Swift Mlx Lm

MLXLMCommon     - Core infra (ModelContainer, ChatSession, Evaluate, KVCache, wired memory helpers)
MLXLLM          - Text-only LLM support (Llama, Qwen, Gemma, Phi, DeepSeek, etc.)
MLXVLM          - Vision-Language Models (Qwen-VL, PaliGemma, Gemma3, etc.)
MLXEmbedders    - Embedding models and pooling utilities

import MLXLLM
import MLXLMCommon
import MLXLMHuggingFace  // from swift-huggingface-mlx
import MLXLMTokenizers   // from swift-tokenizers-mlx

let modelContainer = try await LLMModelFactory.shared.loadContainer(
    from: HubClient.default,
    using: TokenizersLoader(),
    configuration: .init(id: "mlx-community/Qwen3-4B-4bit")
)

let session = ChatSession(modelContainer)

let response = try await session.respond(to: "What is Swift?")
print(response)

for try await chunk in session.streamResponse(to: "Explain structured concurrency") {
    print(chunk, terminator: "")
}

import MLXVLM
import MLXLMCommon
import MLXLMHuggingFace  // from swift-huggingface-mlx
import MLXLMTokenizers   // from swift-tokenizers-mlx

let modelContainer = try await VLMModelFactory.shared.loadContainer(
    from: HubClient.default,
    using: TokenizersLoader(),
    configuration: .init(id: "mlx-community/Qwen2-VL-2B-Instruct-4bit")
)

let session = ChatSession(modelContainer)
let image = UserInput.Image.url(imageURL)

let response = try await session.respond(
    to: "Describe this image",
    image: image,
    video: nil
)

import MLXEmbedders
import MLXEmbeddersHuggingFace  // from swift-huggingface-mlx
import MLXLMTokenizers          // from swift-tokenizers-mlx

let container = try await loadModelContainer(
    from: HubClient.default,
    using: TokenizersLoader(),
    configuration: ModelConfiguration(id: "mlx-community/bge-small-en-v1.5-mlx")
)

let embeddings = await container.perform { model, tokenizer, pooler in
    let tokens = tokenizer.encode(text: "Hello world")
    let input = MLXArray(tokens).expandedDimensions(axis: 0)
    let output = model(input)
    let pooled = pooler(output, normalize: true)
    eval(pooled)
    return pooled
}

let session = ChatSession(
    modelContainer,
    instructions: "You are a helpful assistant",
    generateParameters: GenerateParameters(maxTokens: 500, temperature: 0.7)
)

let r1 = try await session.respond(to: "What is 2+2?")
let r2 = try await session.respond(to: "And if you multiply that by 3?")

await session.clear()

let userInput = UserInput(prompt: "Hello")
let lmInput = try await modelContainer.prepare(input: userInput)

let stream = try await modelContainer.generate(
    input: lmInput,
    parameters: GenerateParameters()
)

for await generation in stream {
    switch generation {
    case .chunk(let text):
        print(text, terminator: "")
    case .toolCall(let call):
        print("Tool call: \(call.function.name)")
    case .info(let info):
        print("\nStop reason: \(info.stopReason)")
        print("\(info.tokensPerSecond) tok/s")
    }
}

struct WeatherInput: Codable { let location: String }
struct WeatherOutput: Codable { let temperature: Double; let conditions: String }

let weatherTool = Tool<WeatherInput, WeatherOutput>(
    name: "get_weather",
    description: "Get current weather",
    parameters: [.required("location", type: .string, description: "City name")]
) { _ in
    WeatherOutput(temperature: 22.0, conditions: "Sunny")
}

let userInput = UserInput(
    prompt: .text("What's the weather in Tokyo?"),
    tools: [weatherTool.schema]
)

let lmInput = try await modelContainer.prepare(input: userInput)
let stream = try await modelContainer.generate(input: lmInput, parameters: GenerateParameters())

for await generation in stream {
    switch generation {
    case .chunk(let text):
        print(text, terminator: "")
    case .toolCall(let call):
        let result = try await call.execute(with: weatherTool)
        print("\nWeather: \(result.conditions)")
    case .info:
        break
    }
}

let params = GenerateParameters(
    maxTokens: 1000,            // nil = unlimited
    maxKVSize: 4096,            // Sliding window (RotatingKVCache)
    kvBits: 4,                  // Quantized cache (4 or 8)
    kvGroupSize: 64,            // Quantization group size
    quantizedKVStart: 0,        // Token index to start KV quantization
    temperature: 0.7,           // 0 = greedy / argmax
    topP: 0.9,                  // Nucleus sampling
    repetitionPenalty: 1.1,     // Penalize repeats
    repetitionContextSize: 20,  // Penalty window
    prefillStepSize: 512        // Prompt prefill chunk size
)

let policy = WiredSumPolicy()
let ticket = policy.ticket(size: estimatedBytes, kind: .active)

let userInput = UserInput(prompt: "Summarize this text")
let lmInput = try await modelContainer.prepare(input: userInput)

let stream = try await modelContainer.generate(
    input: lmInput,
    parameters: GenerateParameters(),
    wiredMemoryTicket: ticket
)

for await generation in stream {
    if case .chunk(let text) = generation {
        print(text, terminator: "")
    }
}

let history: [Chat.Message] = [
    .system("You are helpful"),
    .user("Hello"),
    .assistant("Hi there!")
]

let session = ChatSession(modelContainer, history: history)

let imageFromURL = UserInput.Image.url(fileURL)
let imageFromCI = UserInput.Image.ciImage(ciImage)
let imageFromArray = UserInput.Image.array(mlxArray)

let videoFromURL = UserInput.Video.url(videoURL)
let videoFromAsset = UserInput.Video.avAsset(avAsset)
let videoFromFrames = UserInput.Video.frames(videoFrames)

let response = try await session.respond(to: "What happens in this video?", video: videoFromURL)

let images: [UserInput.Image] = [.url(url1), .url(url2)]
let response = try await session.respond(to: "Compare these two images", images: images, videos: [])

let session = ChatSession(
    modelContainer,
    processing: UserInput.Processing(resize: CGSize(width: 512, height: 512))
)

// DO: Prefer ChatSession for multi-turn chat UX
let session = ChatSession(modelContainer)

// DO: Prepare UserInput before container-level generation
let userInput = UserInput(prompt: "Hello")
let lmInput = try await modelContainer.prepare(input: userInput)

// DO: Use task-handle variants for early-stop scenarios
let (stream, task) = generateTask(
    promptTokenCount: lmInput.text.tokens.size,
    modelConfiguration: context.configuration,
    tokenizer: context.tokenizer,
    iterator: iterator
)
for await item in stream {
    if shouldStop { break }
}
await task.value

// DO: Use wired tickets when coordinating concurrent workloads
let ticket = WiredSumPolicy().ticket(size: estimatedBytes)
let _ = try await modelContainer.generate(input: lmInput, parameters: params, wiredMemoryTicket: ticket)

// DON'T: Skip prepare(input:) before container-level generation.
// ModelContainer.generate expects LMInput, not UserInput.

// DON'T: Share MLXArray across tasks (not Sendable)
let array = MLXArray(...)
Task { _ = array.sum() } // wrong

// DON'T: Ignore task completion after early-break on low-level streams
for await item in stream {
    if shouldStop { break }
}
// await task.value is required for deterministic cleanup

let slidingWindow = GenerateParameters(maxKVSize: 4096)
let quantizedKV = GenerateParameters(kvBits: 4, kvGroupSize: 64)
await session.clear()