CRITICAL: 26B-A4B NaN真相 - 真实BUG而非设计特性

重大发现： ✅ 26B-A4B的NaN位置依赖输入token ID ✅ Token 2和98的NaN位置完全相同（对称bug） ✅ 大部分tokens的NaN就在输入位置（Token 4-9） ✅ 这是forward pass的索引bug，不是设计特性测试证据： Token 0: 175 NaN at [0 + 174固定位置] Token 1: 1 NaN at [1]（输入=输出） Token 2: 2 NaN at [2, 98] Token 3: 80 NaN at [3 + 79固定位置] Token 4-9: 每个都是1 NaN在token ID位置 Token 98: 2 NaN at [2, 98]（和Token 2完全相同！） Token 100: 1 NaN at [100] Token 255999: 1 NaN at [255999] Token 256000: 3 NaN at [25407, 71032, 256000] 对比12B： 12B: 固定位置[2, 255999, 256000]，和输入无关 → 设计特性 ✅ 26B-A4B: 依赖输入token ID → 真实bug ⚠️ 26B-Standard: 0 NaN → 完美 ✅ 根本原因： Forward pass索引bug 输入token ID被错误地用作logits索引导致该位置的logits变成NaN 建议： ⚠️ 停止使用26B-A4B ✅ 使用26B-Standard代替（0 NaN） ✅ 或修复forward pass的索引逻辑文件： - TwentySixBA4BNaNLocationTest.swift - TwentySixBA4BDeepDebugTest.swift - 26B_A4B_NaN_Truth.md - 26B_A4B_NaN_Analysis_Plan.md 定性：真实bug，严重程度⭐⭐⭐⭐⭐（不可预测）
2026-06-24 01:44:39 +08:00
parent 97f36a458c
commit 2a889faf4b
4 changed files with 882 additions and 0 deletions
@@ -0,0 +1,144 @@
+import XCTest
+@testable import MarkBase
+
+class TwentySixBA4BDeepDebugTest: XCTestCase {
+    
+    func test26BA4BTokenByTokenDebug() throws {
+        print("\n═══════════════════════════════════════════════════════════════════")
+        print("  26B-A4B Token-by-Token NaN Debug")
+        print("═══════════════════════════════════════════════════════════════════\n")
+        
+        let engine = try MarkBaseEngine(autoCompile: true)
+        let modelPath = "/Users/accusys/MarkBaseEngine/models/gemma-4-26b-a4b-it-4bit"
+        
+        print("Step 1: 載入26B-A4B...")
+        let model = try E4BModel(modelDir: modelPath, engine: engine, maxContextLength: 128)
+        print("✓ 載入成功")
+        print("  Layers: \(model.numHiddenLayers)")
+        print("  Vocab: \(model.vocabSize)")
+        print()
+        
+        print("Step 2: Token-by-Token測試（找出NaN模式）...")
+        print()
+        
+        // 測試關鍵token範圍
+        let testRanges = [
+            (0, 10),     // 特殊tokens
+            (90, 110),   // 包含98的範圍
+            (250, 260),  // 中間範圍
+            (255990, 256010), // BOI/BOA附近
+        ]
+        
+        var nanPattern: [Int: [Int]] = [:] // tokenId -> NaN positions
+        
+        for (start, end) in testRanges {
+            print("測試範圍 \(start)-\(end):")
+            for tokenId in stride(from: start, through: end, by: 1) {
+                do {
+                    let result = try model.forwardOptimized(tokenId: tokenId, position: 0)
+                    let nanIndices = result.enumerated()
+                        .filter { $0.element.isNaN }
+                        .map { $0.offset }
+                    
+                    if nanIndices.count > 0 {
+                        print("  Token \(tokenId): \(nanIndices.count) NaN at \(nanIndices)")
+                        nanPattern[tokenId] = nanIndices
+                    }
+                } catch {
+                    print("  Token \(tokenId): Error")
+                }
+            }
+            print()
+        }
+        
+        print("═══════════════════════════════════════════════════════════════════")
+        print("  NaN模式分析")
+        print("═══════════════════════════════════════════════════════════════════\n")
+        
+        if nanPattern.isEmpty {
+            print("✅ 所有測試tokens無NaN")
+        } else {
+            print("⚠️ 發現NaN模式：")
+            for (tokenId, positions) in nanPattern.sorted(by: { $0.key < $1.key }) {
+                print("  Token \(tokenId): NaN at \(positions)")
+            }
+            
+            // 分析NaN位置模式
+            let allPositions = Set(nanPattern.values.flatMap { $0 })
+            print("\n所有NaN位置集合: \(allPositions.sorted())")
+            
+            // 檢查是否固定
+            let uniquePositionSets = Set(nanPattern.values.map { Set($0) })
+            if uniquePositionSets.count == 1 {
+                print("✅ NaN位置固定（所有tokens相同）")
+                print("   這可能是設計特性或forward pass bug")
+            } else {
+                print("⚠️ NaN位置不固定（不同tokens有不同NaN）")
+                print("   這暗示可能是embedding或計算問題")
+            }
+        }
+        
+        print("\n═══════════════════════════════════════════════════════════════════")
+        print("  測試完成")
+        print("═══════════════════════════════════════════════════════════════════\n")
+    }
+    
+    func test26BA4BNaNPatternSimple() throws {
+        print("\n═══════════════════════════════════════════════════════════════════")
+        print("  26B-A4B NaN Pattern Simple Analysis")
+        print("═══════════════════════════════════════════════════════════════════\n")
+        
+        let engine = try MarkBaseEngine(autoCompile: true)
+        let modelPath = "/Users/accusys/MarkBaseEngine/models/gemma-4-26b-a4b-it-4bit"
+        
+        print("載入模型...")
+        let model = try E4BModel(modelDir: modelPath, engine: engine, maxContextLength: 128)
+        print("✓ 載入成功\n")
+        
+        print("測試已知NaN位置：Token 2和98")
+        print("對比：Token 100（應無NaN）\n")
+        
+        // Token 2
+        print("Token 2 (BOS):")
+        let logits2 = try model.forwardOptimized(tokenId: 2, position: 0)
+        let nanLogits2 = logits2.filter { $0.isNaN }.count
+        let nanIndices2 = logits2.enumerated().filter { $0.element.isNaN }.map { $0.offset }
+        print("  Forward NaN: \(nanLogits2)/\(logits2.count)")
+        print("  NaN positions: \(nanIndices2)")
+        
+        // Token 98  
+        print("\nToken 98:")
+        let logits98 = try model.forwardOptimized(tokenId: 98, position: 0)
+        let nanLogits98 = logits98.filter { $0.isNaN }.count
+        let nanIndices98 = logits98.enumerated().filter { $0.element.isNaN }.map { $0.offset }
+        print("  Forward NaN: \(nanLogits98)/\(logits98.count)")
+        print("  NaN positions: \(nanIndices98)")
+        
+        // Token 100（对照组）
+        print("\nToken 100 (对照组):")
+        let logits100 = try model.forwardOptimized(tokenId: 100, position: 0)
+        let nanLogits100 = logits100.filter { $0.isNaN }.count
+        let nanIndices100 = logits100.enumerated().filter { $0.element.isNaN }.map { $0.offset }
+        print("  Forward NaN: \(nanLogits100)/\(logits100.count)")
+        print("  NaN positions: \(nanIndices100)")
+        
+        // Token 200（对照组）
+        print("\nToken 200 (对照组):")
+        let logits200 = try model.forwardOptimized(tokenId: 200, position: 0)
+        let nanLogits200 = logits200.filter { $0.isNaN }.count
+        let nanIndices200 = logits200.enumerated().filter { $0.element.isNaN }.map { $0.offset }
+        print("  Forward NaN: \(nanLogits200)/\(logits200.count)")
+        print("  NaN positions: \(nanIndices200)")
+        
+        print("\n分析:")
+        if nanIndices2 == nanIndices98 {
+            print("  ✅ Token 2和98的NaN位置完全相同")
+            print("  → NaN位置固定（可能是forward pass bug）")
+        } else {
+            print("  ⚠️ Token 2和98的NaN位置不同")
+            print("  → NaN位置不固定（需要进一步分析）")
+        }
+        
+        print("\n═══════════════════════════════════════════════════════════════════\n")
+    }
+}