[CoreML Backend] macOS 26.1 ANE regression: fp16 LLaMA inference produces inf/nan (worked on macOS 15.7)

[seyeong-han]

🐛 Describe the bug

Environment

• macOS Version: macOS-26.1-arm64-arm-64bit
• Previous Working Version: macOS-15.7.1-arm64-arm-64bit
• Hardware: Apple Silicon (M-series)
• ExecuTorch: [current version from repo]
• CoreMLTools: [check with pip show coremltools]
• Python: 3.11
• Model: LLaMA 3.2 1B

Issue Description

LLaMA 3.2 1B model exported with fp16 precision and CPU_AND_NE compute units produces inf/nan values during inference on macOS 26.1. The exact same model and code worked correctly on macOS 15.7.

This appears to be a macOS 26.1 Apple Neural Engine (ANE) regression affecting fp16 numerical precision.

I encountered this fp16 issue on Metal as well. issue

Regression Evidence

export.py

compile_specs = CoreMLBackend.generate_compile_specs(  # pyre-fixme[16]
        minimum_deployment_target=ct.target.iOS18,
        compute_precision={
            torch.float16: ct.precision.FLOAT16,
            torch.float32: ct.precision.FLOAT32,
        }[float_dtype],
        compute_unit=ct.ComputeUnit.CPU_AND_NE,
        model_type=CoreMLBackend.MODEL_TYPE.MODEL,  # pyre-fixme[16]
    )

Configuration	macOS 15.7	macOS 26.1
fp16 + CPU_AND_NE	✅ Works	❌ inf/nan
fp16 + CPU_ONLY	✅ Works	✅ Works
fp32 + CPU_AND_NE	✅ Works	✅ Works

Error Message

RuntimeError: probability tensor contains either `inf`, `nan` or element < 0


### Versions

PyTorch version: 2.9.1
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: macOS 26.1 (arm64)
GCC version: Could not collect
Clang version: 17.0.0 (clang-1700.3.19.1)
CMake version: version 4.0.3
Libc version: N/A

Python version: 3.11.14 (main, Oct 21 2025, 18:27:30) [Clang 20.1.8 ] (64-bit runtime)
Python platform: macOS-26.1-arm64-arm-64bit
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
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Apple M3 Pro

Versions of relevant libraries:
[pip3] executorch==1.0.0
[pip3] numpy==2.3.4
[pip3] pytorch_tokenizers==1.0.1
[pip3] torch==2.9.1
[pip3] torchao==0.14.0
[pip3] torchdata==0.11.0
[pip3] torchtune==0.6.1
[conda] executorch                1.0.0                    pypi_0    pypi
[conda] numpy                     2.3.4                    pypi_0    pypi
[conda] pytorch-tokenizers        1.0.1                    pypi_0    pypi
[conda] torch                     2.9.1                    pypi_0    pypi
[conda] torchao                   0.14.0                   pypi_0    pypi
[conda] torchdata                 0.11.0                   pypi_0    pypi
[conda] torchtune                 0.6.1                    pypi_0    pypi

[metascroy]

We are aware of the regression in iOS26 and are looking into it.