llama : expose C API to get layer device type

[okaris]

llama : expose C API to get layer device type

Adds llama_model_dev_layer(model, il) to retrieve the backend device type (CPU, GPU, ACCEL) for a given layer. This allows consumers (e.g. Python bindings) to inspect layer placement without exposing internal structures.

Implements an int32_t-based interface matching the pattern of existing public API functions such as llama_model_n_layer.

Notes:

• Follows naming convention: llama_model_dev_layer
• Pure C-compatible API (int32_t return)
• No changes to third-party deps, no extra headers
• Scoped, focused, standalone change
• Adheres to existing formatting and structure
• Useful for inspection/debugging/perf profiling via bindings

[okaris]

up

[okaris]

@ngxson @slaren @danbev not sure who to tag

[ggerganov]

I think this or similar change was discussed at some point - it does not belong to the public API. From the points that you listed:

Useful for inspection/debugging/perf profiling via bindings

This is the only meaningful one and it means that this is just for debugging purposes which makes it not suitable for the public API of the library. Won't merge unless there is a better reasoning to need such an API.

[okaris]

@ggerganov I understand the concern about keeping the public API minimal. My point is that there’s currently no way to confirm if a model actually (fully) ended up on the chosen accelerator. Since that decision is user driven, some form of feedback feels necessary. If a new function is not acceptable, could this information be returned from the load methods instead? I’m happy to adjust the contribution in whatever way fits best.