Issues

• Limit per-session prompt-end retention in the shared prefix cache

  pytorch/executorch#24914 · Filter by author mergennachin opened on Nov 11, 2026

• [ET-VK] BUFFER softmax and layer_norm abort on oversized tiled dispatch

  pytorch/executorch#24906 · Filter by author flyingdoog opened on Nov 11, 2026

• CUDA backend: aten._grouped_mm cannot be lowered on A100 (sm_80), so MoE models do not export there

  module: cuda
  pytorch/executorch#24902 · Filter by author shoumikhin opened on Nov 11, 2026
  

• CUDA backend: the default export spends more time on host copies than on the model (MoE layer: 2.3 ms of kernels, 4 ms of copies)

  module: cuda
  pytorch/executorch#24901 · Filter by author shoumikhin opened on Nov 11, 2026
  

• CUDA backend: a program crashes with an illegal memory access when CUDA device 1 is current

  bugmodule: cuda
  pytorch/executorch#24887 · Filter by author shoumikhin opened on Nov 11, 2026

• [QNN] HTP fp16 log_softmax returns shifted values when the softmax axis length is 2^n + 1 (65 … 2049)

  pytorch/executorch#24886 · Filter by author ujma1234 opened on Nov 11, 2026

• samsung device farm stopped assigning devices, test-samsung-models-linux times out on every trunk run

  module: cimodule: samsungpartner: samsung
  pytorch/executorch#24869 · Filter by author shoumikhin opened on Nov 10, 2026
  4

• [RFC] A runtime-only ExecuTorch Python package: executorch-runtime runs a .pte without PyTorch

  pytorch/executorch#24820 · Filter by author shoumikhin opened on Nov 10, 2026
  2

• [RFC] ExecuTorch for ROS 2: an installable runtime package and a general inference node

  pytorch/executorch#24819 · Filter by author shoumikhin opened on Nov 10, 2026
  1

• [RFC] Ship the ExecuTorch C++ SDK as a standalone archive with each release, starting with 1.6.0

  pytorch/executorch#24816 · Filter by author shoumikhin opened on Nov 9, 2026

• Slow small-batch quantized inference on M1/M2 GPUs

  pytorch/executorch#24812 · Filter by author mergennachin opened on Nov 9, 2026

• Speculative decoding loses speed to repeated per-round work

  pytorch/executorch#24811 · Filter by author mergennachin opened on Nov 9, 2026

• CUDA backend: no fallback kernels for nonzero and masked_select

  pytorch/executorch#24767 · Filter by author shoumikhin opened on Nov 7, 2026
  1

  

• XNNPACK packs a reused weight once per use, so memory grows with every reuse

  pytorch/executorch#24766 · Filter by author shoumikhin opened on Nov 7, 2026

• No safe way to set the CPU thread count from Python: _unsafe_reset_threadpool after load_method crashes or hangs XNNPACK

  bug
  pytorch/executorch#24765 · Filter by author shoumikhin opened on Nov 7, 2026

• torch.masked_select fails to_edge ("not in Core ATen opset"), although the portable kernel exists

  bug
  pytorch/executorch#24764 · Filter by author shoumikhin opened on Nov 7, 2026
  1

• torch.nonzero returns wrong indices, and indexing with its result fails at run time

  bug
  pytorch/executorch#24763 · Filter by author shoumikhin opened on Nov 7, 2026

• Qualcomm AI Engine Direct: HTP float16 Concat reads a static bool input of all True as False

  module: qnnpartner: qualcomm
  pytorch/executorch#24750 · Filter by author shoumikhin opened on Nov 6, 2026

• Qualcomm AI Engine Direct: a delegated mutable buffer starts from leftover memory

  module: qnnpartner: qualcomm
  pytorch/executorch#24749 · Filter by author shoumikhin opened on Nov 6, 2026

• CUDA AOTI:reinterpret_tensor drops the source tensor's storage offset

  bugmodule: cudatriaged
  Bug

  pytorch/executorch#24737 · Filter by author mshr-h opened on Nov 6, 2026
  4

  

• How can I measure the full memory usage of an ExecuTorch QNN HTP model?

  module: qnnpartner: qualcomm
  Task

  pytorch/executorch#24726 · Filter by author baoliangEss opened on Nov 6, 2026
  1

  

• Arm: how much of each LeRobot policy the Arm backend (VGF, Ethos-U85) takes, and what blocks the rest

  module: arm
  pytorch/executorch#24697 · Filter by author shoumikhin opened on Nov 5, 2026

• Vulkan: how much of each LeRobot policy the Vulkan backend takes, and what blocks the rest

  module: vulkan
  pytorch/executorch#24688 · Filter by author shoumikhin opened on Nov 5, 2026
  

• CUDA: LeRobot policies performance tracker, and what blocks them

  module: cuda
  pytorch/executorch#24684 · Filter by author shoumikhin opened on Nov 5, 2026
  1

  

• CUDA AOTI: Inductor memory planning fails with undefined symbol aoti_torch__alloc_from_pool

  bugmodule: cudatriaged
  Bug

  pytorch/executorch#24682 · Filter by author mshr-h opened on Nov 5, 2026
  2