Using highly optimized Metal Shading Language (MSL) code, I pushed the MacBook Air M2 to its performance limits with the deformable_attention_universal kernel. The results demonstrate both the efficiency of the code and the exceptional power of Apple Silicon. The total computational workload exceeded 8.455 quadrillion FLOPs, equivalent to processing 8,455 trillion operations. On average, the code sustained a throughput of 85.37 TFLOPS, showcasing the chip’s remarkable ability to handle massive workloads. Peak instantaneous performance reached approximately 673.73 TFLOPS, reflecting near-optimal utilization of the GPU cores. Despite this intensity, the cumulative GPU runtime remained under 100 seconds, highlighting the code’s efficiency and time optimization. The fastest iteration achieved a record processing time of only 0.051 ms, demonstrating minimal bottlenecks and excellent responsiveness. Memory management was equally impressive: peak GPU memory usage never exceeded 2 MB, reflecting efficient use of the M2’s Unified Memory. This minimizes data transfer overhead and ensures smooth performance across repeated workloads. Overall, these results confirm that a well-optimized Metal implementation can unlock the full potential of Apple Silicon, delivering exceptional computational density, processing speed, and memory efficiency. The MacBook Air M2, often considered an energy-efficient consumer laptop, is capable of handling highly intensive workloads at performance levels typically expected from much larger GPUs. This test validates both the robustness of the Metal code and the extraordinary capabilities of the M2 chip for high-performance computing tasks.

Hi everyone 👋 I'd like to use coremltools to see how well a model performs on a remote device as part of a CI/CD pipeline. According to the Core ML Tools "Debugging and Performance Utilities" guide, remote devices must be in a "connected" state in order for coremltools to install the ModelRunner application. The devices in our system have a "paired" state, and I'm unable to set the them as "connected." The only way I know how to connect a device is to physically plug it in to a computer and open Xcode. I don't have physical access to the devices in the CI/CD system, and the host computer that interacts with them doesn't have Xcode installed. Here are some questions I've been looking into and would love some help answering: Has anyone managed to use the coremltools performance utilities in a similar system? Can you put a device in a "connected" state if you don't have physical access to the device and if you only have access to Xcode command line tools and not the Xcode app? Is it at all possible to install the coremltools ModelRunner application on a "paired" device, for example, by manually building the app and installing it with devicectl? Would other utilities, such as the MLModelBenchmarker work as expected if the app is installed this way? Thank you!

I'm using python 3.9.6, tensorflow 2.20.0, tensorflow-metal 1.2.0, and when I try to run import tensorflow as tf It gives Traceback (most recent call last): File "/Users/haoduoyu/Code/demo.py", line 1, in <module> import tensorflow as tf File "/Users/haoduoyu/Code/test/lib/python3.9/site-packages/tensorflow/__init__.py", line 438, in <module> _ll.load_library(_plugin_dir) File "/Users/haoduoyu/Code/test/lib/python3.9/site-packages/tensorflow/python/framework/load_library.py", line 151, in load_library py_tf.TF_LoadLibrary(lib) tensorflow.python.framework.errors_impl.NotFoundError: dlopen(/Users/haoduoyu/Code/test/lib/python3.9/site-packages/tensorflow-plugins/libmetal_plugin.dylib, 0x0006): Library not loaded: @rpath/_pywrap_tensorflow_internal.so Referenced from: <8B62586B-B082-3113-93AB-FD766A9960AE> /Users/haoduoyu/Code/test/lib/python3.9/site-packages/tensorflow-plugins/libmetal_plugin.dylib Reason: tried: '/Users/haoduoyu/Code/test/lib/python3.9/site-packages/tensorflow-plugins/../_solib_darwin_arm64/_U@local_Uconfig_Utf_S_S_C_Upywrap_Utensorflow_Uinternal___Uexternal_Slocal_Uconfig_Utf/_pywrap_tensorflow_internal.so' (no such file), '/Users/haoduoyu/Code/test/lib/python3.9/site-packages/tensorflow-plugins/../_solib_darwin_arm64/_U@local_Uconfig_Utf_S_S_C_Upywrap_Utensorflow_Uinternal___Uexternal_Slocal_Uconfig_Utf/_pywrap_tensorflow_internal.so' (no such file) As long as I uninstall tensorflow-metal, nothing goes wrong. How can I fix this problem?

: Hello, I’m seeking clarification on whether Apple provides any framework or API that enables deep integration between Siri and advanced AI assistants (such as ChatGPT), including system-level functions like voice interaction, navigation, cross-platform syncing, and operational access similar to Siri’s own capabilities. If no such option exists today, I would appreciate guidance on the recommended path or approved third-party solutions for building a unified, voice-first experience across Apple’s ecosystem. Thank you for your time and insight.

Hello, I am interested in using jax-metal to train ML models using Apple Silicon. I understand this is experimental. After installing jax-metal according to https://developer.apple.com/metal/jax/, my python code fails with the following error JaxRuntimeError: UNKNOWN: -:0:0: error: unknown attribute code: 22 -:0:0: note: in bytecode version 6 produced by: StableHLO_v1.12.1 My issue is identical to the one reported here https://github.com/jax-ml/jax/issues/26968#issuecomment-2733120325, and is fixed by pinning to jax-metal 0.1.1., jax 0.5.0 and jaxlib 0.5.0. Thank you!

Download the Foundation Models Adaptor Training Toolkit Hi, after I clicked on the download button, I was redirected to this page https://developer.apple.com and did not download the toolkit.

Hi, I'm currently using Metal Performance Shaders Graph (MPSGraphExecutable) to run neural network inference operations as part of a metal rendering pipeline. I also tried to profile the usage of neural engine when running inference using MPSGraphExecutable but the graph shows no sign of neural engine usage. However, when I used the coreML model inspection tool in xcode and run performance report, it was able to use ANE. Does MPSGraphExecutable automatically utilize the Apple Neural Engine (ANE) when running inference operations, or does it only execute on GPU? My model (Core ML Package) was converted from a pytouch model using coremltools with ML program type and support iOS17.0+. Any insights or documentation references would be greatly appreciated!

I am writing a custom package wrapping Foundation Models which provides a chain-of-thought with intermittent self-evaluation among other things. At first I was designing this package with the command line in mind, but after seeing how well it augments the models and makes them more intelligent I wanted to try and build a SwiftUI wrapper around the package. When I started I was using synchronous generation rather than streaming, but to give the best user experience (as I've seen in the WWDC sessions) it is necessary to provide constant feedback to the user that something is happening. I have created a super simplified example of my setup so it's easier to understand. First, there is the Reasoning conversation item, which can be converted to an XML representation which is then fed back into the model (I've found XML works best for structured input) public typealias ConversationContext = XMLDocument extension ConversationContext { public func toPlainText() -> String { return xmlString(options: [.nodePrettyPrint]) } } /// Represents a reasoning item in a conversation, which includes a title and reasoning content. /// Reasoning items are used to provide detailed explanations or justifications for certain decisions or responses within a conversation. @Generable(description: "A reasoning item in a conversation, containing content and a title.") struct ConversationReasoningItem: ConversationItem { @Guide(description: "The content of the reasoning item, which is your thinking process or explanation") public var reasoningContent: String @Guide(description: "A short summary of the reasoning content, digestible in an interface.") public var title: String @Guide(description: "Indicates whether reasoning is complete") public var done: Bool } extension ConversationReasoningItem: ConversationContextProvider { public func toContext() -> ConversationContext { // <ReasoningItem title="${title}"> // ${reasoningContent} // </ReasoningItem> let root = XMLElement(name: "ReasoningItem") root.addAttribute(XMLNode.attribute(withName: "title", stringValue: title) as! XMLNode) root.stringValue = reasoningContent return ConversationContext(rootElement: root) } } Then there is the generator, which creates a reasoning item from a user query and previously generated items: struct ReasoningItemGenerator { var instructions: String { """ <omitted for brevity> """ } func generate(from input: (String, [ConversationReasoningItem])) async throws -> sending LanguageModelSession.ResponseStream<ConversationReasoningItem> { let session = LanguageModelSession(instructions: instructions) // build the context for the reasoning item out of the user's query and the previous reasoning items let userQuery = "User's query: \(input.0)" let reasoningItemsText = input.1.map { $0.toContext().toPlainText() }.joined(separator: "\n") let context = userQuery + "\n" + reasoningItemsText let reasoningItemResponse = try await session.streamResponse( to: context, generating: ConversationReasoningItem.self) return reasoningItemResponse } } I'm not sure if returning LanguageModelSession.ResponseStream<ConversationReasoningItem> is the right move, I am just trying to imitate what session.streamResponse returns. Then there is the orchestrator, which I can't figure out. It receives the streamed ConversationReasoningItems from the Generator and is responsible for streaming those to SwiftUI later and also for evaluating each reasoning item after it is complete to see if it needs to be regenerated (to keep the model on-track). I want the users of the orchestrator to receive partially generated reasoning items as they are being generated by the generator. Later, when they finish, if the evaluation passes, the item is kept, but if it fails, the reasoning item should be removed from the stream before a new one is generated. So in-flight reasoning items should be outputted aggresively. I really am having trouble figuring this out so if someone with more knowledge about asynchronous stuff in Swift, or- even better- someone who has worked on the Foundation Models framework could point me in the right direction, that would be awesome!

I've built an iOS app with a novel approach to AI safety: a deterministic, pre-inference validation layer called Newton Engine. Instead of relying on the LLM to self-moderate, Newton validates every prompt BEFORE it reaches the model. It uses shape theory and semantic analysis to detect: • Corrosive frames (self-harm language patterns) • Logical contradictions (requests that undermine themselves) • Delegation attempts (asking AI to make human decisions) • Jailbreak patterns (prompt injection, role-play escapes) • Hallucination triggers (requests for fabricated citations) The system achieves a 96% adversarial catch rate across 847 test cases, with zero false positives on benign prompts. Key technical details: • Pure Swift/SwiftUI, no external dependencies • Runs entirely on-device (no server calls for validation) • Deterministic (same input always produces same output) • Auditable (full trace logging for every validation) I'm preparing to submit to the App Store and wanted to ask: Are there specific App Review guidelines I should reference for AI safety claims? Is there interest from Apple in deterministic governance layers for Apple Intelligence integration? Any recommendations for demonstrating safety compliance during review? The app is called Ada, and the engine is open source at: github.com/jaredlewiswechs/ada-newton Happy to share technical documentation or discuss the architecture with anyone interested. See: parcri.net

I'm developing a tennis ball tracking feature using Vision Framework in Swift, specifically utilizing VNDetectedObjectObservation and VNTrackObjectRequest. Occasionally (but not always), I receive the following runtime error: Failed to perform SequenceRequest: Error Domain=com.apple.Vision Code=9 "Internal error: unexpected tracked object bounding box size" UserInfo={NSLocalizedDescription=Internal error: unexpected tracked object bounding box size} From my investigation, I suspect the issue arises when the bounding box from the initial observation (VNDetectedObjectObservation) is too small. However, Apple's documentation doesn't clearly define the minimum bounding box size that's considered valid by VNTrackObjectRequest. Could someone clarify: What is the minimum acceptable bounding box width and height (normalized) that Vision Framework's VNTrackObjectRequest expects? Is there any recommended practice or official guidance for bounding box size validation before creating a tracking request? This information would be extremely helpful to reliably avoid this internal error. Thank you!

I am attempting to install Tensorflow on my M1 and I seem to be unable to find the correct matching versions of jax, jaxlib and numpy to make it all work. I am in Bash, because the default shell gave me issues. I downgraded to python 3.10, because with 3.13, I could not do anything right. Current actions: bash-3.2$ python3.10 -m venv ~/venv-metal bash-3.2$ python --version Python 3.10.16 python3.10 -m venv ~/venv-metal source ~/venv-metal/bin/activate python -m pip install -U pip python -m pip install tensorflow-macos And here, I keep running tnto errors like: (venv-metal):~$ pip install tensorflow-macos tensorflow-metal ERROR: Could not find a version that satisfies the requirement tensorflow-macos (from versions: none) ERROR: No matching distribution found for tensorflow-macos What is wrong here? How can I fix that? It seems like the system wants to use the x86 version of python ... which can't be right.

Hello! I'm following the Foundation Models adapter training guide (https://developer.apple.com/apple-intelligence/foundation-models-adapter/) on my NVIDIA DGX Spark box. I'm able to train on my own data but the example notebook fails when I try to export the artifact as an fmadapter. I get the following error for the code block I'm trying to run. I haven't touched any of the code in the export folder. I tried exporting it on my Mac too and got the same error as well (given below). Would appreciate some more clarity around this. Thank you. Code Block: from export.export_fmadapter import Metadata, export_fmadapter metadata = Metadata( author="3P developer", description="An adapter that writes play scripts.", ) export_fmadapter( output_dir="./", adapter_name="myPlaywritingAdapter", metadata=metadata, checkpoint="adapter-final.pt", draft_checkpoint="draft-model-final.pt", ) Error: --------------------------------------------------------------------------- ModuleNotFoundError Traceback (most recent call last) Cell In[10], line 1 ----> 1 from export.export_fmadapter import Metadata, export_fmadapter 3 metadata = Metadata( 4 author="3P developer", 5 description="An adapter that writes play scripts.", 6 ) 8 export_fmadapter( 9 output_dir="./", 10 adapter_name="myPlaywritingAdapter", (...) 13 draft_checkpoint="draft-model-final.pt", 14 ) File /workspace/export/export_fmadapter.py:11 8 from typing import Any 10 from .constants import BASE_SIGNATURE, MIL_PATH ---> 11 from .export_utils import AdapterConverter, AdapterSpec, DraftModelConverter, camelize 13 logger = logging.getLogger(__name__) 16 class MetadataKeys(enum.StrEnum): File /workspace/export/export_utils.py:15 13 import torch 14 import yaml ---> 15 from coremltools.libmilstoragepython import _BlobStorageWriter as BlobWriter 16 from coremltools.models.neural_network.quantization_utils import _get_kmeans_lookup_table_and_weight 17 from coremltools.optimize._utils import LutParams ModuleNotFoundError: No module named 'coremltools.libmilstoragepython'

I have used mlx_lm.lora to fine tune a mistral-7b-v0.3-4bit model with my data. I fused the mistral model with my adapters and upload the fused model to my directory on huggingface. I was able to use mlx_lm.generate to use the fused model in Terminal. However, I don't know how to load the model in Swift. I've used Imports import SwiftUI import MLX import MLXLMCommon import MLXLLM let modelFactory = LLMModelFactory.shared let configuration = ModelConfiguration( id: "pharmpk/pk-mistral-7b-v0.3-4bit" ) // Load the model off the main actor, then assign on the main actor let loaded = try await modelFactory.loadContainer(configuration: configuration) { progress in print("Downloading progress: \(progress.fractionCompleted * 100)%") } await MainActor.run { self.model = loaded } I'm getting an error runModel error: downloadError("A server with the specified hostname could not be found.") Any suggestions? Thanks, David PS, I can load the model from the app bundle // directory: Bundle.main.resourceURL! but it's too big to upload for Testflight

The WWDC25: Explore large language models on Apple silicon with MLX video talks about using your own data to fine-tune a large language model. But the video doesn't explain what kind of data can be used. The video just shows the command to use and how to point to the data folder. Can I use PDFs, Word documents, Markdown files to train the model? Are there any code examples on GitHub that demonstrate how to do this?

Hi all, I'm tuning my app prediction speed with Core ML model. I watched and tried the methods in video: Improve Core ML integration with async prediction and Optimize your Core ML usage. I also use instruments to look what's the bottleneck that my prediction speed cannot be faster. Below is the instruments result with my app. its prediction duration is 10.29ms And below is performance report shows the average speed of prediction is 5.55ms, that is about half time of my app prediction! Below is part of my instruments records. I think the prediction should be considered quite frequent. Could it be faster? How to be the same prediction speed as performance report? The prediction speed on macbook Pro M2 is nearly the same as macbook Air M1!

I followed below url for converting Llama-3.1-8B-Instruct model but always fails even i have 64GB of free space after downloading model from huggingface. https://machinelearning.apple.com/research/core-ml-on-device-llama Also tried with other models Llama-3.1-1B-Instruct & Llama-3.1-3B-Instruct models those are converted but while doing performance test in xcode fails for all compunits. Is there any source code to run llama models in ios app.

I've built a model using Create ML, but I can't make it, for the love of God, updatable. I can't find any checkbox or anything related. It's an Activity Classifier, if it matters. I want to continue training it on-device using MLUpdateTask, but the model, as exported from Create ML, fails with error: Domain=com.apple.CoreML Code=6 "Failed to unarchive update parameters. Model should be re-compiled." UserInfo={NSLocalizedDescription=Failed to unarchive update parameters. Model should be re-compiled.}

I’m considering creating an ILMessageFilterExtension using a mini LLM/SLM to detect fraud and I’ve read it has strict memory limits yet I can’t find it in the documentation. What’s the set limit or any other constraints impacting the feasibility of running 100-500mb model?

Bear with me, please. Please make sure a highly skilled technical person reads and understands this. I want to describe my vision for (AI/Algorithmically) Optimised Operating Systems. To explain it properly, I will describe the process to build it (pseudo). Required Knowledge (no particular order): Processor Logic Circuits, LLM models, LLM tool usage, Python OO coding, Procedural vs OO, NLP fuzzy matching, benchmarking, canvas/artefacts/dynamic HTML interfaces, concepts of how AI models are vastly compressed and miniaturised forms of full data, Algorithmic vs AI. First, take all OO Python code (example) on GitHub (example), then separate each function from each object into its own procedure (procedural logic) by making a logical procedural list of actions to perform only that function based on its entire dependency chain (i.e. all other objects it relies on). Relate all compiled functions using (for example) fuzzy matching on the name, or AI-based functional profiling to get multiple instances of each function. Starting with the most used function, test each one against the others that perform the same task for bugs and completeness. Determine the fastest, most optimal version of that function (and every function). Add a single instance of each most optimal function to the centralised tool codebase, which will later be utilised by the language models. This ensures we rely only on the most optimised function for each and every use case — with every program using one shared instance of that function instead of compiling it separately. In this way, we optimise all existing (Python OO) code based on maximum availability of certain resources. Use the optimised code to teach the AI how to map each user task directly to optimised code (i.e. each user-facing function in unoptimised program code to its optimised version). Later, teach the codebase how to optimise these functions directly to the logic circuits on the processor — as an Operating System — using essentially the same process. Eventually, use the most optimised form for tasks via a dynamic canvas/artefacts UI interface generated by the Start Menu using the language model. Allow custom interfaces to emerge dynamically from the available functions based on language understanding. Grow and optimise the algorithmic codebase until the entire operating system — with versioning — evolves based on the AI-generated user interfaces (compressing them into new versions of the tools codebase). Please don’t give up until you understand this concept completely and nothing in the description is beyond your full comprehension. It’s a highly significant step forward.

Hi there, I have a custom keypoint detection model and want to use it via vision's CoremlRequest API. Here's some complication for input and output: For input My model expect 512x512 a image. Which would be resized and padded from a 1920x1080 frame. I use the .scaleToFit option, but can I also specify the color used for padding? For output: My model output a CoreMLFeatureValueObservation, can I have it output in a format vision recognizes? such as joints/keypoints If my model is able to output in a format vision recognizes, would it take care to restoring the coordinates back to the original frame? (undo the padding) If not, how do I restore it from .scaletofit option? Best,