pytorch/executorch

🐛 Describe the bug Summary In exported LFM2 and LFM2.5 models, ShortConv.conv_state (a mutable buffer holding the last two conv inputs) is never cleared when a new sequence starts at input_pos = 0. After TextLLMRunner.reset() (Python), LlmModule.resetContext() (Android) or TextLLMRunner::warmup(), the next conversation starts from the previous conversation's convolution state. Continuing a conversation at the current position is not affected. Expected: after reset(), a prompt gives the same output as on a freshly loaded model. Actual: the output depends on what ran before the reset. To reproduce Environment: executorch==1.5.1 and torch==2.14.0 from pip. Model: a public export, lfm_2_5_350m_xnnpack_8da4w.pte (sha256 e8c38e0d933932b02dddbcf92df813045899d0f71bf5346dbceecf5c4078f05d), and its tokenizer.json (sha256 2221c71b5dce048a8abae62843b92bd7deec13cc153f95fa6e3327a47b79a7da). 1. Module level (python module_repro.py lfm_2_5_350m_xnnpack_8da4w.pte) import sys import torch from exec

Enable the native Kev example on the WebGPU backend, extending the XNNPACK and MLX support added in #23023. Kev runs prefill followed by pointer-head scoring, with an immutable prefix snapshot reused across question batches. Start with the existing C++ Module API and a native Dawn build, without Python at inference. Use the WebGPU backend guide and operator tests as starting points. Qwen3.5 MoE's GatedDeltaNet implementation provides a reference for GDN math and state layouts. Coordinate shared export/lowering work with the Vulkan issue, #23098. Work to cover: Audit Kev's prefill and score graphs against WebGPU runtime support. Add missing WGSL kernels and lowering support, including GDN, with focused backend regression tests. Export success alone is insufficient: WebGPUPartitioner shares Vulkan's frontend, whose operator coverage is broader. Add a WebGPU export option and CMake linkage for the existing kev_runner and kev_benchmark. Respect the current mutually exclusive Vulkan/WebGPU

Enable the native Kev example on the Vulkan backend, extending the XNNPACK and MLX support added in #23023. Kev runs prefill followed by pointer-head scoring, with an immutable prefix snapshot reused across question batches. The goal is to support this workflow through the existing C++ API on Vulkan-capable GPUs, without Python at inference. Start with Kev's model/export code, the Vulkan export example, and the Vulkan backend. Qwen3.5 MoE's GatedDeltaNet implementation provides a reference for GDN math and state layouts; Vulkan will need its own lowering or shader support where coverage is missing. Work to cover: Identify operator and dynamic-shape gaps in Kev's prefill and score graphs. Add the required Vulkan support, including GDN, with focused backend regression tests. A small GDN parity test is a useful first step. Add a Vulkan export option and CMake linkage so the existing kev_runner and kev_benchmark run the exported model through Module. Preserve system_one, explicit prefill/

Enable the native Kev example on NVIDIA GPUs through ExecuTorch's AOTI CUDA backend. The example landed in #23023 with XNNPACK and MLX support. Kev runs prefill followed by a pointer head over option-boundary hidden states. A prefix snapshot is reused across question batches. The goal is to run this same workflow through the existing C++ API, without Python at inference. Use Qwen3.5 MoE as the primary reference for CUDA export, native loading, and Gated DeltaNet (GDN). Its chunked GDN kernel and backend tests are useful starting points. Work to cover: Add a CUDA export option for Kev's prefill and score methods using CudaPartitioner. Reuse the existing GDN prefill kernel, adapting layouts, gate/scaling conventions, and explicit initial/final state to Kev. Wire up aoti_cuda_backend in the example's CMake build and load the .pte plus CUDA delegate .ptd data through Module. Handle device memory and synchronization when retaining prefix snapshots and reading scores on the host. Preserve s

Problem The Apple and pybind presets use xcrun -sdk macosx --find metal. On current Xcode releases this can locate a launcher even when the separately installed Metal Toolchain component is unavailable, so configuration enables MLX and fails later during compilation. The documentation correction is part of #22620. The preset behavior change is intentionally excluded from that documentation-only PR. Proposed change Probe xcrun -sdk macosx metal --version in both presets and update the fallback message to point to xcodebuild -downloadComponent MetalToolchain. Test plan Configure on Apple silicon with Xcode and the Metal Toolchain installed; verify MLX remains enabled. Configure with Xcode but without the Metal Toolchain; verify MLX is disabled cleanly. Run Apple pybind and MLX build CI. cc @larryliu0820 @GregoryComer @metascroy