Fix crashes, wrong outputs and lowering failures in several backends and in exir by shoumikhin · Pull Request #24222 · pytorch/executorch

Catch missing scatter and index_put kernels when lowering to AOTI backends

shoumikhin

committed

eebd74a

Keep the OpenVINO blob alive while its compiled model is in use

shoumikhin

committed

58ebe32

Qualcomm AI Engine Direct - Fix DecomposeWrapWithAutocast for real regions

shoumikhin

committed

b0f7f53

Lower detach in AOTI backends

shoumikhin

committed

44991dd

Copy a strided CUDA delegate output into its ETensor

shoumikhin

committed

80aa633

Do not fold a tensor that an in-place op writes in constant_prop_pass

shoumikhin

committed

7e5178f

[MLX] Check tensor list inputs for unsupported dtypes

shoumikhin

committed

d48cbd1

[MLX] Share one constant slot between placeholders of the same weight

shoumikhin

committed

d1a5fc5

Skip CoreML ops on bfloat16 tensors

shoumikhin

committed

5e7f3e8

Skip CoreML ops on tensors above rank 5

shoumikhin

committed

87a0991

Vulkan: load programs with a constant of more than 4 dims

shoumikhin

committed

34059c9

Arm backend: Keep unused delegate inputs in the TOSA graph

shoumikhin

committed

f869d87

NXP backend: quantize programs from a non-strict torch.export

shoumikhin

committed

2bdf12d

Do not quantize bfloat16 tensors in XNNPACKQuantizer

shoumikhin

committed

31a6348

Keep the output node last when to_backend fuses a partition

shoumikhin

committed

5dc180d

Lower torch._check_tensor_all and torch.sym_ite through to_edge

shoumikhin

committed

401a564

Add the aoti_torch_as_strided shim to the CUDA backend

shoumikhin

committed

1078f5a

Stop GroupBasedPartitioner from making partitions that depend on each other

shoumikhin

committed

8363aae

Cortex-M: Handle a constant operand in MatmulToBmmPass

shoumikhin

committed

3dc94ab

exir: Give tensor constants static shapes in ExportPass

shoumikhin

committed

5350e76

[MLX] Read the base of an out-of-place slice update as an input

shoumikhin

committed

43849be

Qualcomm AI Engine Direct - Decompose a stack that HTP Pack cannot take

shoumikhin

committed

41052c8

Qualcomm AI Engine Direct - Keep partitions with no runtime input on CPU

shoumikhin

committed

f1b70d3

Qualcomm AI Engine Direct - Remove a 0-d select only when its input can replace it

shoumikhin

committed

c112a14

Qualcomm AI Engine Direct - Recompose RMSNorm only over the last dim with its weight

shoumikhin

committed

27993e9

Qualcomm AI Engine Direct - Leave bfloat16 nodes on CPU

shoumikhin

committed

2db708f

Qualcomm AI Engine Direct - Leave nodes with data-dependent shapes on CPU

shoumikhin

committed

cb58a4b

Qualcomm AI Engine Direct - Cast a top-of-graph int64 constant before its user

shoumikhin

committed

593b908

Delegate ELU to XNNPACK

shoumikhin

committed

c9c90f6

Stop XNNPACK from dropping the scales of weight-only quantized linears

shoumikhin

committed

3ee64c6

Add a portable kernel for cummin.out

shoumikhin

committed

91894de

OpenVINO backend: convert partitions with their input types

shoumikhin

committed

4ec7b9f

CUDA backend: lower transposed convolution without ATen

shoumikhin

committed

d0cf5b0

Navigation Menu

Repository navigation

Fix crashes, wrong outputs and lowering failures in several backends and in exir - #24222

Fix crashes, wrong outputs and lowering failures in several backends and in exir#24222
shoumikhin wants to merge 33 commits into
pytorch:mainfrom
shoumikhin:fix-lerobot-export-bugs

Commits on Oct 29, 2026

Catch missing scatter and index_put kernels when lowering to AOTI backends

Keep the OpenVINO blob alive while its compiled model is in use

Qualcomm AI Engine Direct - Fix DecomposeWrapWithAutocast for real regions

Lower detach in AOTI backends

Copy a strided CUDA delegate output into its ETensor

Do not fold a tensor that an in-place op writes in constant_prop_pass

[MLX] Check tensor list inputs for unsupported dtypes

[MLX] Share one constant slot between placeholders of the same weight

Skip CoreML ops on bfloat16 tensors

Skip CoreML ops on tensors above rank 5

Vulkan: load programs with a constant of more than 4 dims

Arm backend: Keep unused delegate inputs in the TOSA graph

NXP backend: quantize programs from a non-strict torch.export

Do not quantize bfloat16 tensors in XNNPACKQuantizer

Keep the output node last when to_backend fuses a partition

Lower torch._check_tensor_all and torch.sym_ite through to_edge

Add the aoti_torch_as_strided shim to the CUDA backend

Stop GroupBasedPartitioner from making partitions that depend on each other

Cortex-M: Handle a constant operand in MatmulToBmmPass

exir: Give tensor constants static shapes in ExportPass

[MLX] Read the base of an out-of-place slice update as an input

Qualcomm AI Engine Direct - Decompose a stack that HTP Pack cannot take

Qualcomm AI Engine Direct - Keep partitions with no runtime input on CPU

Qualcomm AI Engine Direct - Remove a 0-d select only when its input can replace it

Qualcomm AI Engine Direct - Recompose RMSNorm only over the last dim with its weight

Qualcomm AI Engine Direct - Leave bfloat16 nodes on CPU

Qualcomm AI Engine Direct - Leave nodes with data-dependent shapes on CPU

Qualcomm AI Engine Direct - Cast a top-of-graph int64 constant before its user

Delegate ELU to XNNPACK

Stop XNNPACK from dropping the scales of weight-only quantized linears

Add a portable kernel for cummin.out

OpenVINO backend: convert partitions with their input types

CUDA backend: lower transposed convolution without ATen