
│       ├── api.py              # Python API layer
│       ├── query/
│       │   ├── builder.py      # Query builder
│       │   ├── executor.py     # Query executor
│       │   ├── mapper.py       # Result mapper
│       │   ├── registry.py     # Query registry
│       │   ├── condition.py    # Query conditions
│       │   ├── config.py       # Query configuration
│       │   └── templates/      # Query templates
│       ├── models/             # Data models
│       └── mcp/                # MCP server for agent integration
├── tests/                  # Test files
├── examples/               # Example data
└── docs/                   # Documentation

Development

pytest                           # Run all tests
./tests/run_tests.sh             # Full test run with coverage
black . && ruff check .          # Format and lint
python -m build                  # Build sdist and wheel

CLI/MCP semantic contract

The CLI and MCP server must keep using the shared API/core services as the single semantic source of truth. The CLI may render results as terminal text and the MCP server may return JSON-like tool payloads, but normalized meanings must stay in sync for focus handling, query template listing, template metadata, query execution, and shared error cases.

tests/test_contract_cli_mcp.py compares normalized CLI output with MCP tool results on the same inputs and fixtures. When adding a new shared CLI/MCP capability, update or add a contract case in that file so CI fails if either adapter drifts from the shared behavior.