[metascroy]

Adds a coreai_compute_unit backend option for the Core AI delegate, set the same way as coreai_cache_dir. It selects which compute units Core AI may use when it specializes a model on the device.

Stacked on #24374.

Option values

Each value is exactly one of the configurations the Core AI SDK provides through SpecializationOptions:

Value	Core AI SpecializationOptions	Allowed compute units
default (or absent)	.default	CPU, GPU, Neural Engine
cpu_only	.cpuOnly	CPU
prefer_cpu	init(preferredComputeUnitKind: .cpu)	CPU, GPU, Neural Engine
prefer_gpu	init(preferredComputeUnitKind: .gpu)	CPU, GPU, Neural Engine
prefer_neural_engine	init(preferredComputeUnitKind: .neuralEngine)	CPU, GPU, Neural Engine

The option mirrors the SDK and adds nothing of its own:

• cpu_only is the only restriction. A preference maximizes use of that unit but falls back to any allowed unit for operations it can't run.
• No other combinations. The SDK can't express ones like "CPU and Neural Engine only", because its allowed units can't be set.
• No separate allowed/preferred settings. One string covers every configuration the SDK can build.

BackendOptions<2> options;
ET_CHECK_OK_OR_RETURN_ERROR(options.set_option("coreai_compute_unit", "cpu_only"));
LoadBackendOptionsMap map;
ET_CHECK_OK_OR_RETURN_ERROR(map.set_options("CoreAIBackend", options.view()));
ET_CHECK_OK_OR_RETURN_ERROR(module.load(map));

Implementation

• coreai_backend.mm reads the option during init, next to coreai_cache_dir. Unknown values, non-string values and null fail with InvalidArgument before a loader is created. The error lists the valid values.
• ETCoreAIBridge.h adds an ETCoreAIComputeUnit enum, and ETCoreAICreateModelLoader now takes it. ETCoreAIModelLoader.swift maps each value to the matching SpecializationOptions.
• Caching: the cache key already includes the loader's specialization identity, so each value gets its own bookmark, source directory and lock stripe. default keeps its existing identity, so bookmarks from before this change still restore. clear_cache_for_pte still removes every variant of a PTE.
• Creating a loader only to evict a bookmark uses default, because eviction doesn't depend on specialization.
• The README documents the option.

Test plan

• coreai_host_test passes. A new delegate test checks that each value reaches the loader factory, that a missing option means default, and that unknown and non-string values are rejected before a loader is created.
• coreai_swift_bridge_test passes. A new Swift test checks that each value builds exactly the expected SpecializationOptions, that the five values have five different cache identities, and that default keeps the "default" identity.
• Manual multithreaded check, using a standalone runner that loads a two-delegate fp16 PTE with the five values assigned round-robin across threads:
  • Cold, 10 threads × 50 inferences: all outputs matched the reference, with one cache key per compute unit.
  • Warm, 50 threads × 200: all 10,000 outputs matched, with no new specializations; every load restored its own compute unit's bookmark.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24375

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[shoumikhin]

@claude review effort=max

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.