Expose all models

[atdrendel]

This pull request exposes a simple actor for each supported model. The actual implementation of the models is LLM, which is just a convenience wrapper around MLX's stuff.

Each model conforms to a new protocol called ModelProtocol. This allows us to add extra functions for each model in just a single place: ModelProtocol.swift. The first example of this is request(_:maxTokenCount:).

Because of the reentrancy problem with Swift Actors, ModelProtocol.llm is wrapped inside of ActorLock, which is taken from Apple's swift-build/ASyncLock.swift.

According to MLX's documentation, the AI models themselves are not thread-safe, which means calls to them need to be serialized. However, because Swift Actors are reentrant, calling try await llm.request(_:maxTokenCount:) could immediately suspend and allow another reference to the same actor be called. This may not be a problem with library today, but I think it may be in the future, especially when we add support for KVCache. I think it's better to ensure that every call to ModelProtocol.someFunc is transactional, which is what we are doing by wrapping the implementation of every method on ModelProtocol inside of a AsyncLock.

[atdrendel]

I can't get Phi 3.5 MoE to work. When I try to run it, I see this error:

Caught error: keyNotFound(base: "SuScaledRotaryEmbedding", key: "_freqs")

Given the size of the model and the limited amount of RAM I have available, I'm not sure I'd even be able to run it if this error weren't thrown. So, I'm just continuing on to the next model.

[atdrendel]

I can't get Qwen 1.5 or Qwen 2.5 1.5B to run. They both fail with the same error:

Caught error: keyNotFound(base: "Linear", key: "weight")

As with the above, I'm not going to spend any time on trying to fix these errors. We have plenty of models to pick from, and we can fix it later or wait for an upstream update that fixes them.

[atdrendel]

These errors are known:

• Qwen 2.5 fails to load: Key weight not found in Linear ml-explore/mlx-swift-examples#214
• Error: "Failed: keyNotFound(base: "Linear", key: "weight")" for Qwen LLMs with mlx-swift 0.21.3 ml-explore/mlx-swift-examples#210

[atdrendel]

Yeah, I've got the same error for OpenELM:

Caught error: keyNotFound(base: "Linear", key: "weight")