tim/kangkang
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(AI): MNN 文本模型升到 Qwen3.5-4B(taobao-mnn 预转换)

Browse Source 
现场机 iPhone 17(A19/SME2)内存与加速均可承载 4B,质量优于 2B。

- ModelKind.mnnLLM rawValue → "Qwen3.5-4B-MNN",displayName → Qwen3.5-4B (MNN/SME2)
- ModelManifest:7 个运行时文件(llm.mnn.weight ~2.45GB + 拆分的
  visual.mnn.weight 188MB),总计 2,836,770,850 bytes(~2.64GiB)
- ModelManifestTests:文件数 7 / 总字节 / URL 更新到 Qwen3.5-4B-MNN
- CLAUDE.md §2:MNN 主模型记为 Qwen3.5-4B,MLX 兜底仍 2B

模拟器 ModelManifestTests TEST SUCCEEDED。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
link2026
2026-06-08 20:28:14 +08:00
parent 39b1521f00
commit cbacd9461a
4 changed files with 15 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
CLAUDE.md
View File

@@ -24,7 +24,7 @@
| 图表 | Swift Charts | iOS 16+ 原生 | | 图表 | Swift Charts | iOS 16+ 原生 |
| **AI 运行时(主)** | **MNN (alibaba) + Arm SME2 + CPU** | 挑战赛考核点:Qwen + MNN + SME2 端侧 CPU 推理。device-only(xcframework 见 `scripts/build-mnn-xcframework.sh`),A19/iPhone17 启用 SME2、A17 回退 NEON。经 `MNNLLMBridge`(ObjC++)→ `MNNBackend` | | **AI 运行时(主)** | **MNN (alibaba) + Arm SME2 + CPU** | 挑战赛考核点:Qwen + MNN + SME2 端侧 CPU 推理。device-only(xcframework 见 `scripts/build-mnn-xcframework.sh`),A19/iPhone17 启用 SME2、A17 回退 NEON。经 `MNNLLMBridge`(ObjC++)→ `MNNBackend` |
| **AI 运行时(兜底)** | **MLX Swift (Apple 官方,Metal GPU)** | 双后端:`InferenceEngine` 切换,模拟器/兜底用 MLX。不要建议 Core ML / llama.cpp / Ollama | | **AI 运行时(兜底)** | **MLX Swift (Apple 官方,Metal GPU)** | 双后端:`InferenceEngine` 切换,模拟器/兜底用 MLX。不要建议 Core ML / llama.cpp / Ollama |
| LLM | Qwen3.5-2B 4bit(MNN 格式 + MLX `mlx-community/Qwen3.5-2B-4bit`) | 文本生成、关键词抽取、趋势解读 | | LLM | MNN 主:Qwen3.5-4B(`taobao-mnn/Qwen3.5-4B-MNN`,~2.64GiB);MLX 兜底:Qwen3.5-2B-4bit | 文本生成、关键词抽取、趋势解读 |
| VL | Qwen3-VL-4B-Instruct 4bit (MLX `mlx-community/Qwen3-VL-4B-Instruct-4bit`) | 拍照→结构化指标。MNN VL 需 OMNI 构建,暂走 MLX | | VL | Qwen3-VL-4B-Instruct 4bit (MLX `mlx-community/Qwen3-VL-4B-Instruct-4bit`) | 拍照→结构化指标。MNN VL 需 OMNI 构建,暂走 MLX |
| 文档扫描 | VisionKit `VNDocumentCameraView` | 不要自己写透视校正 | | 文档扫描 | VisionKit `VNDocumentCameraView` | 不要自己写透视校正 |
| Face ID | LocalAuthentication | | | Face ID | LocalAuthentication | |

13
康康/AI/ModelManifest.swift
View File

@@ -60,17 +60,18 @@ nonisolated enum ModelManifest {
ModelFile(path: "video_preprocessor_config.json", bytes: 817), ModelFile(path: "video_preprocessor_config.json", bytes: 817),
] ]
case .mnnLLM: case .mnnLLM:
// taobao-mnn/Qwen3.5-2B-MNN 预转换 MNN 格式(HF API 实测,2026-06)。 // taobao-mnn/Qwen3.5-4B-MNN 预转换 MNN 格式(HF API 实测,2026-06)。
// 运行时必需:config.json(MNN llm 配置)+ llm_config.json(超参)+ llm.mnn(图) // 运行时必需:config.json(MNN llm 配置)+ llm_config.json(超参)+ llm.mnn(图)
// + llm.mnn.weight(量化权重 ~1.1GB)+ tokenizer.txt + visual.mnn(多模态,文本路径不用但配置含 mllm)。 // + llm.mnn.weight(量化权重 ~2.45GB)+ tokenizer.txt + visual.mnn/visual.mnn.weight(多模态,
// 排除 README/.gitattributes 与可读 dump(llm.mnn.json / export_args.json)。 // 文本路径不用但配置含 mllm,带上避免缺文件)。排除 README/.gitattributes 与可读 dump。
return [ return [
ModelFile(path: "config.json", bytes: 652), ModelFile(path: "config.json", bytes: 652),
ModelFile(path: "llm_config.json", bytes: 8_692), ModelFile(path: "llm_config.json", bytes: 8_693),
ModelFile(path: "llm.mnn", bytes: 2_148_136), ModelFile(path: "llm.mnn", bytes: 3_651_096),
ModelFile(path: "llm.mnn.weight", bytes: 1_176_647_702), ModelFile(path: "llm.mnn.weight", bytes: 2_629_387_626),
ModelFile(path: "tokenizer.txt", bytes: 6_465_727), ModelFile(path: "tokenizer.txt", bytes: 6_465_727),
ModelFile(path: "visual.mnn", bytes: 488_096), ModelFile(path: "visual.mnn", bytes: 488_096),
ModelFile(path: "visual.mnn.weight", bytes: 196_768_960),
] ]
} }
} }

6
康康/AI/ModelStore.swift
View File

@@ -4,16 +4,16 @@ nonisolated enum ModelKind: String, CaseIterable {
/// 也是沙盒 Models/ 下的子目录名 / CDN 路径段。 /// 也是沙盒 Models/ 下的子目录名 / CDN 路径段。
/// - llm:MLX(GPU)文本兜底,Qwen3.5-2B(多模态权重,走 qwen3_5 文本路径)。 /// - llm:MLX(GPU)文本兜底,Qwen3.5-2B(多模态权重,走 qwen3_5 文本路径)。
/// - vl :MLX(GPU)拍照识别,Qwen3-VL-4B。 /// - vl :MLX(GPU)拍照识别,Qwen3-VL-4B。
/// - mnnLLM:MNN(CPU/SME2,挑战赛考核路径)文本,Qwen3.5-2B 预转换 MNN 格式(taobao-mnn)。 /// - mnnLLM:MNN(CPU/SME2,挑战赛考核路径)文本,Qwen3.5-4B 预转换 MNN 格式(taobao-mnn)。
case llm = "Qwen3.5-2B-4bit" case llm = "Qwen3.5-2B-4bit"
case vl = "Qwen3-VL-4B-Instruct-4bit" case vl = "Qwen3-VL-4B-Instruct-4bit"
case mnnLLM = "Qwen3.5-2B-MNN" case mnnLLM = "Qwen3.5-4B-MNN"
var displayName: String { var displayName: String {
switch self { switch self {
case .llm: return "Qwen3.5-2B (MLX)" case .llm: return "Qwen3.5-2B (MLX)"
case .vl: return "Qwen3-VL-4B" case .vl: return "Qwen3-VL-4B"
case .mnnLLM: return "Qwen3.5-2B (MNN/SME2)" case .mnnLLM: return "Qwen3.5-4B (MNN/SME2)"
} }
} }

8
康康Tests/ModelManifestTests.swift
View File

@@ -20,12 +20,12 @@ struct ModelManifestTests {
#expect(ModelManifest.totalBytes(for: .vl) == 3_109_729_929) #expect(ModelManifest.totalBytes(for: .vl) == 3_109_729_929)
} }
@Test func mnnHasSixFunctionalFiles() { @Test func mnnHasSevenFunctionalFiles() {
#expect(ModelManifest.files(for: .mnnLLM).count == 6) #expect(ModelManifest.files(for: .mnnLLM).count == 7)
} }
@Test func mnnTotalBytesMatchesManifest() { @Test func mnnTotalBytesMatchesManifest() {
#expect(ModelManifest.totalBytes(for: .mnnLLM) == 1_185_759_005) #expect(ModelManifest.totalBytes(for: .mnnLLM) == 2_836_770_850)
} }
@Test func mnnHasEssentialRuntimeFiles() { @Test func mnnHasEssentialRuntimeFiles() {
@@ -39,7 +39,7 @@ struct ModelManifestTests {
@Test func mnnFileURLUsesRepoPath() { @Test func mnnFileURLUsesRepoPath() {
let file = ModelFile(path: "config.json", bytes: 652) let file = ModelFile(path: "config.json", bytes: 652)
let url = ModelManifest.fileURL(for: .mnnLLM, file: file) let url = ModelManifest.fileURL(for: .mnnLLM, file: file)
#expect(url.absoluteString == "https://file.myv0.com/Qwen3.5-2B-MNN/config.json") #expect(url.absoluteString == "https://file.myv0.com/Qwen3.5-4B-MNN/config.json")
} }
@Test func excludesReadmeAndGitattributes() { @Test func excludesReadmeAndGitattributes() {