The Evolution of ONNX: Driving Interoperability in AI...
summary
ONNX becomes a practical bridge: TurnkeyML converts PyTorch (example Python) to ONNX, auto-recognizes model inputs, optimizes and benchmarks across runtimes (ONNX Runtime, TensorRT, DirectML) and saves a detailed cache and stats file. The project converts about 2,000 new models into the ONNX Model Zoo (~10x increase). Plugins enable vendor execution providers (AMD/Intel/Ryzen AI), profilers (VTune, UPR, nvidia-smi), and optimizations like quantization/sparsity; Intel reports AMX-tuned matrix multiplies ~4x faster on Gen4.
AI-written summary · about
chapters (11)
- 0:00ONNX introduction
- 2:50Turnkey ML repository
- 5:18Turnkey ML workflow
- 9:31Demo: Llama with Turnkey ML
- 11:56AMD’s open source investments
- 13:28ONNX benefits for devs
- 16:59Cloud deployment scenarios
- 22:42Edge deployment challenges
- 25:49Intel’s ONNX optimizations
- 27:56ONNX runtime GPU/NPU
- 29:02Summary and future
related talks
45mLeverage Power of Machine Learning with ONNX
05mONNX Runtime
30mScalable ML acceleration with ONNX Runtime
33mWhat on earth is ONNX and ONNXRUNTIME
05mPyTorch ONNX Export Support
41mLeverage Power of Machine Learning with ONNX
54mOne Universal Language for All Machine Learning Models
45mMaking Machine Learning Applications Fast and Simple with...