Daeseon Yoo

문제

Phase 9.0 Week 2-3 wire 박는 직후 — Workflow w02pv083c (4-agent MLXVLM research, 167K tokens, 6분)가 박은 full code sketch를 그대로 박은 후 swift build 시도.

error: extra argument 'images' in call (UserInput init)
error: value of type 'GenerateParameters' has no member 'topK'
error: value of type 'GenerateCompletionInfo' has no member 'stopReason'
error: cannot find 'Memory' in scope
error: mutation of captured var 'raw' / 'info' in concurrently-executing code [#SendableClosureCaptures]
error: reference to captured var 'params' (let needed)

→ 6 errors, code 안 박힘.

진짜 원인 — 시점 mismatch

Workflow research 박은 거 (Probe A):
  "REPO: ml-explore/mlx-swift-lm (formerly ml-explore/mlx-swift-examples
   Libraries/MLX{LMCommon,VLM,LLM}). MLXVLM 라이브러리는 2026년에 mlx-swift-lm에 split."
 
→ Workflow가 *최신 main 2026 source*에서 API 박았음
→ 단 우리 Package.swift는 *mlx-swift-examples 2.21+ pinned*
→ 실제 박힌 API는 *2.21 시점 stale*

4 API drift + 2 Swift 6 concurrency

Drift 1 — UserInput.init(chat:images:)

// Workflow 박은 거 (최신 2026):
let userInput = UserInput(
    chat: chat,
    images: [userImage],      // ← parameter 없음 (pinned 2.21)
    processing: ...
)
 
// Pinned 2.21 실제 (.build/checkouts/mlx-swift-examples/.../UserInput.swift):
public init(chat: [Chat.Message], processing: Processing = .init())
public init(prompt: String, images: [Image] = [], ...)
public init(messages: [Message], images: [Image] = [], ...)
// → chat: 형식엔 images 별도 X

Chat.Message.user(_:images:) 안에 박는 게 진짜 path (.build/checkouts/.../Chat.swift):

public static func user(
    _ content: String, images: [UserInput.Image] = [], videos: [UserInput.Video] = []
) -> Self

Drift 2 — GenerateParameters.topK

// Workflow 박은 거:
params.topK = 1
 
// Pinned 2.21 실제 (.build/checkouts/.../Evaluate.swift):
public var maxTokens: Int?
public var maxKVSize: Int?
public var temperature: Float = 0.6
public var topP: Float = 1.0
public var repetitionPenalty: Float?
// → topK 없음

해결: temperature 0.0 + topP 0.001로 effective greedy.

Drift 3 — GenerateCompletionInfo.stopReason

// Workflow 박은 거:
Log "stop=\(info.stopReason)"
 
// Pinned 2.21 실제:
public struct GenerateCompletionInfo: Sendable {
    public let promptTokenCount: Int
    public let generationTokenCount: Int
    public let promptTime: TimeInterval
    public let generateTime: TimeInterval
    public var promptTokensPerSecond: Double { ... }
    public var tokensPerSecond: Double { ... }
    // → stopReason 없음
}

해결: tokensPerSecond + generationTokenCount + promptTokenCount만 log.

Drift 4 — MLX cache control namespace

// Workflow 박은 거:
Memory.cacheLimit = 20 * 1024 * 1024
 
// mlx-swift 0.29 실제 (.build/checkouts/mlx-swift/Source/MLX/GPU.swift):
extension MLX.GPU {
    public static func set(cacheLimit: Int)
    public static var cacheLimit: Int { get }
}

해결: MLX.GPU.set(cacheLimit: 20 * 1024 * 1024).

추가 — Package.swift에 mlx-swift 0.29.0+ 직접 dep 추가. mlx-swift-examples는 transitive만 resolve → MLX product re-export X.

.package(url: "https://github.com/ml-explore/mlx-swift", from: "0.29.0"),
 
// + ScreenBridge target dependencies:
.product(name: "MLX", package: "mlx-swift"),

Swift 6 concurrency 1 — Sendable closure 외부 var mutation

// Workflow 박은 거:
var raw = ""
var info: GenerateCompletionInfo?
try await container.perform { context in
    for await event in stream {
        case .chunk(let s): raw += s         // ← Sendable closure 안 mutation error
        case .info(let i): info = i          // ← 동일
    }
}

container.perform은 @Sendable closure. outer var mutation 거부.

해결: closure 안 local var로 박은 후 tuple return.

let result: (raw: String, info: GenerateCompletionInfo?)
result = try await container.perform { context -> (String, GenerateCompletionInfo?) in
    var localRaw = ""
    var localInfo: GenerateCompletionInfo?
    for await event in stream { ... }
    return (localRaw, localInfo)
}
let raw = result.raw
let info = result.info

Swift 6 concurrency 2 — var 외부 capture

// Workflow 박은 거:
var params = GenerateParameters()
params.temperature = 0.0
// ...
try await container.perform { context in
    let stream = try MLXLMCommon.generate(
        parameters: params,  // ← var capture in @Sendable closure
        ...
    )
}

해결: let immutable copy로 박은 후 capture.

var paramsBuilder = GenerateParameters()
paramsBuilder.temperature = 0.0
paramsBuilder.topP = 0.001
paramsBuilder.maxTokens = 512
let params = paramsBuilder
// closure에서 params capture OK

진단 방식 — pinned source 직접 grep

# 박힌 거 확인:
grep -A 12 "init(\\s*chat:" .build/checkouts/mlx-swift-examples/Libraries/MLXLMCommon/UserInput.swift
 
# GenerateParameters 실제 field:
grep -E "public var (temperature|topP|topK|maxTokens)" \
  .build/checkouts/mlx-swift-examples/Libraries/MLXLMCommon/Evaluate.swift
 
# Chat.Message API:
grep -B 1 -A 5 "public static func user" \
  .build/checkouts/mlx-swift-examples/Libraries/MLXLMCommon/Chat.swift
 
# mlx-swift cache API:
grep -rE "cacheLimit|setCacheLimit" .build/checkouts/mlx-swift/Source/

→ 4 grep 한 번에 진짜 API 박힘. Workflow 추측보다 .build/checkouts source가 single source of truth.

결과

fix 후:
  swift build  6.62s ✓
  swift test   133/133 ✓

Phase 9.0 Week 2-3 wire complete. 실 inference는 ~2GB HF download 후 가능 (사용자 환경에서 첫 호출 시 자동).

패턴 — Workflow research 신뢰의 한계

Workflow가 박은 full code sketch는 최신 main source 기반일 때 많음 (LLM training cutoff + WebSearch가 main 표시).
Pinned version과 다름 — 특히 빠르게 변하는 lib (mlx-swift 2026 mlx-swift-lm split, MLXLMCommon API churn) 의존 시.
.build/checkouts/<pkg>/에서 pinned version source 직접 verify가 single source of truth. Workflow synthesis는 guidance로만 사용, 실제 박는 코드는 grep으로 검증.
Swift 6 strict concurrency는 swift build가 박은 후만 명확 — Workflow agent는 컴파일러 실행 X — @Sendable capture 같은 issue는 내가 박은 후 즉시 발견.

비용

Workflow research: 6분 (167K tokens, 4 agents) — 가치 큼 (전체 path + API surface + chat template + example 박힘)
API drift fix: 5분 (.build/checkouts grep 4 + code 6 edit)
Swift 6 concurrency fix: 5분 (tuple return + let params)
첫 build pass: 6.62s
총 박는 시간: ~20분 (Workflow 6분 + fix 15분)

→ 박혀있던 Workflow 박은 full code sketch + 4 API grep 결과 = 30분 안 박을 수 있음. 다음에는 코드 박기 전 grep 먼저가 더 빠를 수도. 또는 Workflow에 grep까지 시키는 것도 — 단 .build/checkouts는 swift package resolve 후만 박혀있음.

Commit

59cc0cbc6494158368619294dfd92f05e837dacd (2026-05-31)

Qwen2.5-VL-3B MLXVLM wire — research synthesis vs pinned source 6 errors

AI 버전

문제