lmstrix 1.0.64

A toolkit for managing and testing LM Studio models with automatic context limit discovery

LMStrix

LMStrix is a professional Python toolkit designed to supercharge your interaction with LM Studio. It provides a powerful command-line interface (CLI) and Python API for managing, testing, and running local language models, with a standout feature: Adaptive Context Optimization.

Key Features

• 🔍 Automatic Context Discovery: Binary search algorithm to find the true operational context limit of any model
• 📊 Beautiful Verbose Logging: Enhanced stats display with emojis showing inference metrics, timing, and token usage
• 🚀 Smart Model Management: Models persist between calls to reduce loading overhead
• 🎯 Flexible Inference Engine: Run inference with powerful prompt templating and percentage-based output control
• 📋 Comprehensive Model Registry: Track models, their context limits, and test results with JSON persistence
• 🛡️ Safety Controls: Configurable thresholds and fail-safes to prevent system crashes
• 💻 Rich CLI Interface: Beautiful terminal output with progress indicators and formatted tables

Installation

# Using pip
pip install lmstrix

# Using uv (recommended)
uv pip install lmstrix

Quick Start

Command-Line Interface

# Scan for available models in LM Studio
lmstrix scan

# List all models with their context limits and test status
lmstrix list

# Test context limit for a specific model
lmstrix test llama-3.2-3b-instruct

# Test all untested models with safety threshold
lmstrix test --all --threshold 102400

# Run inference with enhanced verbose logging
lmstrix infer "What is the capital of Poland?" -m llama-3.2-3b-instruct --verbose

# Run inference with percentage-based output tokens
lmstrix infer "Explain quantum computing" -m llama-3.2-3b-instruct --out_ctx "25%"

# Use file-based prompts with templates
lmstrix infer summary -m llama-3.2-3b-instruct --file_prompt adam.toml --text_file document.txt

# Direct text input for prompts
lmstrix infer "Summarize: {{text}}" -m llama-3.2-3b-instruct --text "Your content here"

Enhanced Verbose Output

When using --verbose, LMStrix provides comprehensive statistics:

════════════════════════════════════════════════════════════
🤖 MODEL: llama-3.2-3b-instruct
🔧 CONFIG: maxTokens=26214, temperature=0.7
📝 PROMPT (1 lines, 18 chars): Capital of Poland?
════════════════════════════════════════════════════════════
⠸ Running inference...
════════════════════════════════════════════════════════════
📊 INFERENCE STATS
⚡ Time to first token: 0.82s
⏱️  Total inference time: 11.66s
🔢 Predicted tokens: 338
📝 Prompt tokens: 5
🎯 Total tokens: 343
🚀 Tokens/second: 32.04
🛑 Stop reason: eosFound
════════════════════════════════════════════════════════════

Python API

from lmstrix.loaders.model_loader import load_model_registry
from lmstrix.core.inference_manager import InferenceManager

# Load model registry
registry = load_model_registry()

# List available models
models = registry.list_models()
print(f"Available models: {len(models)}")

# Run inference
manager = InferenceManager(verbose=True)
result = manager.infer(
    model_id="llama-3.2-3b-instruct",
    prompt="What is the meaning of life?",
    out_ctx=100,
    temperature=0.7
)

if result["succeeded"]:
    print(f"Response: {result['response']}")
    print(f"Tokens used: {result['tokens_used']}")
    print(f"Time: {result['inference_time']:.2f}s")

Context Testing & Optimization

LMStrix uses a sophisticated binary search algorithm to discover true model context limits:

Safety Features

• Threshold Protection: Configurable maximum context size to prevent system crashes
• Progressive Testing: Starts with small contexts and increases safely
• Persistent Results: Saves test results to avoid re-testing

Testing Commands

# Test specific model
lmstrix test llama-3.2-3b-instruct

# Test all models with custom threshold
lmstrix test --all --threshold 65536

# Test at specific context size
lmstrix test --all --ctx 32768

# Reset and re-test a model
lmstrix test llama-3.2-3b-instruct --reset

Model Management

Registry Commands

# Scan for new models
lmstrix scan --verbose

# List models with different sorting
lmstrix list --sort size        # Sort by size
lmstrix list --sort ctx         # Sort by tested context
lmstrix list --show json        # Export as JSON

# Check system health
lmstrix health --verbose

Model Persistence

Models stay loaded between inference calls for improved performance:

• When no explicit context is specified, models remain loaded
• Last-used model is remembered for subsequent calls
• Explicit context changes trigger model reloading

Prompt Templating

LMStrix supports flexible prompt templating with TOML files:

# adam.toml
[aps]
prompt = """
You are an AI assistant skilled in Abstractive Proposition Segmentation.
Convert the following text: {{text}}
"""

[summary] 
prompt = "Create a comprehensive summary: {{text}}"

Use with CLI:

lmstrix infer aps --file_prompt adam.toml --text "Your text here"
lmstrix infer summary --file_prompt adam.toml --text_file document.txt

Development

# Clone repository
git clone https://github.com/twardoch/lmstrix.git
cd lmstrix

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

# Run tests
pytest

# Run linting
hatch run lint:all

Project Structure

src/lmstrix/
├── cli/main.py              # CLI interface
├── core/
│   ├── inference_manager.py # Unified inference engine
│   ├── models.py            # Model registry
│   └── context_tester.py    # Context limit testing
├── api/client.py            # LM Studio API client
├── loaders/                 # Data loading utilities
└── utils/                   # Helper utilities

Features in Detail

Adaptive Context Optimizer

• Binary search algorithm for efficient context limit discovery
• Safety thresholds to prevent system crashes
• Automatic persistence of test results
• Resume capability for interrupted tests

Enhanced Logging

• Beautiful emoji-rich output in verbose mode
• Comprehensive inference statistics
• Progress indicators for long operations
• Clear error messages with context

Smart Model Management

• Automatic model discovery from LM Studio
• Persistent registry with JSON storage
• Model state tracking (loaded/unloaded)
• Batch operations for multiple models

Requirements

• Python 3.11+
• LM Studio installed and configured
• Models downloaded in LM Studio

License

MIT License - see LICENSE file for details.

Contributing

Contributions welcome! Please read our contributing guidelines and submit pull requests for any improvements.