mlx-swift-lm Skill

1. Overview & Triggers

mlx-swift-lm is a Swift package for running Large Language Models (LLMs) and Vision-Language Models (VLMs) on Apple Silicon using MLX. It supports local inference, streaming generation, wired-memory coordination, tool calling, LoRA/DoRA fine-tuning, and embeddings.

When to Use This Skill

• Running LLM/VLM inference on macOS/iOS with Apple Silicon
• Streaming text generation from local models
• Coordinating concurrent inference with wired-memory policies and tickets
• Tool calling / function calling with models
• LoRA adapter training and fine-tuning
• Text embeddings for RAG/semantic search
• Porting model architectures from Python MLX-LM to Swift

Architecture Overview

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

2. Key File Reference

Purpose	File Path
Thread-safe model wrapper	Libraries/MLXLMCommon/ModelContainer.swift
Simplified chat API	Libraries/MLXLMCommon/ChatSession.swift
Generation & streaming APIs	Libraries/MLXLMCommon/Evaluate.swift
KV cache types	Libraries/MLXLMCommon/KVCache.swift
Wired-memory policies	Libraries/MLXLMCommon/WiredMemoryPolicies.swift
Wired-memory measurement helpers	Libraries/MLXLMCommon/WiredMemoryUtils.swift
Model configuration	Libraries/MLXLMCommon/ModelConfiguration.swift
Chat message types	Libraries/MLXLMCommon/Chat.swift
Tool call processing	Libraries/MLXLMCommon/Tool/ToolCallFormat.swift
Concurrency utilities	Libraries/MLXLMCommon/Utilities/SerialAccessContainer.swift
LLM factory & registry	Libraries/MLXLLM/LLMModelFactory.swift
VLM factory & registry	Libraries/MLXVLM/VLMModelFactory.swift
LoRA configuration	Libraries/MLXLMCommon/Adapters/LoRA/LoRAContainer.swift
LoRA training	Libraries/MLXLLM/LoraTrain.swift

3. Quick Start

LLM Chat (Simplest API)

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: "")
}

VLM with Image

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
)

Embeddings

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
}

4. Primary Workflow: LLM Inference

ChatSession API (Recommended)

ChatSession manages conversation history and KV cache automatically:

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()

Streaming with ModelContainer.generate(...)

For lower-level control, prepare UserInput and generate directly:

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")
    }
}

Generation API Surface (Evaluate.swift)

Use these depending on your control needs:

• generate(input:..., context:..., wiredMemoryTicket:) -> AsyncStream<Generation>: decoded text + tool calls.
• generateTask(..., wiredMemoryTicket:) -> (AsyncStream<Generation>, Task<Void, Never>): same output, plus task handle for deterministic cleanup when consumers stop early.
• generateTokens(..., wiredMemoryTicket:) -> AsyncStream<TokenGeneration>: raw token IDs.
• generateTokensTask(..., wiredMemoryTicket:) -> (AsyncStream<TokenGeneration>, Task<Void, Never>): raw tokens + task handle.
• GenerateStopReason: .stop, .length, .cancelled in final .info.

See references/generation.md for full patterns.

Tool Calling

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
    }
}

See references/tool-calling.md for multi-turn tool loops.

GenerateParameters

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
)

Wired Memory (Optional)

Use policy tickets to coordinate concurrent inference memory:

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: "")
    }
}

For policy selection, reservations, and measurement-based budgeting, see references/wired-memory.md.

Prompt Caching / History Re-hydration

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

let session = ChatSession(modelContainer, history: history)

5. Secondary Workflow: VLM Inference

Image Input Types

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

Video Input

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)

Multiple Images

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

VLM-Specific Processing

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

6. Best Practices

DO

// 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

// 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

Thread Safety

• ModelContainer is Sendable and thread-safe.
• ChatSession is not thread-safe; use one session per task/flow.
• MLXArray is not Sendable; keep it inside one isolation domain or use SendableBox transfer patterns.

Memory Management

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

7. Reference Links

Reference	When to Use
references/model-container.md	Loading models, ModelContainer API, ModelConfiguration
references/generation.md	generate, generateTask, raw token streaming APIs
references/wired-memory.md	Wired tickets, policies, budgeting, reservations
references/kv-cache.md	Cache types, memory optimization, cache serialization
references/concurrency.md	Thread safety, SerialAccessContainer, async patterns
references/tool-calling.md	Function calling, tool formats, ToolCallProcessor
references/tokenizer-chat.md	Tokenizer, Chat.Message, EOS tokens
references/supported-models.md	Model families, registries, model-specific config
references/lora-adapters.md	LoRA/DoRA/QLoRA, loading adapters
references/training.md	LoRATrain API, fine-tuning
references/embeddings.md	EmbeddingModel, pooling, use cases
references/model-porting.md	Porting models from Python MLX-LM to Swift

8. Deprecated Patterns Summary

If you see...	Use instead...
generate(... didGenerate:) callback	AsyncStream-based generation APIs
perform { model, tokenizer in }	perform { context in }
TokenIterator(prompt: MLXArray)	TokenIterator(input: LMInput)
ModelRegistry typealias	LLMRegistry or VLMRegistry
createAttentionMask(h:cache:[KVCache]?)	createAttentionMask(h:cache:KVCache?)

9. Automatic vs Manual Configuration

Automatic Behaviors

Feature	Details
EOS token loading	Loaded from config.json
EOS override	generation_config.json > config.json > defaults
EOS merging	All sources merged at generation time
EOS detection	Stops generation when EOS encountered
Chat template application	Applied by tokenizer / processor path
Tool call format detection	Inferred from model_type in config.json
Cache type selection	Driven by GenerateParameters (maxKVSize, kvBits)
Tokenizer loading	Loaded automatically from model assets
Model weight loading	Downloaded and loaded from Hugging Face/local directory

Optional Configuration

Feature	When to Configure
extraEOSTokens	Model has unlisted stop tokens
toolCallFormat	Override auto-detected tool parser format
maxKVSize	Enable sliding window cache
kvBits, kvGroupSize, quantizedKVStart	Enable and tune KV quantization
prefillStepSize	Tune prompt prefill chunking/perf tradeoff
wiredMemoryTicket	Coordinate policy-based wired-memory limits