mobilenetv2 doesn't work with Vulkan backend

[sternezsl]

🐛 Describe the bug

I can successfully export the vulkan pte. When I run the model with

./backends/vulkan/vulkan_executor_runner --model_path /scratch/models/vulkan_mobilenetv2.pte

I get the error:

I 00:00:00.001717 executorch:executor_runner.cpp:82] Model file /scratch/models/vulkan_mobilenetv2.pte is loaded.
I 00:00:00.001730 executorch:executor_runner.cpp:91] Using method forward
I 00:00:00.001732 executorch:executor_runner.cpp:138] Setting up planned buffer 0, size 606112.
libc++abi: terminating due to uncaught exception of type vkcompute::vkapi::Error: Exception raised from check_conv_args at /scratch/code/meta/executorch/backends/vulkan/runtime/graph/ops/impl/Convolution.cpp:225: (check_packed_dim_is(in, WHCN::kChannelsDim)) is false!
[1] 544559 IOT instruction (core dumped) ./backends/vulkan/vulkan_executor_runner --model_path

I inspect the code and find that the input tensor's packed_dim is 0 rather than 2(WHCN::kChannelsDim). If I comment out the check_conv_args function, then I run into another problem:

I 00:00:00.000351 executorch:executor_runner.cpp:82] Model file /scratch/models/vulkan_mobilenetv2.pte is loaded.
I 00:00:00.000358 executorch:executor_runner.cpp:91] Using method forward
I 00:00:00.000360 executorch:executor_runner.cpp:138] Setting up planned buffer 0, size 606112.
libc++abi: terminating due to uncaught exception of type vkcompute::vkapi::Error: Exception raised from mean at /scratch/code/meta/executorch/backends/vulkan/runtime/graph/ops/impl/Reduce.cpp:112: (dims_list->size() == 1) is false!

I noticed that @SS-JIA replaced the mobilenet demo with a Add one in the tutorial a few days ago. I guess you know the problem. I try to fix the problem. unfortunately, presently I do not know much about Vulkan backend, could you please give me some hints and I'll try to fix it.

The following is the model conversion script:

import torch
import torchvision.models as models

from torch.export import export, ExportedProgram
from torchvision.models.mobilenetv2 import MobileNet_V2_Weights
from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
from executorch.exir import EdgeProgramManager, ExecutorchProgramManager, to_edge_transform_and_lower
from executorch.exir.backend.backend_api import to_backend
from executorch.backends.vulkan import VulkanPartitioner

mobilenet_v2 = models.mobilenetv2.mobilenet_v2(weights=MobileNet_V2_Weights.DEFAULT).eval()
sample_inputs = (torch.randn(1, 3, 224, 224), )

exported_program: ExportedProgram = export(mobilenet_v2, sample_inputs)
edge: EdgeProgramManager = to_edge_transform_and_lower(
    exported_program,
    partitioner=[VulkanPartitioner()],
)

# print(edge.exported_program().graph_module)

exec_prog = edge.to_executorch()

with open("vulkan_mobilenetv2.pte", "wb") as file:
    exec_prog.write_to_file(file)

Versions

PyTorch version: 2.5.0+git8a0ce38
Is debug build: False
CUDA used to build PyTorch: 11.8
ROCM used to build PyTorch: N/A
VULKAN used to build PyTorch: True

OS: Fedora Linux Asahi Remix 40 (KDE Plasma) (aarch64)
GCC version: (GCC) 14.2.1 20240912 (Red Hat 14.2.1-3)
Clang version: 18.1.7 (https://github.com/llvm/llvm-project 768118d1ad38bf13c545828f67bd6b474d61fc55)
CMake version: version 3.20.0
Libc version: glibc-2.39
Vulkan Driver Version: Mesa 24.3.0-devel (git-f05157f591),
Vulkan Instance Version: 1.3.290

Python version: 3.10.10 | packaged by conda-forge | (main, Mar 24 2023, 19:56:21) [GCC 11.3.0] (64-bit runtime)
Python platform: Linux-6.11.3-dimilar-4-1-edge-ARCH+-aarch64-with-glibc2.39
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU: Apple M1 Max

Versions of relevant libraries:
[pip3] flake8==6.1.0
[pip3] flake8-breakpoint==1.1.0
[pip3] flake8-bugbear==23.9.16
[pip3] flake8-comprehensions==3.14.0
[pip3] flake8-plugin-utils==1.3.3
[pip3] flake8-pyi==23.5.0
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.25.0
[pip3] pytorch-sphinx-theme==0.0.19
[pip3] torchao==0.7.0+git6b529961
[pip3] torchvision==0.20.0a0+f851df1
[conda] numpy 1.25.0 pypi_0 pypi
[conda] pytorch-sphinx-theme 0.0.19 pypi_0 pypi
[conda] torchao 0.7.0+git6b529961 pypi_0 pypi
[conda] torchfix 0.5.0 pypi_0 pypi
[conda] torchvision 0.20.0a0+f851df1 pypi_0 pypi

cc @SS-JIA @manuelcandales

[sternezsl]

• The first error is due to the memory layouts of input and output tensor are not TENSOR_CHANNELS_PACKED
• About the second error, I don't think the reduce implementation of sum.dim_IntList is the alternative of the old one. I restore the old Sum.cpp and the staff of sum_dim*.glsl and sum_dim*.yaml, then the mobilenetv2 works now and I get the same result as same as the xnnpack backend

[SS-JIA]

Hi @sternezsl, thanks for raising this issue!

I noticed that @SS-JIA replaced the mobilenet demo with a Add one in the tutorial a few days ago. I guess you know the problem. I try to fix the problem. unfortunately, presently I do not know much about Vulkan backend, could you please give me some hints and I'll try to fix it.

I definitely appreciate the help on this :) I did encounter an issue when trying out the mobilenet example, but it was separate from the issues that you have described. The problem that I encountered back then was related to the runner binary, not the Vulkan export, and I decided to replace the example with a simpler one because I didn't have time to look into the failure before the beta release of ExecuTorch. The fact that you were able to get MobileNet V2 working suggests that the problem with the runner binary has been fixed, however. If we can verify that mobilenet works on main then it would make sense to make the example mobilenet once more (y)

About the second error, I don't think the reduce implementation of sum.dim_IntList is the alternative of the old one. I restore the old Sum.cpp and the staff of sum_dim*.glsl and sum_dim*.yaml, then the mobilenetv2 works now and I get the same result as same as the xnnpack backend

I also became aware of this regression while trying to commit the new reduce shader, but decided to go ahead with it for the sake of updating the implementation to use a more modern pattern of being agnostic to memory layouts, and because the new implementation is more performant for the single reduction dim case. Also, at the time mobilenet v2 was the only model that I was working with at the time that used sum with multiple reduction dimensions so I figured the impact of the regression was not super significant.

When I first committed the new reduce shader, there was an issue with the regression since the new implementation cannot handle reducing over multiple tensor dimensions. However, I have since landed #6488 as well which addresses this. The linked PR essentially adds the ability for operators to implement a specific check function to detect whether they are supported by Vulkan, and after the PR was committed the sum op in mobilenet will no longer be delegated to Vulkan.

Please let me know if you're still running into problems on the current main branch.

The first error is due to the memory layouts of input and output tensor are not TENSOR_CHANNELS_PACKED

I am currently working on #6636 which will fix errors like these. However, in the meantime if you lower to Vulkan with the following settings:

VulkanPartitioner(
    compile_options={
        "memory_layout_override": vulkan_graph_schema.VkMemoryLayout.TENSOR_CHANNELS_PACKED,
    }

It should fix the issue as well.

[alexggener]

hi @SS-JIA @sternezsl

Just tested mobilenetv2 network on vulkan delegate and worked perfectly. See below output of my test:

(executorch_env) jetson@ubuntu:~/alex/executorch$ ./cmake-out/backends/vulkan/vulkan_executor_runner --model_path ../example_files/vulkan_mobilenetv2.pte 
I 00:00:00.008907 executorch:executor_runner.cpp:82] Model file ../example_files/vulkan_mobilenetv2.pte is loaded.
I 00:00:00.008956 executorch:executor_runner.cpp:91] Using method forward
I 00:00:00.008974 executorch:executor_runner.cpp:138] Setting up planned buffer 0, size 852992.
I 00:00:01.341410 executorch:executor_runner.cpp:161] Method loaded.
I 00:00:01.343461 executorch:executor_runner.cpp:171] Inputs prepared.
I 00:00:01.377504 executorch:executor_runner.cpp:180] Model executed successfully.
I 00:00:01.377541 executorch:executor_runner.cpp:184] 1 outputs: 
Output 0: tensor(sizes=[1, 1000], [
  -0.509862, 0.300651, 0.0953805, 0.147718, 0.231189, 0.338554, 0.206877, -0.0575791, -0.389279, -0.0606882, 
  -0.0213991, -0.12104, -0.288959, 0.13405, -0.171975, -0.0603708, 0.0203585, -0.058535, 0.337849, -0.0718671, 
  0.490757, 0.524136, 0.197857, 0.122061, -0.359134, 0.109462, 0.347758, 0.478523, 0.226562, 0.03636, 
  0.0159195, 0.351976, 0.259124, -0.0542909, 0.285072, -0.221404, 0.237161, -0.378553, 0.395111, -0.0668834, 
  0.357152, 0.400402, 0.389977, -0.189019, 0.243561, -0.103937, 0.592332, 0.00744483, -0.183821, -0.446262, 
  -0.182813, -0.679569, 0.663809, 0.560708, 0.362915, -0.0855762, 0.142371, 0.172888, 0.593107, 0.305179, 
  0.44763, -0.138471, -0.149111, 0.063244, -0.12325, 0.511517, 0.519203, 0.392336, 0.731634, 0.765339, 
  0.460784, 0.611435, -0.209273, 0.328244, -0.142377, 0.699485, 0.0476283, 0.562083, 1.51458, 0.825767, 
  0.126672, 0.0498361, -0.0896511, -0.142821, -0.0252667, 0.00359163, 0.081919, -0.214226, 0.0404544, 0.105451, 
  -0.268513, -0.0829375, 0.331347, -0.345984, -0.134049, -0.291834, -0.118031, -0.102933, 0.158994, -0.0496359, 
  ...,
  -0.188106, -0.694425, 0.22042, -0.0921049, 0.761134, 0.212517, 0.0171885, 0.461983, 0.685643, -0.122819, 
  0.352455, 2.10311, 0.211244, 0.0732195, -0.366484, -0.500692, -0.00569095, -0.186634, 0.256008, 0.101067, 
  -0.112603, 0.0634007, 0.519891, -0.543176, -0.223353, 0.155177, -0.2306, -0.180301, -0.402723, -0.102213, 
  0.331341, -0.0324454, 0.428082, -0.253921, -0.192847, -0.207007, 0.521829, 0.121384, 0.284403, -0.160636, 
  0.0179912, 0.290297, 0.32837, 0.154164, 0.193875, 0.287706, -0.0284017, -0.119623, 0.955584, 0.581988, 
  0.808401, 0.669408, 0.272974, 0.161549, 0.379885, 0.212438, -0.325234, 0.100537, 0.292663, -0.382237, 
  -0.389108, 0.447187, -0.124377, 0.214354, 0.592613, -0.367157, 0.191232, 0.423556, 0.349301, 0.0348431, 
  -0.227161, 0.567002, 0.2029, 0.710087, 0.421643, -0.00655486, 0.0114809, 0.398904, 0.0349747, -0.163218, 
  0.187845, -0.15439, -0.227166, 0.150875, 0.265098, 0.0874977, -0.188219, 0.021303, -0.029385, -0.279626, 
  0.421217, 0.100453, -0.506771, -0.115819, -0.693013, -0.183252, 0.154796, -0.410685, 0.011931, 0.449709, 
])

My environment:

PyTorch version: 2.6.0.dev20241111+cpu
Is debug build: False
CUDA used to build PyTorch: Could not collect
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.5 LTS (aarch64)
GCC version: (Ubuntu 13.1.0-8ubuntu1~22.04) 13.1.0
Clang version: Could not collect
CMake version: version 3.31.2
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov  6 2024, 20:22:13) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.148-tegra-aarch64-with-glibc2.35
Is CUDA available: False
CUDA runtime version: 12.2.140
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: GPU 0: Orin (nvgpu)
Nvidia driver version: 540.4.0
cuDNN version: Probably one of the following:
/usr/lib/aarch64-linux-gnu/libcudnn.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_adv.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_adv_infer.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn_adv_train.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn_cnn.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_cnn_infer.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn_cnn_train.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn_engines_precompiled.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_graph.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_heuristic.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_ops.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_ops_infer.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn_ops_train.so.8.9.4
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                       aarch64
CPU op-mode(s):                     32-bit, 64-bit
Byte Order:                         Little Endian
CPU(s):                             8
On-line CPU(s) list:                0-3
Off-line CPU(s) list:               4-7
Vendor ID:                          ARM
Model name:                         Cortex-A78AE
Model:                              1
Thread(s) per core:                 1
Core(s) per cluster:                4
Socket(s):                          -
Cluster(s):                         1
Stepping:                           r0p1
CPU max MHz:                        1984.0000
CPU min MHz:                        115.2000
BogoMIPS:                           62.50
Flags:                              fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm lrcpc dcpop asimddp uscat ilrcpc flagm paca pacg
L1d cache:                          256 KiB (4 instances)
L1i cache:                          256 KiB (4 instances)
L2 cache:                           1 MiB (4 instances)
L3 cache:                           2 MiB (1 instance)
NUMA node(s):                       1
NUMA node0 CPU(s):                  0-3
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit:        Not affected
Vulnerability L1tf:                 Not affected
Vulnerability Mds:                  Not affected
Vulnerability Meltdown:             Not affected
Vulnerability Mmio stale data:      Not affected
Vulnerability Retbleed:             Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass:    Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:           Mitigation; __user pointer sanitization
Vulnerability Spectre v2:           Mitigation; CSV2, but not BHB
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected

Versions of relevant libraries:
[pip3] executorch==0.5.0a0+2ed5ce3
[pip3] numpy==1.26.4
[pip3] torch==2.6.0.dev20241111+cpu
[pip3] torchao==0.8.0+git2e032c6b
[pip3] torchaudio==2.5.0.dev20241111
[pip3] torchsr==1.0.4
[pip3] torchvision==0.20.0.dev20241112
[pip3] triton==3.1.0
[conda] Could not collect

Hope this helps with the mobilenetv2 example!

Alex.

[sternezsl]

hi @SS-JIA @sternezsl

Just tested mobilenetv2 network on vulkan delegate and worked perfectly. See below output of my test:

(executorch_env) jetson@ubuntu:~/alex/executorch$ ./cmake-out/backends/vulkan/vulkan_executor_runner --model_path ../example_files/vulkan_mobilenetv2.pte 
I 00:00:00.008907 executorch:executor_runner.cpp:82] Model file ../example_files/vulkan_mobilenetv2.pte is loaded.
I 00:00:00.008956 executorch:executor_runner.cpp:91] Using method forward
I 00:00:00.008974 executorch:executor_runner.cpp:138] Setting up planned buffer 0, size 852992.
I 00:00:01.341410 executorch:executor_runner.cpp:161] Method loaded.
I 00:00:01.343461 executorch:executor_runner.cpp:171] Inputs prepared.
I 00:00:01.377504 executorch:executor_runner.cpp:180] Model executed successfully.
I 00:00:01.377541 executorch:executor_runner.cpp:184] 1 outputs: 
Output 0: tensor(sizes=[1, 1000], [
  -0.509862, 0.300651, 0.0953805, 0.147718, 0.231189, 0.338554, 0.206877, -0.0575791, -0.389279, -0.0606882, 
  -0.0213991, -0.12104, -0.288959, 0.13405, -0.171975, -0.0603708, 0.0203585, -0.058535, 0.337849, -0.0718671, 
  0.490757, 0.524136, 0.197857, 0.122061, -0.359134, 0.109462, 0.347758, 0.478523, 0.226562, 0.03636, 
  0.0159195, 0.351976, 0.259124, -0.0542909, 0.285072, -0.221404, 0.237161, -0.378553, 0.395111, -0.0668834, 
  0.357152, 0.400402, 0.389977, -0.189019, 0.243561, -0.103937, 0.592332, 0.00744483, -0.183821, -0.446262, 
  -0.182813, -0.679569, 0.663809, 0.560708, 0.362915, -0.0855762, 0.142371, 0.172888, 0.593107, 0.305179, 
  0.44763, -0.138471, -0.149111, 0.063244, -0.12325, 0.511517, 0.519203, 0.392336, 0.731634, 0.765339, 
  0.460784, 0.611435, -0.209273, 0.328244, -0.142377, 0.699485, 0.0476283, 0.562083, 1.51458, 0.825767, 
  0.126672, 0.0498361, -0.0896511, -0.142821, -0.0252667, 0.00359163, 0.081919, -0.214226, 0.0404544, 0.105451, 
  -0.268513, -0.0829375, 0.331347, -0.345984, -0.134049, -0.291834, -0.118031, -0.102933, 0.158994, -0.0496359, 
  ...,
  -0.188106, -0.694425, 0.22042, -0.0921049, 0.761134, 0.212517, 0.0171885, 0.461983, 0.685643, -0.122819, 
  0.352455, 2.10311, 0.211244, 0.0732195, -0.366484, -0.500692, -0.00569095, -0.186634, 0.256008, 0.101067, 
  -0.112603, 0.0634007, 0.519891, -0.543176, -0.223353, 0.155177, -0.2306, -0.180301, -0.402723, -0.102213, 
  0.331341, -0.0324454, 0.428082, -0.253921, -0.192847, -0.207007, 0.521829, 0.121384, 0.284403, -0.160636, 
  0.0179912, 0.290297, 0.32837, 0.154164, 0.193875, 0.287706, -0.0284017, -0.119623, 0.955584, 0.581988, 
  0.808401, 0.669408, 0.272974, 0.161549, 0.379885, 0.212438, -0.325234, 0.100537, 0.292663, -0.382237, 
  -0.389108, 0.447187, -0.124377, 0.214354, 0.592613, -0.367157, 0.191232, 0.423556, 0.349301, 0.0348431, 
  -0.227161, 0.567002, 0.2029, 0.710087, 0.421643, -0.00655486, 0.0114809, 0.398904, 0.0349747, -0.163218, 
  0.187845, -0.15439, -0.227166, 0.150875, 0.265098, 0.0874977, -0.188219, 0.021303, -0.029385, -0.279626, 
  0.421217, 0.100453, -0.506771, -0.115819, -0.693013, -0.183252, 0.154796, -0.410685, 0.011931, 0.449709, 
])

My environment:

PyTorch version: 2.6.0.dev20241111+cpu
Is debug build: False
CUDA used to build PyTorch: Could not collect
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.5 LTS (aarch64)
GCC version: (Ubuntu 13.1.0-8ubuntu1~22.04) 13.1.0
Clang version: Could not collect
CMake version: version 3.31.2
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov  6 2024, 20:22:13) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.148-tegra-aarch64-with-glibc2.35
Is CUDA available: False
CUDA runtime version: 12.2.140
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: GPU 0: Orin (nvgpu)
Nvidia driver version: 540.4.0
cuDNN version: Probably one of the following:
/usr/lib/aarch64-linux-gnu/libcudnn.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_adv.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_adv_infer.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn_adv_train.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn_cnn.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_cnn_infer.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn_cnn_train.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn_engines_precompiled.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_graph.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_heuristic.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_ops.so.9.3.0
/usr/lib/aarch64-linux-gnu/libcudnn_ops_infer.so.8.9.4
/usr/lib/aarch64-linux-gnu/libcudnn_ops_train.so.8.9.4
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                       aarch64
CPU op-mode(s):                     32-bit, 64-bit
Byte Order:                         Little Endian
CPU(s):                             8
On-line CPU(s) list:                0-3
Off-line CPU(s) list:               4-7
Vendor ID:                          ARM
Model name:                         Cortex-A78AE
Model:                              1
Thread(s) per core:                 1
Core(s) per cluster:                4
Socket(s):                          -
Cluster(s):                         1
Stepping:                           r0p1
CPU max MHz:                        1984.0000
CPU min MHz:                        115.2000
BogoMIPS:                           62.50
Flags:                              fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm lrcpc dcpop asimddp uscat ilrcpc flagm paca pacg
L1d cache:                          256 KiB (4 instances)
L1i cache:                          256 KiB (4 instances)
L2 cache:                           1 MiB (4 instances)
L3 cache:                           2 MiB (1 instance)
NUMA node(s):                       1
NUMA node0 CPU(s):                  0-3
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit:        Not affected
Vulnerability L1tf:                 Not affected
Vulnerability Mds:                  Not affected
Vulnerability Meltdown:             Not affected
Vulnerability Mmio stale data:      Not affected
Vulnerability Retbleed:             Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass:    Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:           Mitigation; __user pointer sanitization
Vulnerability Spectre v2:           Mitigation; CSV2, but not BHB
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected

Versions of relevant libraries:
[pip3] executorch==0.5.0a0+2ed5ce3
[pip3] numpy==1.26.4
[pip3] torch==2.6.0.dev20241111+cpu
[pip3] torchao==0.8.0+git2e032c6b
[pip3] torchaudio==2.5.0.dev20241111
[pip3] torchsr==1.0.4
[pip3] torchvision==0.20.0.dev20241112
[pip3] triton==3.1.0
[conda] Could not collect

Hope this helps with the mobilenetv2 example!

Alex.

The issue is that Metal does not allow linear tiling on 3D textures. ExecuTorch includes a check for this in

executorch/backends/vulkan/runtime/vk_api/Adapter.cpp

Line 191 in f8d67fc

bool test_linear_tiling_3d_image_support(VkDevice device) {

, but there vkCreateImage always returns true on Asahi Linux(MJesa hk vulkan driver). I checked both Mesa HoneyKrisp driver and MoltenVk driver, both hk_CreateImage and MVKDevice::createImage don't check the validation of the combination. But they do check in the implementation of vkGetPhysicalDeviceImageFormatProperties2. I also tested the LLVMPIPE driver, it does work.