Bug fixes and improvements
Add opt-in synchronous GPU allocator (nv_use_sync_gpu_allocator) with a Technical Notes doc
Support iGPU zero-copy: pass resolved device pointers from host-accessible tensors directly to TRT context
Enable Vulkan resource import and CiG; add D3D12 external resource import, disable CPU fallback
Add compile-only mode support for ORT 1.27+ (EP ABI changes to reduce compile time)
Pass Phi-4 LongRoPE cache offset through the EP
Add IHV Weightless EP Context: generate EP context models without embedded weights
Add profiler support for MSFT integration
Add Windows-on-Arm binary and wheel build support in build.bat
Fix OOB write on high-rank inputs and use-after-free in EP destructor
Pin EP DLL in DllMain throughout ORT process lifetime to prevent premature unload
Fix path traversal via model-supplied name in runtime cache; remove legacy RefitEngineImpl fallback
Fix runtime cache teardown lifetime; use MinSupportedOrtApiVersion for ORT version comparison
Contributors to this release of TensorRT RTX EP ABI:
@keshavv27, @gedoensmax, @ishwar-raut1, @umangb-09, @yen-shi, @praneshgo, @wenbingl, @mklimenko-nv, @hthadicherla, @sandeep-nv
