Fix compatibility with MLX-Swift-Examples v2.25.7

[yayashuxue]

MLX-Swift v2.25.7 Compatibility Changes

Overview

This document summarizes the changes made to ensure FastVLM compatibility with MLX-Swift-Examples v2.25.7 (Latest package)

Files Modified

• .gitignore (+4 lines)
• app/FastVLM/FastVLM.swift (+27 lines, -9 lines)

Detailed Changes

1. .gitignore Updates

Added additional Xcode-related entries to prevent tracking of user-specific files:

xcuserdata/
*.xcuserstate
*.xcuserdatad/
project.pbxproj.orig

2. FastVLM.swift - Attention Mask API Changes

Before (v2.25.6 and earlier):

let mask = createAttentionMask(h: h, cache: cache)
h = layer(h, mask: mask, cache: cache?[i])

After (v2.25.7 compatible):

let mask = createAttentionMask(h: h, cache: cache, returnArray: true)
let maskArray: MLXArray? = {
    switch mask {
    case .array(let array):
        return array
    case .causal, .none:
        return nil
    case .arrays(_):
        return nil
    }
}()

for (i, layer) in layers.enumerated() {
    h = layer(h, mask: maskArray, cache: cache?[i])
}

Key Changes:

• Added returnArray: true parameter to createAttentionMask()
• Extract MLXArray from the returned AttentionMask enum
• Pass MLXArray? instead of AttentionMask to layer functions

3. FastVLMProcessor - Message Handling Improvements

Before:

var messages = prompt.asMessages()
if messages[0]["role"] != "system" {
    // ...
}
var lastMessage = messages[lastIndex]["content"] ?? ""

After:

var messages: [Message]
switch prompt {
case .text(let text):
    messages = [["role": "user", "content": text]]
case .messages(let msgs):
    messages = msgs
case .chat(let chatMsgs):
    messages = chatMsgs.map { ["role": $0.role.rawValue, "content": $0.content] }
}
if (messages[0]["role"] as? String) != "system" {
    // ...
}
var lastMessage: String = (messages[lastIndex]["content"] as? String) ?? ""

Key Changes:

• Explicit type casting for dictionary values
• Proper handling of different prompt types
• Improved type safety with explicit String casting

4. Image Token Processing

Before:

lastMessage += Array(repeating: config.imageToken, count: numImageTokens)
    .joined()

After:

let imageTokens = String(repeating: config.imageToken, count: numImageTokens)
lastMessage += imageTokens

Key Changes:

• Simplified string repetition using String(repeating:count:)
• More efficient than creating array and joining

5. LMInput.ProcessedImage API Update

Before:

let image = LMInput.ProcessedImage(pixels: pixels, imageGridThw: [thw])
let gridThw = input.image?.imageGridThw

After:

let image = LMInput.ProcessedImage(pixels: pixels, frames: [thw])
let gridThw = input.image?.frames

Key Changes:

• Property renamed from imageGridThw to frames
• Maintains same functionality with updated API

Impact

• ✅ Maintains backward compatibility in functionality
• ✅ Fixes compilation errors with MLX-Swift-Examples v2.25.7
• ✅ Improves type safety in message processing
• ✅ Tested and working on iOS device

Testing Status

• Compilation successful with MLX-Swift-Examples v2.25.7
• iOS device testing completed

[yayashuxue]

@federicobucchi @rwebb key changes focus on making the ios build compatible with the latest XLM package