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 +1,99 @@
+import XCTest
+@testable import MarkBase
+
+final class AudioTowerLoadTest: XCTestCase {
+    func testAudioTowerLoad() throws {
+        print("\n═══════════════════════════════════════")
+        print("      AudioTower Loading Test")
+        print("═══════════════════════════════════════\n")
+        
+        let modelDir = "/Users/accusys/MarkBaseEngine/models/E4B-MarkBase"
+        
+        print("Step 1: Load AudioConfig...")
+        let acfg = loadAudioConfig(modelDir: modelDir)
+        print("  hiddenSize: \(acfg.hiddenSize)")
+        print("  numHiddenLayers: \(acfg.numHiddenLayers)")
+        print("  subsamplingConvChannels: \(acfg.subsamplingConvChannels)")
+        
+        print("\nStep 2: Load safetensors...")
+        let reader = try SafeTensorsReader(path: modelDir + "/model.safetensors")
+        let descriptors = reader.allDescriptors()
+        let audioKeys = descriptors.filter { $0.name.hasPrefix("audio_tower.") }
+        print("  Found \(audioKeys.count) audio tensors")
+        
+        print("\nStep 3: Create engine...")
+        let engine = try MarkBaseEngine(autoCompile: true)
+        
+        print("\nStep 4: Load AudioTower...")
+        do {
+            let tower = try loadAudioTower(reader: reader, config: acfg, engine: engine)
+            print("  ✓ AudioTower loaded: \(tower.config.numHiddenLayers) layers")
+        } catch {
+            print("  ✗ AudioTower failed: \(error)")
+            throw error
+        }
+        
+        print("\n═══════════════════════════════════════")
+        print("✓ AudioTower loading test passed")
+        print("═══════════════════════════════════════\n")
+    }
+    
+    func testAudioForward() throws {
+        print("\n═══════════════════════════════════════")
+        print("      AudioTower Forward Test")
+        print("═══════════════════════════════════════\n")
+        
+        let modelDir = "/Users/accusys/MarkBaseEngine/models/E4B-MarkBase"
+        
+        print("Step 1: Load AudioTower...")
+        let engine = try MarkBaseEngine(autoCompile: true)
+        let acfg = loadAudioConfig(modelDir: modelDir)
+        let reader = try SafeTensorsReader(path: modelDir + "/model.safetensors")
+        let tower = try loadAudioTower(reader: reader, config: acfg, engine: engine)
+        print("  ✓ Loaded: \(tower.config.numHiddenLayers) layers, hidden=\(tower.config.hiddenSize)")
+        
+        print("\nStep 2: Create fake mel spectrogram...")
+        let seqLen = 98
+        let nMels = 128
+        var melFeatures: [Float] = Array(repeating: 0.1, count: seqLen * nMels)
+        for i in 0..<melFeatures.count {
+            melFeatures[i] = Float(i % 100) / 100.0
+        }
+        print("  Input shape: [\(seqLen), \(nMels)]")
+        
+        print("\nStep 3: Run forward pass...")
+        let inputBuffer = engine.device.makeBuffer(bytes: melFeatures, length: melFeatures.count * 4)!
+        let outputProjDims = tower.config.outputProjDims
+        let outputBuffer = engine.device.makeBuffer(length: seqLen / 4 * outputProjDims * 4)!
+        
+        do {
+            try tower.forward(inputBuffer: inputBuffer, seqLen: seqLen, outputBuffer: outputBuffer)
+            print("  ✓ Forward pass completed")
+        } catch {
+            print("  ✗ Forward failed: \(error)")
+            throw error
+        }
+        
+        print("\nStep 4: Check output...")
+        let ptr = outputBuffer.contents().assumingMemoryBound(to: Float.self)
+        let outputLen = seqLen / 4 * outputProjDims
+        var hasNaN = false
+        var maxVal: Float = 0
+        var minVal: Float = 0
+        for i in 0..<outputLen {
+            let v = ptr[i]
+            if v.isNaN { hasNaN = true }
+            if v > maxVal { maxVal = v }
+            if v < minVal { minVal = v }
+        }
+        print("  Output shape: [\(seqLen/4), \(outputProjDims)]")
+        print("  Range: [\(minVal), \(maxVal)]")
+        print("  Has NaN: \(hasNaN)")
+        
+        XCTAssertFalse(hasNaN, "Output should not have NaN")
+        
+        print("\n═══════════════════════════════════════")
+        print("✓ AudioTower forward test passed")
+        print("═══════════════════════════════════════\n")
+    }
+}