Bug: Metal bfloat kernel crash when using Swift package

[jmousseau]

What happened?

I encountered a crash in ggml_metal_init when the first bfloat kernel is added. Below is an explanation of what I believe is happening. My device is running macOS 15.0.1 and has an M3 Max.

Because the macOS platform version in Package.swift is .v12, the condition to disable bfloat, __METAL_VERSION__ < 310, will always be true. However, at runtime the device family is used to set has_bfloat. As a result, there can be a bfloat mismatch between the Swift package's compiled Metal library and the runtime flag. This leads to a crash when adding bfloat Metal kernels as they don't exist.

Some ideas:

1. Consider bumping the macOS platform version to .v14 so consumers of the Swift package can use bfloat.
2. Set has_bfloat by also checking for the existence of an empty kernel in the MTLLibrary that will only exist if the library was compiled with bfloat support.

Name and Version

version: 4043 (60e17ce)
built with Apple clang version 16.0.0 (clang-1600.0.26.4) for arm64-apple-darwin24.0.0

What operating system are you seeing the problem on?

Mac

Relevant log output

No response

[ggerganov]

Option 1. seems simple enough - should we do that?

[jmousseau]

Option 1. seems simple enough - should we do that?

Seems reasonable to me. Only downside is that projects with v12 or v13 requirements would need to maintain a version with modified Package.swift and manually disabled has_bfloat.

[ggerganov]

Could you take a look at #10218?