huggingface · pcuenca · Oct 6, 2025 · Oct 5, 2025
diff --git a/Sources/Generation/Decoders.swift b/Sources/Generation/Decoders.swift
@@ -3,7 +3,7 @@ import CoreML
 
 // MARK: Greedy Decoding
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 func selectNextTokenUsingGreedyDecoding(from scores: MLTensor) -> MLTensor {
     let indices = scores.argmax(alongAxis: -1).reshaped(to: [1, 1])
     // Ensure indices are Int32 for concatenation with input tokens
@@ -19,7 +19,7 @@ func selectNextTokenUsingGreedyDecoding(from scores: MLTensor) -> MLTensor {
 ///
 /// - Parameter scores: Processed logits tensor [batch_size, vocab_size]
 /// - Returns: Sampled token ID tensor [batch_size, 1]
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 func selectNextTokenUsingSampling(from scores: MLTensor) -> MLTensor {
     // Convert logits to probabilities
     let probs = scores.softmax(alongAxis: -1)

diff --git a/Sources/Generation/Generation.swift b/Sources/Generation/Generation.swift
@@ -38,7 +38,7 @@ public typealias GenerationOutput = [Int]
 /// - Parameter tokens: Input token sequence
 /// - Parameter config: Generation configuration
 /// - Returns: Logits array for next token prediction
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public typealias NextTokenModel = (MLTensor, GenerationConfig) async -> MLTensor
 
 /// Callback for receiving generated tokens during streaming.
@@ -48,7 +48,7 @@ public typealias PredictionTokensCallback = (GenerationOutput) -> Void
 public typealias PredictionStringCallback = (String) -> Void
 
 /// Protocol for text generation implementations.
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public protocol Generation {
     /// Generates text from a prompt string.
     ///
@@ -62,7 +62,7 @@ public protocol Generation {
     func generate(config: GenerationConfig, prompt: String, model: NextTokenModel, tokenizer: Tokenizer, callback: PredictionStringCallback?) async -> String
 }
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 extension Generation {
     public func generate(
         config: GenerationConfig,
@@ -162,7 +162,7 @@ extension Generation {
     }
 }
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public extension Generation {
     /// Performs greedy or sampling-based text generation based on generation configuration.
     ///

diff --git a/Sources/Generation/LogitsWarper/LogitsProcessor.swift b/Sources/Generation/LogitsWarper/LogitsProcessor.swift
@@ -8,7 +8,7 @@ import CoreML
 /// such as temperature scaling, top-k/top-p filtering, and repetition penalties.
 ///
 /// Based on: https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public protocol LogitsProcessor {
     /// Processes logits for next token prediction.
     ///
@@ -28,7 +28,7 @@ public protocol LogitsProcessor {
 /// This class provides a convenient way to chain multiple logits processors together.
 /// Each processor is applied in order to the logits tensor, with the output of one
 /// processor becoming the input to the next.
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public struct LogitsProcessorList {
     public var processors: [any LogitsProcessor]
 

diff --git a/Sources/Generation/LogitsWarper/MinPLogitsWarper.swift b/Sources/Generation/LogitsWarper/MinPLogitsWarper.swift
@@ -15,7 +15,7 @@ import CoreML
 ///
 /// Based on:
 /// - https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py#L460
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public struct MinPLogitsWarper: LogitsProcessor {
     public let minP: Float
     public let minTokensToKeep: Int

diff --git a/Sources/Generation/LogitsWarper/RepetitionPenaltyLogitsProcessor.swift b/Sources/Generation/LogitsWarper/RepetitionPenaltyLogitsProcessor.swift
@@ -22,7 +22,7 @@ public enum LogitsProcessorError: Error {
 /// Based on:
 /// - https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py#L297
 /// - Paper: https://arxiv.org/abs/1909.05858
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public struct RepetitionPenaltyLogitsProcessor: LogitsProcessor {
     public let penalty: Float
 

diff --git a/Sources/Generation/LogitsWarper/TemperatureLogitsWarper.swift b/Sources/Generation/LogitsWarper/TemperatureLogitsWarper.swift
@@ -11,7 +11,7 @@ import CoreML
 /// Often used together with `TopPLogitsWarper` and `TopKLogitsWarper`.
 ///
 /// Based on: https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py#L231
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public struct TemperatureLogitsWarper: LogitsProcessor {
     public let temperature: Float
 

diff --git a/Sources/Generation/LogitsWarper/TopKLogitsWarper.swift b/Sources/Generation/LogitsWarper/TopKLogitsWarper.swift
@@ -10,7 +10,7 @@ import CoreML
 /// Pro tip: In practice, LLMs use top_k in the 5-50 range.
 ///
 /// Based on: https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py#L532
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public struct TopKLogitsWarper: LogitsProcessor {
     public let topK: Int
     public let filterValue: Float

diff --git a/Sources/Generation/LogitsWarper/TopPLogitsWarper.swift b/Sources/Generation/LogitsWarper/TopPLogitsWarper.swift
@@ -14,7 +14,7 @@ import CoreML
 /// Based on:
 /// - https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py#L465
 /// - Paper: https://arxiv.org/abs/1904.09751 (Nucleus Sampling)
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public struct TopPLogitsWarper: LogitsProcessor {
     public let topP: Float
     public let filterValue: Float

diff --git a/Sources/Models/LanguageModel.swift b/Sources/Models/LanguageModel.swift
@@ -12,7 +12,7 @@ import Generation
 import Hub
 import Tokenizers
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 /// A high-level interface for language model inference using CoreML.
 ///
 /// `LanguageModel` provides a convenient way to load and interact with pre-trained
@@ -72,7 +72,7 @@ public class LanguageModel {
     }
 }
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 private extension LanguageModel {
     static func contextRange(from model: MLModel) -> (min: Int, max: Int) {
         contextRange(from: model, inputKey: Keys.inputIds)
@@ -109,7 +109,7 @@ private extension LanguageModel {
     }
 }
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 extension LanguageModel {
     struct Configurations {
         var modelConfig: Config
@@ -118,7 +118,7 @@ extension LanguageModel {
     }
 }
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 extension LanguageModel {
     enum Keys {
         // Input keys
@@ -135,7 +135,7 @@ extension LanguageModel {
     }
 }
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public extension LanguageModel {
     /// Loads a compiled CoreML model from disk.
     ///
@@ -155,7 +155,7 @@ public extension LanguageModel {
     }
 }
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 extension LanguageModel {
     enum KVCacheAvailability {
         /// Language models that support KV cache via state. Implementation details for handling state
@@ -167,7 +167,7 @@ extension LanguageModel {
     }
 }
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public extension LanguageModel {
     /// Metadata fields associated to the Core ML model.
     var metadata: [MLModelMetadataKey: Any] {
@@ -296,7 +296,7 @@ public extension LanguageModel {
 // MARK: - Configuration Properties
 
 /// Asynchronous properties that are downloaded from the Hugging Face Hub configuration.
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public extension LanguageModel {
     /// The model configuration dictionary.
     ///
@@ -402,7 +402,7 @@ public extension LanguageModel {
 
 // MARK: - TextGenerationModel Conformance
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 extension LanguageModel: TextGenerationModel {
     /// The default generation configuration for this model.
     ///
@@ -424,7 +424,7 @@ extension LanguageModel: TextGenerationModel {
 ///
 /// Maintains a KV Cache as sequence generation progresses,
 /// using stateful Core ML buffers to minimize latency.
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public class LanguageModelWithStatefulKVCache: LanguageModel {
     private enum Mode {
         case prefilling

diff --git a/Sources/Models/LanguageModelTypes.swift b/Sources/Models/LanguageModelTypes.swift
@@ -15,7 +15,7 @@ import Tokenizers
 ///
 /// This protocol establishes the fundamental requirements for any language model
 /// that can perform next-token prediction and text generation tasks.
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public protocol LanguageModelProtocol {
     /// The name or path of the model.
     ///
@@ -50,7 +50,7 @@ public protocol LanguageModelProtocol {
     func predictNextTokenScores(_ input: MLTensor, config: GenerationConfig) async -> MLTensor
 }
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public extension LanguageModelProtocol {
     /// Function call syntax for next token prediction.
     ///
@@ -69,7 +69,7 @@ public extension LanguageModelProtocol {
 ///
 /// This protocol extends `LanguageModelProtocol` and `Generation` to provide
 /// high-level text generation functionality with configurable parameters.
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public protocol TextGenerationModel: Generation, LanguageModelProtocol {
     /// The default generation configuration for this model.
     ///
@@ -92,7 +92,7 @@ public protocol TextGenerationModel: Generation, LanguageModelProtocol {
     ) async throws -> String
 }
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 public extension TextGenerationModel {
     /// Default implementation of text generation that uses the underlying generation framework.
     ///

diff --git a/Tests/GenerationTests/GenerationIntegrationTests.swift b/Tests/GenerationTests/GenerationIntegrationTests.swift
@@ -4,7 +4,7 @@ import XCTest
 
 @testable import Generation
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 final class GenerationIntegrationTests: XCTestCase {
 
     // MARK: - Mock Model for Testing
@@ -343,7 +343,7 @@ final class GenerationIntegrationTests: XCTestCase {
 
 // MARK: - Test Helper
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 struct TestGeneration: Generation {
     func generate(
         config: GenerationConfig,

diff --git a/Tests/GenerationTests/LogitsProcessorTests.swift b/Tests/GenerationTests/LogitsProcessorTests.swift
@@ -3,7 +3,7 @@ import XCTest
 
 @testable import Generation
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 final class LogitsProcessorTests: XCTestCase {
     private let accuracy: Float = 0.0001
 
@@ -319,7 +319,7 @@ final class LogitsProcessorTests: XCTestCase {
 
 // MARK: - Test Helpers
 
-@available(macOS 15.0, iOS 18.0, *)
+@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, watchOS 11.0, *)
 func assertMLTensorEqual(
     _ tensor: MLTensor,
     expected: [Float],