Initial commit: E4B-MarkBase model integration with passing tests

- E4B-MarkBase model (42 layers, 4.4GB) loaded successfully - All Phase 1-6 tests passed (model loading, forward pass, vision/audio towers, token generation, performance) - All stress tests passed (5/5 in 127.6s) - Concurrent inference - Memory stress (67.5 tok/s, 0 NaN) - Continuous generation - Batch processing - Long-running stability - Swift Metal inference engine with multimodal support
2026-06-23 18:12:35 +08:00
commit ac75faa0cc
301 changed files with 63426 additions and 0 deletions
@@ -0,0 +1,47 @@
+import XCTest
+@testable import MarkBase
+
+class Model31BForwardTest: XCTestCase {
+    
+    func testModel31BForwardNaN() throws {
+        print("\n═══════════════════════════════════════════════════════════════════")
+        print("  31B Forward Pass NaN Test")
+        print("═══════════════════════════════════════════════════════════════════\n")
+        
+        let modelPath = "/Users/accusys/MarkBaseEngine/models/gemma-4-31b-it-4bit"
+        
+        guard FileManager.default.fileExists(atPath: modelPath) else {
+            print("⚠ Model not found")
+            return
+        }
+        
+        let engine = try MarkBaseEngine(autoCompile: true)
+        
+        print("Loading 31B...")
+        let model = try E4BModel(modelDir: modelPath, engine: engine, maxContextLength: 128)
+        print("✓ Model loaded (Layers: \(model.numHiddenLayers), Hidden: \(model.hiddenSize))")
+        
+        // Test multiple tokenIds (like 26B-A4B)
+        print("\nTesting tokenIds 0-10...")
+        var nanPattern: [Int: Int] = [:]
+        for testTokenId in 0..<10 {
+            let result = try model.forwardOptimized(tokenId: testTokenId, position: 0)
+            let nanCount = result.filter { $0.isNaN }.count
+            if nanCount > 0 {
+                nanPattern[testTokenId] = nanCount
+            }
+        }
+        
+        if nanPattern.isEmpty {
+            print("  ✓ No NaN for tokenIds 0-10")
+            print("  31B may handle scales differently (no NaN despite wrong scales)")
+        } else {
+            print("  ⚠ NaN detected:")
+            for (tokenId, count) in nanPattern.sorted(by: { $0.key < $1.key }) {
+                print("    tokenId=\(tokenId): NaN=\(count)")
+            }
+        }
+        
+        print("\n═══════════════════════════════════════════════════════════════════")
+    }
+}