What's Changed

• Fix WhisperProcessor divide-by-zero when single prompt is provided by @Copilot in #2068
• Fix lm_head tensor loading order dependency in quantized model builder by @thpereir in #2061
• Fail to build Whisper model by @xiaofeihan1 in #2075
• Rename NemotronCacheConfig to NemotronConfig and add blank penalty to the decoder by @nenad1002 in #2042
• Fix YaRN RoPE bugs in model builder and add parity tests by @titaiwangms in #2076
• Add Transformers v5 Support by @sayanshaw24 in #2089
• macOS ARM64 ADO pipeline by @Copilot in #2091
• Reduce CPU-side per-token overhead in GenerateNextToken and SampleTopP by @hanbitmyths in #2085
• Add onStageComplete by @apsonawane in #2074
• [WebGPU] Support continuous decoding (RewindTo) with graph capture by @qjia7 in #2083
• [Mistral3] Add VLM support with multi-image inference by @titaiwangms in #2077
• Add k_quant_linear mixed-precision quantization for hybrid attention … by @apsonawane in #2100
• Removes QNN packaging from onnxruntime-genai pipelines by @baijumeswani in #2109
• Add Gemma4 multimodal support (vision + audio) by @apsonawane in #2103
• Update GUIDs during az login by @kunal-vaishnavi in #2122
• Add CODEOWNERS file for repository ownership by @kunal-vaishnavi in #2119
• Qwen3.5: drop fp32 cast around RMSNorm in builder by @xiaofeihan1 in #2101
• Add support for LFM2 in ORT GenAI by @xenova in #1979
• Enable CUDA graph capture for CUDA EP to improve decode throughput by @apsonawane in #2070
• [Qwen3.5] dedup position ids by @daijh in #2102
• Address win-cuda pipeline errors by @baijumeswani in #2154
• Update Extensions Commit to Fix Id2Token Bugs by @sayanshaw24 in #2159
• Limit the CUDA cmake architectures to 86 for CI builds by @baijumeswani in #2161
• Gate leaked-object error reporting in Shutdown() to debug builds or when logging is enabled by @baijumeswani in #2162
• Update Copilot instructions for reviewing model builder by @kunal-vaishnavi in #2164
• Fix DecoderState input_ids check regression introduced in #2103 by @titaiwangms in #2148
• Fix memory leaks by @skottmckay in #2153
• [Qwen3.5] Use LpNormalization for L2-norm in linear-attention Q/K by @xiaofeihan1 in #2127
• Fix: Win32 build failure when paths contain spaces by @nsubaru in #2053
• Fix CUDA build with MSVC by enabling /Zc:preprocessor for nvcc host compilation on VS 16.5 or greater by @nsubaru in #2054
• Apply linear rope_scaling in model builder for Neutts/nano by @VishalX in #2142
• Fix Quark/AWQ weight loading for Qwen3-VL-4B text model by @anilmartha in #2143
• Fix WebGPU inference crash in embedding and multi-modal feature allocation by @feich-ms in #2163
• Support Visual Studio 18 2026 build by @Copilot in #2017
• Add QNN EP documentation to OGA including Genie note by @qti-kromero in #2158
• Use windowsml package and make winml usage simpler by @baijumeswani in #2155
• Cleanup TensorObject created by OrtxTensorResultGetAt by @skottmckay in #2168
• Fix nemotron leaks by @skottmckay in #2169
• [RyzenAI] make speech sub-model optional in PhiMultiModalProcessor by @manasablrm in #2167
• Enable graph capture for WebGPU models and DML continuous decoding tests by @qjia7 in #2099
• [Qwen3] Allow packed QKV MatMul under QK-Norm via post-MatMul Split by @xiaofeihan1 in #2137
• Enable Linux ARM64 builds and packaging by @baijumeswani in #2107
• Add gemma4 unit tests by @apsonawane in #2151
• Auto-detect fixed kv-cache shape in DefaultKeyValueCache by @akholodnamdcom in #2166
• Add text-only mode support for Qwen 3.5 model builder by @apsonawane in #2157
• Fix heap overflow issue by @apsonawane in #2110
• [Benchmark] Add --use_random_tokens flag to C benchmark by @VishalX in #2170
• Add HunYuan Dense V1 (hunyuan_v1_dense) model support by @anilmartha in #2144
• Nvidia Parakeet Tdt ASR support by @nenad1002 in #2150
• Multilingual Streaming Nemotron ASR + CUDA support by @nenad1002 in #2171
• Add Csharp binding for Multi-lingual ASR by @rui-ren in #2176
• Add VideoChat-Flash (OpenGVLab) language model support by @anilmartha in #2147
• Update Nemotron ASR docs by @rui-ren in #2178
• Validate sliding window size before creating KV cache by @baijumeswani in #2181
• Fix external weights loading for in-memory models without changing cwd by @baijumeswani in #2180
• Enable Qwen3.5 TRT-RTX EP path with CUDA graph by @yen-shi in #2139
• Add Qwen3.5-MoE (35B-A3B) model support by @tanzeel-amd in #2146
• Update ort-extensions commit by @baijumeswani in #2182

New Contributors

• @titaiwangms made their first contribution in #2076
• @nsubaru made their first contribution in #2053
• @VishalX made their first contribution in #2142
• @anilmartha made their first contribution in #2143
• @feich-ms made their first contribution in #2163
• @qti-kromero made their first contribution in #2158
• @manasablrm made their first contribution in #2167
• @yen-shi made their first contribution in #2139
• @tanzeel-amd made their first contribution in #2146

Full Changelog: v0.13.1...v0.14.0

What's Changed

• update WebGPU buffer memory info name by @fs-eire in #1957
• Add enable_profiling in Runtime Options by @xiaofeihan1 in #1949
• Fix uninitialized tools variable and improve exception debug messages by @sheller-ms in #1971
• Add common download to Phi-3 tutorial by @kunal-vaishnavi in #1973
• Add support for InternLM2 model architecture by @amdrajeevp1 in #1958
• Update cmake cuda architecture and use win-arm64 pool workaround by @baijumeswani in #1976
• Update examples after 0.12.0 release by @kunal-vaishnavi in #1980
• Add CI pipeline for WebGPU EP model testing by @qjia7 in #1956
• Fix Python nightly build by @kunal-vaishnavi in #1981
• Add missing Quark 0.11 weight patterns for ChatGLM3 output layer by @poganesh in #1983
• Support Qwen2.5-VL pre-quantized models in qwen.py by @poganesh in #1985
• [VitisAI] external_ep_libray support fix for WinML by @akholodnamdcom in #1984
• Fix guidance bug by @baijumeswani in #1988
• Fix incorrect batch responses when using multiple prompts by @lnigam in #1986
• Enable webgpu graph capture in base.py by @qjia7 in #1991
• Harden CUDA error checking across the codebase by @Copilot in #1994
• allow pruned models for prefill by @fs-eire in #1995
• Fix WinML Packaging Pipeline by @baijumeswani in #1998
• Add small changes after pruning prefill by @kunal-vaishnavi in #2000
• webgpu: Optimize Copyfrom by @qjia7 in #1992
• Add support for CUDA 13 by @baijumeswani in #2001
• add webgpu to qmoe path by @guschmue in #2005
• Fix ERNIE 4.5 model builder: rope_attrs and config architecture name by @xiaoyao9184 in #2007
• Bug fix in Continuous Decoding by @chilukam-qti in #2008
• Update Phi-4 mm README links by @kunal-vaishnavi in #2014
• Add Qwen3-VL model support + multi-image input support in Qwen VL family by @hanbitmyths in #2003
• Add Qwen3.5 model support and optimize multi-image handling by @apsonawane in #2019
• Reuse a single generator via RewindTo(0) in benchmark instead of creating multiple generators by @qjia7 in #2002
• [RyzenAI] WinML compatibility fix by @akholodnamdcom in #2026
• Nemotron ASR Support for Streaming by @nenad1002 in #1997
• [WebGPU] Fix the prefill regression when graph capture is ON by @qjia7 in #2021
• Support 4 inputs for nemotron model by @jiafatom in #2036
• Updated java packaging based on python packaging logic by @EPNW-Eric in #2029
• Fix android packaging pipeline by @baijumeswani in #2039
• Add OpenAI's Whisper to model builder by @kunal-vaishnavi in #2018
• [Java] Add a dependency on onnxruntime (#2030) by @EPNW-Eric in #2040
• Fix mutually exclusive inputs for language models by @kunal-vaishnavi in #2046
• Decouple plugin execution providers (EPs) from the USE_WINML pre-processor macro by @baijumeswani in #2038
• Route pipeline model RunOptions through SetRunOption for proper special key handling by @Copilot in #2044
• Add ort_build_version and ort_build_source parameters to nuget and python packaging pipelines, remove ROCm support by @Copilot in #2049
• Add batched multi-image vision path and window_size config for Qwen VL by @hanbitmyths in #2050
• docs: fix formatting and syntax highlighting in documentation by @riddles-the-one in #2051
• Add Silero VAD Support to Nemotron Streaming ASR by @sayanshaw24 in #2035
• Add Qwen3.5 hybrid decoder export support (GatedDeltaNet + Attention) by @apsonawane in #2043
• Add support for QNN stateful models by @qti-ashimaj in #2012
• Allocate recurrent state via device allocator to enable CUDA graph capture by @apsonawane in #2057
• Speed up CI pipelines by @Copilot in #2052
• Fix tool calling for TRT-RTX models by @kunal-vaishnavi in #2048
• Fix vision pipeline EP hardcoding and pixel_values rank mismatch for Qwen VL models by @apsonawane in #2060

New Contributors

• @sheller-ms made their first contribution in #1971
• @amdrajeevp1 made their first contribution in #1958
• @poganesh made their first contribution in #1983
• @xiaoyao9184 made their first contribution in #2007
• @chilukam-qti made their first contribution in #2008
• @EPNW-Eric made their first contribution in #2029

Full Changelog: v0.12.0...v0.13.0

• Update examples after 0.12.0 release
• Add missing Quark 0.11 weight patterns for ChatGLM3 output layer
• Support Qwen2.5-VL pre-quantized models in qwen.py
  
• Fix incorrect batch responses when using multiple prompts
  
• Harden CUDA error checking across the codebase
• allow pruned models for prefill
• Add small changes after pruning prefill
  

• #1988
• #1984

What's Changed

• Update versions after making 0.11.0 branch by @kunal-vaishnavi in #1867
• Fix guidance usage in continuous decoding by @kunal-vaishnavi in #1870
• Fix HelloPhi C# example by @kunal-vaishnavi in #1871
• Fix regex by @apsonawane in #1875
• Update extensions commit by @apsonawane in #1874
• Revert removal of eps_without_if_support by @xiaofeihan1 in #1878
• Fix condition for NPU by @apsonawane in #1880
• Model builder refactoring by @tianleiwu in #1862
• Add lintrunner to format code by @tianleiwu in #1884
• Remove empty submodule leftover. by @xkszltl in #1883
• Fix build for lack of RTLD_DI_ORIGIN support by @jaeyoonjung in #1888
• Enable graph capture for webgpu by @qjia7 in #1848
• Generic shared emb_tokens/lm_head implementation by @jixiongdeng in #1885
• Fix bug in Squeeze for getting the value of total_seq_len by @Honry in #1886
• Extra_options disable_qkv_fusion to untie qkv_projs from upstream choice by @jixiongdeng in #1893
• Fix mac pipeline by @apsonawane in #1904
• whisper: Support a variant of the whisper pipeline where encoder / decoder are stateful. by @RyanMetcalfeInt8 in #1857
• Add model builder for Qwen2_5_VLTextModel by @tianleiwu in #1882
• Integrate FARA-7B model by @apsonawane in #1902
• Fix gpt-oss model export by @apsonawane in #1861
• OpenVINO: Add support for model caching via 'cache_dir' provider option by @RyanMetcalfeInt8 in #1900
• WinML - Remove the inclusive Microsoft.WindowsAppSDK.ML range check by @chrisdMSFT in #1907
• Run the model in text mode by @apsonawane in #1908
• Update extensions commit by @apsonawane in #1914
• Fix gpt-oss export by @apsonawane in #1915
• Support Olive new uint8 quantization format by @xiaoyu-work in #1916
• Disable CUDA graph for Phi LongRoPE models with IF nodes on TRT-RTX by @anujj in #1921
• Add support for CUDA and CPU arch for Qwen-2.5-VL and Fara-7B by @apsonawane in #1919
• Add Gemma-3 vision tutorial to ONNX Runtime GenAI by @kunal-vaishnavi in #1793
• Quark GPT-OSS support by @thpereir in #1903
• Fix sliding window alignment regression in QNN models by @apsonawane in #1938
• AMD RyzenAI EP Support by @akholodnamdcom in #1935
• Update README by @natke in #1934
• [RyzenAI] Non-pruned models backward compatibility by @akholodnamdcom in #1942
• [VitisAI] EP loader by @akholodnamdcom in #1918
• Set default top_k and top_p if it is None by @xiaoyu-work in #1944
• Ensure dlls are signed in the c and nuget packages. by @baijumeswani in #1947
• Bump torch from 2.7.1 to 2.7.1+cpu in /test/python/directml/torch by @dependabot[bot] in #1868
• Add linker flags for 16 KB page size on Android by @sheetalarkadam in #1860
• Only manually load DLLs if onnxruntime.dll is not already loaded. by @chemwolf6922 in #1800
• Add a doc showing how to run GPT OSS 20B with WebGPU by @natke in #1945
• Add C#, Java, and Objective-C APIs for Config by @kunal-vaishnavi in #1946
• Fix GatherBlockQuantized node to support symmetric quantized LM_HEAD by @sushraja-msft in #1951
• Fix QMoE blockwise quantization support for TRT-RTX execution provider by @anujj in #1926
• Revert "Add a doc showing how to run GPT OSS 20B with WebGPU" by @kunal-vaishnavi in #1950
• Add custom model path support for unit tests by @mpasumarthi-git in #1917
• fix: patch llguidance to remove reference to ring crate by @sanaa-hamel-microsoft in #1948
• Implement graph models for EPs by @qjia7 in #1895
• Update handling EOS token id detection by @kunal-vaishnavi in #1925
• Remove onnxruntime-genai-cuda from the foundry package by @baijumeswani in #1954
• Include linux builds in the foundry ort-genai package by @baijumeswani in #1955
• Support pre-registered plug-in NvTensorRtRtx execution provider library by @anujj in #1889
• [RyzenAI] Linux compatibility fixes by @akholodnamdcom in #1959
• Use cuda 12.8 to build ort-genai by @baijumeswani in #1960
• Bump protobuf from 5.29.5 to 6.33.5 in /test/python by @dependabot[bot] in #1961
• Add RAII wrappers for ORT Model Editor API types by @qjia7 in #1953
• Rewrite all examples using standardization by @kunal-vaishnavi in #1939
• Add versioning to the onnxruntime-genai-cuda.dll by @baijumeswani in #1965
• [Build][Packaging] macOS packaging to skip building x86_64 by @baijumeswani in #1966
• Sync packaging changes with ONNX Runtime by @baijumeswani in #1967
• Release 0.12.0 cherry-pick PR by @baijumeswani in #1978

New Contributors

• @xkszltl made their first contribution in #1883
• @jaeyoonjung made their first contribution in #1888
• @jixiongdeng made their first contribution in #1885
• @Honry made their first contribution in #1886
• @thpereir made their first contribution in #1903
• @akholodnamdcom made their first contribution in #1935
• @sheetalarkadam made their first contribution in #1860
• @sanaa-hamel-microsoft made their first contribution in #1948

Full Changelog: v0.11.4...v0.12.0

What's Changed

• WinML - Remove the inclusive Microsoft.WindowsAppSDK.ML range check by @chrisdMSFT in #1907
• Run the model in text mode by @apsonawane in #1908

Full Changelog: v0.11.3...v0.11.4

What's Changed

• Model builder refactoring by @tianleiwu in #1862
• Add lintrunner to format code by @tianleiwu in #1884
• Remove empty submodule leftover. by @xkszltl in #1883
• Fix build for lack of RTLD_DI_ORIGIN support by @jaeyoonjung in #1888
• Enable graph capture for webgpu by @qjia7 in #1848
• Generic shared emb_tokens/lm_head implementation by @jixiongdeng in #1885
• Fix bug in Squeeze for getting the value of total_seq_len by @Honry in #1886
• Extra_options disable_qkv_fusion to untie qkv_projs from upstream choice by @jixiongdeng in #1893
• Fix mac pipeline by @apsonawane in #1904
• whisper: Support a variant of the whisper pipeline where encoder / decoder are stateful. by @RyanMetcalfeInt8 in #1857
• Add model builder for Qwen2_5_VLTextModel by @tianleiwu in #1882
• Integrate FARA-7B model by @apsonawane in #1902
• Set version as 0.11.3 by @kunal-vaishnavi in #1905

New Contributors

• @xkszltl made their first contribution in #1883
• @jaeyoonjung made their first contribution in #1888
• @jixiongdeng made their first contribution in #1885
• @Honry made their first contribution in #1886

Full Changelog: v0.11.2...v0.11.3

What's Changed

• Revert removal of eps_without_if_support by @xiaofeihan1 in #1878
• Fix condition for NPU by @apsonawane in #1880
• Set version as 0.11.2 by @kunal-vaishnavi in #1881

Full Changelog: v0.11.1...v0.11.2