0.8.3

• #331: adds a helper to convert an onnx model into dot
• #330: fixes access rope_parameters for transformers>=5
• #329: supports lists with OnnxruntimeEvaluator
• #326: use ConcatFromSequence in LoopMHA with the loop
• #325: adds plug for LoopMHA, extends the unit tests to measure the discrepancies
• #324: supports FunctionProto with arguments in OnnxruntimeEvaluator
• #323: drops torch 2.8 on CI
• #322: support rerunning onnx kernels with torch intermediate results in side-by-side
• #314: fix modelbuilder download needed after this change microsoft/onnxruntime-genai#1862
• #311: use custom and local function to use PackedMultiHeadAttention from onnxruntime
• #310: splits patches into multiple files
• #308: add option --save_ep to dump the exported program as well as torch input
• #304, #306, #316, #317, #318, #319: improves side-by-side comparison, creates command line sbs

0.8.2

• #303: fix inputs for summarization, feature extraction tasks
• #302: adds helpers to analyse onnxruntime profiling
• #297: experiment around a higher ops loop_for
• #292, #293, #294, #295: first version of new patches for Qwen models

0.8.1

• #290: adds one prompt for text2text-generation
• #289: adds command line options --exppo to give the exporter additional options
• #287: adds input 'inputs_prompt' to test a LLM, meant to be used during validation
• #288: add .contiguous in torch.cond branch (attention patch for sdpa implementation)
• #286: adds variable to track random nodes in models

0.8.0

• #283: fix historical aggregation when multiple input sets are used
• #282: add tools to understand better which functions were patched
• #280: fixes patches for sdpa_attention_forward for different version of transformers
• #278: implements onnx_generate_with_genai
• #277: changes the serialization for all caches to reorder the model outputs (key_1, value_1, key_2, ...)
• #276: implements onnx_generate which implements method generate for an onnx model,
• #275: fixes function patched_vmap

0.7.16

• #273: enables export with FakeTensor
• #272: makes patches work with FakeTensor
• #270: add export sample code to export a specific model id with the appropriate inputs
• #269: adds one unit test to track a patch fixing broadcast output shape
• #267: patches sdpa_attention_forward because of a control flow (transformers>=5.0)
• #266: makes patch_torch an integer in torch_export_patches to enable more patches

0.7.15

• #264: allows to validate a model with inputs defined from another task
• #261: first updates to support transformers>=5.0, not yet completed

0.7.14

• #257: patch to disable one exception in pytorch
• #256: extract subfolder from modelid//subfolder
• #252: adds new sets of inputs for task texgt-generation
• #250: add variables to track sequence nodes
• #249: patches _maybe_broadcast to support a corner case

0.7.13

• #247: supports more gemma models with ModelBuilder
• #246: add a set of inputs checking models works for an empty cache on task text-generation
• #237: dummy inputs for google/gemma-3-4b-it (task image-text-to-text)
• #244: add a patch to bypass the exception raised when the dynamic dimension is in {0,1}

0.7.12

• #232: fixes --patch argument so that --patch=0 works
• #231: better statistics about fusions
• #227: better support for model_id//pretrained, adds speed up when running command validate
• #226: fix input order for models created with modelbuilder

0.7.11

0.7.11

• #224: support model_id with // to specify a subfolder
• #223: adds task image-to-video
• #220: adds option --ort-logs to display onnxruntime logs when creating the session
• #220: adds a patch for PR #40791 huggingface/transformers#40791_ in transformers

0.7.10

• #218: patches used sdpa_mask_recent_torch used from _vmap_for_bhqkv

0.7.9

• #214: fix modelbuilder export
• #213: use DYNAMIC on batch size

0.7.8

• #210: add utilities to investigate models
• #208: add a patch for Qwen3 (rewrite a loop)