Issues

New issue

Label tracking meta-issue (edit me to get automatically CC'ed on issues! cc bot)
#7679 · mergennachin opened on Jan 15, 2025
[v1.5.0] Release Schedule and Tracker
#22188 · JacobSzwejbka opened on Aug 26, 2026
5

PTE and PTD serialization drop stricter alignment requests for deduplicated named data
pytorch/executorch#23302 · opened on Oct 1, 2026
LFM2/LFM2.5: short-conv state is not cleared when a new sequence starts, so output after reset() depends on the previous prompt
pytorch/executorch#23262 · opened on Sep 29, 2026
XNNPACK never delegates slice_copy with stride != 1, but op-support.csv only excludes zero-dim/dynamic shapes
pytorch/executorch#23258 · opened on Sep 29, 2026
[RFC] ExecuTorch Persisting Device Specialized Delegate Artifacts
pytorch/executorch#23192 · opened on Sep 28, 2026
6
[QNN] Enable ConvTranspose + BatchNorm fusion after #23170
pytorch/executorch#23185 · opened on Sep 28, 2026
1
[Tracking] Complete audio task coverage across CPU, GPU, and NPU backends
pytorch/executorch#23164 · opened on Sep 25, 2026
7
Qualcomm: 8-bit per-channel weight scales are floored at the 16-bit eps, and the HTP miscomputes near-zero channels
Task
pytorch/executorch#23160 · opened on Sep 25, 2026
1
[cpu kernels] native_layer_norm: layer_norm_scalar returns NaN on large-mean rows; Half/BF16 at N>=256 slow after #23153
pytorch/executorch#23159 · opened on Sep 25, 2026
2
[Vulkan] scalar_tensor uploads float bytes to integer scalar shaders
pytorch/executorch#23158 · opened on Sep 25, 2026
[Vulkan] Missing/partial op support blocks full delegation of a dynamic-shape transformer encoder
pytorch/executorch#23156 · opened on Sep 25, 2026
Enable Nemotron-3-Diarization on MLX, CUDA, XNNPACK, and Vulkan
pytorch/executorch#23131 · opened on Sep 24, 2026
2
Qualcomm: 16a16w quantization is much less accurate than 16a8w
Bug
pytorch/executorch#23108 · opened on Sep 24, 2026
2
Enable native Kev inference with the WebGPU backend
pytorch/executorch#23099 · opened on Sep 23, 2026
2
Enable native Kev inference with the Vulkan backend
pytorch/executorch#23098 · opened on Sep 23, 2026
2
Enable native Kev inference with the AOTI CUDA backend
pytorch/executorch#23097 · opened on Sep 23, 2026
2
Qualcomm: HTP hangs in fp16 on a dilated conv when the dilated extent exceeds the input and cin is large
Bug
pytorch/executorch#23096 · opened on Sep 23, 2026
3
Add BF16 convolution support to the XNNPACK backend
pytorch/executorch#23058 · opened on Sep 23, 2026
Track Muse Glimmer DFlash2 support
pytorch/executorch#23011 · opened on Sep 22, 2026
[Vulkan] SDPA buffer IO uses the wrong stride when head_dim is not divisible by 4
pytorch/executorch#23001 · opened on Sep 22, 2026
Add a decision-model ("System One") example, aka Jev-like model
pytorch/executorch#22985 · opened on Sep 22, 2026
1
Metal backend: generated kernels index a packed copy of a non-packed view out of bounds (wrong results, crash in MPSGraph)
Bug
pytorch/executorch#22983 · opened on Sep 22, 2026
RFC: run an ExecuTorch model from Python without PyTorch
pytorch/executorch#22971 · opened on Sep 21, 2026
4
[Vulkan] LLM decode ~12% slower than XNNPACK on Adreno
Task
pytorch/executorch#22968 · opened on Sep 21, 2026
13
Metal backend: ops read a view with a storage offset before the GPU has written it (silently wrong results)
Bug
pytorch/executorch#22956 · opened on Sep 20, 2026
2
Metal backend: exported .pte does not load outside a process that already has PyTorch's libomp (embedded AOTI object links /opt/llvm-openmp/lib/libomp.dylib)
Task
pytorch/executorch#22955 · opened on Sep 20, 2026
1

Navigation Menu

Issues

Issues

PTE and PTD serialization drop stricter alignment requests for deduplicated named data

LFM2/LFM2.5: short-conv state is not cleared when a new sequence starts, so output after reset() depends on the previous prompt

XNNPACK never delegates slice_copy with stride != 1, but op-support.csv only excludes zero-dim/dynamic shapes

[RFC] ExecuTorch Persisting Device Specialized Delegate Artifacts

[QNN] Enable ConvTranspose + BatchNorm fusion after #23170

[Tracking] Complete audio task coverage across CPU, GPU, and NPU backends

Qualcomm: 8-bit per-channel weight scales are floored at the 16-bit eps, and the HTP miscomputes near-zero channels

[cpu kernels] native_layer_norm: layer_norm_scalar returns NaN on large-mean rows; Half/BF16 at N>=256 slow after #23153

[Vulkan] scalar_tensor uploads float bytes to integer scalar shaders

[Vulkan] Missing/partial op support blocks full delegation of a dynamic-shape transformer encoder

Enable Nemotron-3-Diarization on MLX, CUDA, XNNPACK, and Vulkan

Qualcomm: 16a16w quantization is much less accurate than 16a8w

Enable native Kev inference with the WebGPU backend

Enable native Kev inference with the Vulkan backend

Enable native Kev inference with the AOTI CUDA backend

Qualcomm: HTP hangs in fp16 on a dilated conv when the dilated extent exceeds the input and cin is large

Add BF16 convolution support to the XNNPACK backend

Track Muse Glimmer DFlash2 support

[Vulkan] SDPA buffer IO uses the wrong stride when head_dim is not divisible by 4

Add a decision-model ("System One") example, aka Jev-like model

Metal backend: generated kernels index a packed copy of a non-packed view out of bounds (wrong results, crash in MPSGraph)

RFC: run an ExecuTorch model from Python without PyTorch

[Vulkan] LLM decode ~12% slower than XNNPACK on Adreno

Metal backend: ops read a view with a storage offset before the GPU has written it (silently wrong results)

Metal backend: exported `.pte` does not load outside a process that already has PyTorch's libomp (embedded AOTI object links `/opt/llvm-openmp/lib/libomp.dylib`)