0.9.3

• #422: add remove_inputs to InputObserver
• #421: fix a few patches for MoE
• #426: remove MambaCache

0.9.2

• #415: improves function make_model_with_local_functions to support ill-defined partitions
• #413: fix InputObserver in the generic case
• #412: patches for ViTModel (through rewriting)

0.9.1

• #410: add patch for _get_range_constraints
• #409: improves ModelBuilder wrapper
• #408: fix torch_deepcopy for empty DynamicCache and transformers==5.1.0, 5.2.0 (see huggingface/transformers#43765)

0.9.0

• #403: update the serialization of SlidingWindowCache to include parameter sliding_window, patch for sdpa_mask
• #400, #401, #402: improves InputObserver (investigations), add it the documentation

0.8.11

• #396: fix serialization for DynamicCache with different layer classes
• #394: add function make_model_with_local_functions to partition a model into local functions

0.8.10

• #384: add weights_only=False when using torch.load

0.8.9

• #383: removed bool, int, float, None as input dummies for the exporter in method_to_onnx
• #382: make the ordering of the inferred dynamic shapes more robust
• #381: add parameter expand_batch_for to method_to_onnx
• #378: implements the computation of discrepancies in method_to_onnx
• #379: update the handling of cache after the removal of HybridCache, SlidingWindowCache in transformers>=5,

0.8.8

• #375: export a method to onnx in order to export using method generate
• #376: fix patched lazy_initialization for transformers>=5
• #372: fix patch on rotary embedding
• #371: fix make_fake_with_dynamic_dimensions

0.8.7

• #366: add command line to optimize a model
• #363: patch for DynamicDimConstraintPrinter
• #360, #364: preliminary work for phi4

0.8.6

• #357: complete simple_loop_for, an easier to rewrite loops
• #356: include qwen embedding part
• #355: better command line to export models
• #353, #354: add command line to compare two onnx models