v3.4.0

• Features:
  • (OpenVINO) Added GroupedMatMul support to the AWQ and Scale Estimation algorithms, enabling data-aware weight compression of grouped_mm-based MoE models (#4176).
• Fixes:
  • (OpenVINO) Extended the RoPE ignored pattern to recognize the matmul -> transpose -> cos/sin structure used by models such as GPT-OSS (#4175).
  • (PyTorch) Removed the __getitem__ node that follows split-like operations (e.g. chunk) from the TorchFX inference graph, fixing statistic collection errors for models such as YOLO11 (#4155).
  • (ONNX) Fixed a KeyError in bias attribute resolution when a tensor is a model input and therefore has no parent node (#4169).
  • (ONNX) Fixed squeeze axis determination for ONNXEmbeddingMetatype nodes (#4144).
• Improvements:
  • (PyTorch) Replaced the deprecated export_for_training with export in TorchFX examples, documentation, and tests (#4191).
• Tutorials:
  • Post-Training Optimization of Ministral-3 Models
  • Post-Training Optimization of Muse Glimmer-30B Model
• Requirements:
  • Updated torch to 2.13.0 (#4197).
  • Updated torchao to 0.18.0 (#4196).
  • Updated transformers to 5.5.0, optimum to 2.3.0, and optimum-intel to 2.1.0 (#4200).
  • Updated onnx-ir to 1.0.0 and onnxscript to 0.7.1 (#4190).
  • Updated datasets to 5.0.1 (#4182).
  • Updated the pydot upper bound to <= 4.0.1 (#4181).