Skip to content

RELEASE_NOTES.md

Latest commit

 

History

History
151 lines (107 loc) · 5.21 KB

RELEASE_NOTES.md

File metadata and controls

151 lines (107 loc) · 5.21 KB

Release Notes

v0.3.1 - MAX 26.2.0 Update (January 2026)

MAX/Mojo Version: 26.2.0.dev2026012505 (locked)

What's New

Breaking API Changes

  • Tensor → Buffer rename: Updated all max.driver.Tensor imports to max.driver.Buffer
    • Runtime data handling now uses Buffer class
    • Graph specifications still use TensorType (unchanged)
    • All 49 tests passing with updated API

New Features

  • 🎉 Apple Silicon GPU matmul support!
    • Linear layer operations now work on Apple Silicon GPU
    • MAX nightly includes fallback path for matrix multiplication
    • Previously blocked ops.matmul now functional on GPU
    • Verified working: pixi run python examples/python/02_linear_layer/linear_layer.py --device gpu

Tooling

  • Automated version update script: scripts/update_max_version.py
    • Safely updates MAX to latest nightly with automatic rollback
    • Runs full test suite before committing version change
    • Saves detailed failure reports for breaking changes
    • Documented in scripts/README.md

Requirements

  • MAX/Mojo: 26.2.0.dev2026012505 (locked in pixi.toml)
  • Pixi package manager
  • Python 3.11+

Migration from v0.3.0

To update from v0.3.0:

  1. Replace from max.driver import Tensor with from max.driver import Buffer
  2. Replace Tensor.from_numpy() with Buffer.from_numpy()
  3. Keep TensorType imports unchanged

v0.3.0 - Public Release (January 2026)

MAX/Mojo Version: 26.1.0.dev2026010718 (locked)

Status: Superseded by v0.3.1

What's New

Six Progressive Examples

  • 1️⃣ Element-wise Operations - Basic graph construction with mul, add, relu
  • 2️⃣ Linear Layer - Matrix operations with matmul, transpose
  • 3️⃣ DistilBERT Sentiment - Production transformer (66M parameters, 5.58x speedup vs PyTorch)
  • 4️⃣ MLP Regression - Multi-layer perceptron for housing prices
  • 5️⃣ CNN MNIST - Convolutional network for digit classification
  • 6️⃣ RNN Forecast - (WIP) Parked due to API limitations

Minimal Examples

Each example includes a minimal version showing pure MAX Graph API without abstractions:

  • elementwise_minimal.py - 116 lines of clear, educational code
  • linear_layer_minimal.py - 140 lines showing matrix operations

Infrastructure Improvements

  • Comprehensive testing: 49 pytest tests with correctness validation
  • Performance benchmarks: MAX vs PyTorch comparisons for all models
  • Package restructuring: All Python modules installable, no sys.path manipulation
  • Systematic benchmarking: TOML configs, MD/JSON/CSV reports with machine IDs
  • Version locking: MAX version locked to prevent breaking API changes
  • Code quality: Pre-commit hooks, ruff formatting

Documentation

  • Learning journey framing: Repository positioned as one developer's exploration
  • Mojo context: Documentation explaining Mojo Graph API deprecation (May 2025)
  • GPU findings: Apple Silicon GPU experiments (element-wise ops working, matmul blocked)
  • Terminology feedback: Submitted InferenceSession clarification to Modular docs team

Performance Highlights

  • DistilBERT (M1 CPU): 5.58x faster than PyTorch (45.88ms vs 255.85ms)
  • MLP Regression: PyTorch 253x faster (honest reporting of both wins and losses)
  • CNN MNIST: PyTorch ~5x faster (both produce identical predictions)
  • Apple Silicon GPU: Element-wise ops working ✅, matmul now working ✅

Requirements

  • MAX/Mojo: 26.1.0.dev2026010718 (locked in pixi.toml)
  • Pixi package manager
  • Python 3.11+

Breaking Changes

None - this is the initial public release.

Known Limitations

  • RNN example parked due to MAX Graph API sequence processing limitations
  • MAX slower than PyTorch on some workloads (MLP, CNN)

v0.2.0 - MAX Graph DistilBERT (December 2025)

Internal version - not publicly released

Features

  • Full MAX Graph implementation of DistilBERT
  • 5.58x speedup over PyTorch on M1
  • Comprehensive documentation and guides
  • Numbered examples for learning
  • Apple Silicon GPU experiments (element-wise ops working)

v0.1.0 - Lexicon-based Baseline (November 2025)

Internal version - not publicly released

Features

  • Pure Mojo sentiment classifier
  • Simple lexicon-based approach
  • Benchmarking foundation

Version Support Policy

MAX API Version Locking: This repository locks the MAX version to prevent breaking changes. Users should:

  1. Use the locked version for guaranteed compatibility
  2. Update modular = "==X.X.X" in pixi.toml when ready to modernise
  3. Test all examples after updating MAX version
  4. Report any breaking changes as GitHub issues

Learning Resource: This is a learning repository documenting one developer's exploration of MAX. Not an authoritative guide.

Changelog Format

We follow these principles:

  • Honest reporting: Document both performance wins and losses
  • Version transparency: Lock and document MAX versions used
  • Learning focus: Emphasize educational value over production readiness
  • Community engagement: Welcome corrections and improvements

Maintained by: DataBooth
Licence: Apache 2.0
Repository: https://github.com/DataBooth/max-learning