[build failed]Compiler encountered an internal error

## env
* system
- ProductName: macOS
- ProductVersion: 12.6.1
- Intel version.
* python:
- Python 3.12.1 (v3.12.1:2305ca5144, Dec 7 2023, 17:23:39) [Clang 13.0.0 (clang-1300.0.29.30)] on darwin

## log

```

$python train_gpt2.py
using device: mps
loading weights from pretrained gpt: gpt2
model.safetensors: 100%|██████████████████████████████████████████████████████████████████| 548M/548M [00:38<00:00, 14.2MB/s]
generation_config.json: 100%|████████████████████████████████████████████████████████████████| 124/124 [00:00<00:00, 784kB/s]
loading cached tokens in data/tiny_shakespeare_val.bin
/AppleInternal/Library/BuildRoots/a0876c02-1788-11ed-b9c4-96898e02b808/Library/Caches/com.apple.xbs/Sources/MetalPerformanceShaders/MPSCore/Utility/MPSLibrary.mm:504: failed assertion `MPSKernel MTLComputePipelineStateCache unable to load function copyNDArrayData.
Compiler encountered an internal error: (null)
'
Abort trap: 6
/Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/multiprocessing/resource_tracker.py:254: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
warnings.warn('resource_tracker: There appear to be %d '
$
$
$
$
$
$python train_gpt2.py
using device: mps
loading weights from pretrained gpt: gpt2
loading cached tokens in data/tiny_shakespeare_val.bin
/AppleInternal/Library/BuildRoots/a0876c02-1788-11ed-b9c4-96898e02b808/Library/Caches/com.apple.xbs/Sources/MetalPerformanceShaders/MPSCore/Utility/MPSLibrary.mm:504: failed assertion `MPSKernel MTLComputePipelineStateCache unable to load function copyNDArrayData.
Compiler encountered an internal error: (null)
'
Abort trap: 6
```

Start by reproducing `python train_gpt2.py` on the reported macOS 12.6.1 and Python 3.12.1 environment, using the MPS device and the cached `data/tiny_shakespeare_val.bin` input. Inspect the failure around the `train_gpt2.py` entry point and determine the compatibility boundary; done means the command no longer aborts or the supported configuration and limitation are documented.

Written by the indexing model from the issue text.