Device mismatch error when evaluating converted PT2E quantized ViT model on CUDA/MPS

[ofirgo]

🐛 Describe the bug

After converting a quantized model using convert_pt2e(), attempting to evaluate the model on CUDA or MPS devices results in a device mismatch error. The error indicates that some tensors remain on CPU while others are on the target device (CUDA/MPS), specifically during dequantization operations.

Error Message
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

The error occurs during torch.ops.aten.mul.Tensor operation between dequantized tensors.

Reproduce Steps

import torch
from executorch.backends.arm.ethosu import EthosUCompileSpec
from executorch.backends.arm.quantizer import EthosUQuantizer, get_symmetric_quantization_config
from torchao.quantization.pt2e.quantize_pt2e import convert_pt2e, prepare_pt2e
from torchvision.models import vit_b_16, ViT_B_16_Weights

device = 'cuda'  # or 'mps'
batch_size = 4
example_input = torch.randn(batch_size, 3, 224, 224)

weights = ViT_B_16_Weights.IMAGENET1K_V1
model = vit_b_16(weights=weights)
model.eval()

exported_program = torch.export.export(model, (example_input,))
graph_model = exported_program.module(check_guards=False)

# Configure quantizer and prepare model
compile_spec = EthosUCompileSpec(
    target="ethos-u55-128",
    system_config="Ethos_U55_High_End_Embedded",
    memory_mode="Shared_Sram",
)

quantizer = EthosUQuantizer(compile_spec)
operator_config = get_symmetric_quantization_config(is_per_channel=True)
quantizer.set_global(operator_config)

prepared_model = prepare_pt2e(graph_model, quantizer)

# Calibrate (simplified)
with torch.no_grad():
    prepared_model(example_input)

# Convert to quantized model
quantized_graph_model = convert_pt2e(prepared_model, fold_quantize=True)
quantized_graph_model._exported_training = False
quantized_graph_model.to(device)

# Error occurs here when trying to run on CUDA/MPS
test_input = torch.randn(batch_size, 3, 224, 224).to(device)
output = quantized_graph_model(test_input)  # RuntimeError!

Current Workaround
A helper function move_after_dequant_to_device() is required to manually move dequantization parameters to the target device:

def move_after_dequant_to_device(gm: fx.GraphModule, device="mps"):
    dev = torch.device(device)
    g = gm.graph
    for n in list(g.nodes):
        if n.op == "call_function" and n.target in DEQUANTS:
            with g.inserting_after(n):
                to_n = g.call_function(
                    torch.ops.aten._to_copy.default,
                    args=(n,),
                    kwargs=dict(dtype=None, layout=None, device=dev,
                                pin_memory=False, non_blocking=False, memory_format=None),
                )
            n.replace_all_uses_with(to_n)
            to_n.replace_input_with(to_n, n)
    gm.recompile()
    return gm

quantized_graph_model = move_after_dequant_to_device(quantized_graph_model, device)

Versions

PyTorch version: 2.9.1+cu128
Is debug build: False
CUDA used to build PyTorch: 12.8
ROCM used to build PyTorch: N/A

OS: Ubuntu 24.04.3 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version: Could not collect
CMake version: version 3.31.6
Libc version: glibc-2.39

Python version: 3.12.3 (main, Aug 14 2025, 17:47:21) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.12.55-74.119.amzn2023.x86_64-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to:
GPU models and configuration:
GPU 0: NVIDIA A100-SXM4-80GB
GPU 1: NVIDIA A100-SXM4-80GB
GPU 2: NVIDIA A100-SXM4-80GB
GPU 3: NVIDIA A100-SXM4-80GB
GPU 4: NVIDIA A100-SXM4-80GB
GPU 5: NVIDIA A100-SXM4-80GB
GPU 6: NVIDIA A100-SXM4-80GB
GPU 7: NVIDIA A100-SXM4-80GB

Nvidia driver version: 580.105.08
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.14.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.14.0
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 46 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 96
On-line CPU(s) list: 0-95
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Platinum 8275CL CPU @ 3.00GHz
CPU family: 6
Model: 85
Thread(s) per core: 2
Core(s) per socket: 24
Socket(s): 2
Stepping: 7
BogoMIPS: 5999.99
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch pti fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid mpx avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves ida arat pku ospke
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 1.5 MiB (48 instances)
L1i cache: 1.5 MiB (48 instances)
L2 cache: 48 MiB (48 instances)
L3 cache: 71.5 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-23,48-71
NUMA node1 CPU(s): 24-47,72-95
Vulnerability Gather data sampling: Unknown: Dependent on hypervisor status
Vulnerability Indirect target selection: Mitigation; Aligned branch/return thunks
Vulnerability Itlb multihit: KVM: Mitigation: VMX unsupported
Vulnerability L1tf: Mitigation; PTE Inversion
Vulnerability Mds: Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Meltdown: Mitigation; PTI
Vulnerability Mmio stale data: Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Vulnerable
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Vulnerable
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Retpoline
Vulnerability Srbds: Not affected
Vulnerability Tsa: Not affected
Vulnerability Tsx async abort: Not affected
Vulnerability Vmscape: Not affected

Versions of relevant libraries:
[pip3] executorch==1.0.1
[pip3] intel-openmp==2021.4.0
[pip3] mkl==2021.1.1
[pip3] mkl-devel==2021.1.1
[pip3] mkl-include==2021.1.1
[pip3] mypy_extensions==1.1.0
[pip3] numpy==2.3.5
[pip3] nvidia-cublas-cu12==12.8.4.1
[pip3] nvidia-cuda-cupti-cu12==12.8.90
[pip3] nvidia-cuda-nvrtc-cu12==12.8.93
[pip3] nvidia-cuda-runtime-cu12==12.8.90
[pip3] nvidia-cudnn-cu12==9.10.2.21
[pip3] nvidia-cudnn-frontend==1.14.1
[pip3] nvidia-cufft-cu12==11.3.3.83
[pip3] nvidia-curand-cu12==10.3.9.90
[pip3] nvidia-cusolver-cu12==11.7.3.90
[pip3] nvidia-cusparse-cu12==12.5.8.93
[pip3] nvidia-cusparselt-cu12==0.7.1
[pip3] nvidia-nccl-cu12==2.27.5
[pip3] nvidia-nvjitlink-cu12==12.8.93
[pip3] nvidia-nvtx-cu12==12.8.90
[pip3] onnx==1.18.0
[pip3] onnx-ir==0.1.11
[pip3] onnxscript==0.5.4
[pip3] optree==0.17.0
[pip3] pytorch_tokenizers==1.0.1
[pip3] pytorch-triton==3.4.0+gitc817b9b6
[pip3] tbb==2021.13.1
[pip3] torch==2.9.1+cu128
[pip3] torch_tensorrt==2.9.0a0
[pip3] torchao==0.14.0+git
[pip3] torchprofile==0.0.4
[pip3] torchvision==0.24.1+cu128
[pip3] triton==3.5.1

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

[SS-JIA]

@ofirgo I see that you are currently using the Arm Ethos U Quantizer to quantize the model. Quantizer classes are specialized to a target backend, so it is not recommended to use a particular quantizer if you are not planning to lower/run the model via the backend it was designed for. If you remove the quantization step, do you still run into the same issue?

Since you are trying to execute the model with CUDA and MPS, @larryliu0820 and @manuelcandales may have suggestions on the right way to quantize/execute your model.

[ofirgo]

@ofirgo I see that you are currently using the Arm Ethos U Quantizer to quantize the model. Quantizer classes are specialized to a target backend, so it is not recommended to use a particular quantizer if you are not planning to lower/run the model via the backend it was designed for. If you remove the quantization step, do you still run into the same issue?

Since you are trying to execute the model with CUDA and MPS, @larryliu0820 and @manuelcandales may have suggestions on the right way to quantize/execute your model.

Hi, thank you for the response.
I eventually intend to deploy quantized models onto the Ethos U target platform.
I'm currently just working on learning the system and its functionality. One of the steps that I need is to be able to evaluate the unlowered model.