v0.1.5 🎄 2024/12/19 🎄 Benchmarks & Speedups

• Added internal/bechmarks package: See progress in https://docs.google.com/spreadsheets/d/1ikpJH6rVVHq8ES-IA8U4lkKH4XsTSpRyZewXwGTgits/edit?gid=1753191050#gid=1753191050
  • Benchmark ONNX models with XLA, ONNX Runtime (ORT), CPU and GPU
  • Very simple models
  • KnightsAnalytics/all-MiniLM-L6-v2
  • Slices (parts of) KnightsAnalytics/all-MiniLM-L6-v2
• Updated dependencies to GoMLX 0.16.1 with lots of accelerations.

v0.1.4 Added Flatten Op

• Added Flatten op support.

v0.1.3 Saving back to ONNX model after fine-tuning

• Added ContextToONNX to save variables back to ONNX model (in memory).
• Refactored internal/togomlx to inside onnx/ subdir.
• Added Model.Write and Model.SaveToFile.

v0.1.2 Added LSTM support.

• Updated dependency to GoMLX v0.15.2

v0.1.1 New ops, guess variables being used as constants

• Assume some variables are constant during constant-expression evaluation.
• Improved pretty-printing of attributes: include their values for small values.
• New ops: Range, Tile, CumSum, Not, Tanh, GatherElements, several standard unary and binary operators.
• Fixed ops: Where.

v0.1.0

API, basic conversion mechanism, on-the-fly constant expression evaluation.

Models already successful converted:

• "sentence-transformers/all-MiniLM-L6-v2",
• "KnightsAnalytics/distilbert-base-uncased-finetuned-sst-2-english",
• "KnightsAnalytics/distilbert-NER",

But still missing many Ops before most models can be converted.