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 +1,78 @@
+import XCTest
+@testable import MarkBase
+
+final class MoEForwardTests: XCTestCase {
+    
+    func test26BA4BModelLoading() throws {
+        print("\n═══════════════════════════════════════")
+        print("      26B-A4B MoE Model Loading Test")
+        print("═══════════════════════════════════════\n")
+        
+        let modelDir = "/Users/accusys/MarkBaseEngine/models/gemma-4-26b-a4b-it-4bit"
+        
+        print("Step 1: Initialize Metal engine...")
+        let engine = try MarkBaseEngine()
+        try engine.compileSource(MetalKernels.combinedSource)
+        print("  ✓ Engine ready\n")
+        
+        print("Step 2: Load 26B-A4B MoE model...")
+        
+        do {
+            let start = Date()
+            let model = try E4BModel(modelDir: modelDir, engine: engine, maxContextLength: 512)
+            let loadTime = Date().timeIntervalSince(start)
+            
+            print("  ✓ Model loaded successfully in \(String(format: "%.3f", loadTime))s\n")
+            
+            print("Model parameters:")
+            print("  Layers: \(model.numHiddenLayers)")
+            print("  Hidden size: \(model.hiddenSize)")
+            print("  Vocab size: \(model.vocabSize)")
+            
+            // Verify 26B-A4B parameters
+            XCTAssertEqual(model.numHiddenLayers, 30, "Should have 30 layers")
+            XCTAssertEqual(model.hiddenSize, 2816, "Should have hidden_size=2816")
+            XCTAssertEqual(model.vocabSize, 262144, "Should have vocab_size=262144")
+            
+            print("\n═══════════════════════════════════════")
+            print("✓ MoE model loading test passed")
+            print("═══════════════════════════════════════\n")
+            
+        } catch {
+            print("  ✗ Model loading failed: \(error)\n")
+            throw error
+        }
+    }
+    
+    func test26BA4BSimpleGeneration() throws {
+        print("\n═══════════════════════════════════════")
+        print("      26B-A4B MoE Simple Generation")
+        print("═══════════════════════════════════════\n")
+        
+        let modelDir = "/Users/accusys/MarkBaseEngine/models/gemma-4-26b-a4b-it-4bit"
+        let engine = try MarkBaseEngine()
+        try engine.compileSource(MetalKernels.combinedSource)
+        
+        let model = try E4BModel(modelDir: modelDir, engine: engine, maxContextLength: 128)
+        let tokenizer = try TokenizerFactory.load(modelDir: modelDir)
+        let generator = StreamingGenerator(model: model, tokenizer: tokenizer, engine: engine)
+        
+        print("Testing simple generation (5 tokens)...")
+        let prompt = "Hello"
+        let config = GenerationConfig(maxTokens: 5, temperature: Float(0.7))
+        
+        let start = Date()
+        let response = try generator.generateComplete(prompt: prompt, config: config)
+        let genTime = Date().timeIntervalSince(start)
+        
+        print("  ✓ Generated in \(String(format: "%.3f", genTime))s")
+        print("  Response: \"\(response)\"")
+        print("  Speed: \(String(format: "%.2f", 5.0 / genTime)) tok/s")
+        
+        XCTAssertGreaterThan(response.count, 0, "Should generate some text")
+        
+        print("\n═══════════════════════════════════════")
+        print("✓ Simple generation test passed")
+        print("═══════════════════════════════════════\n")
+    }
+}