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 +1,55 @@
+import XCTest
+@testable import MarkBase
+
+final class VisionQuickTest: XCTestCase {
+    
+    func test12BVisionQuickLoad() throws {
+        print("\n═══════════════════════════════════════")
+        print("  12B Vision QUICK Load Test")
+        print("═══════════════════════════════════════\n")
+        
+        let modelDir = "/Users/accusys/.cache/huggingface/hub/models--mlx-community--gemma-4-12B-it-4bit/snapshots/73bcf09092aa277861d5a191b989b666f7f32e8f"
+        
+        print("Step 1: Create engine...")
+        let engine = try MarkBaseEngine(autoCompile: true)
+        print("  ✓ Engine created")
+        
+        print("\nStep 2: Load VisionTower12B...")
+        let loadStart = Date()
+        let visionTower = try VisionTower12B.load(modelDir: modelDir, engine: engine)
+        let loadTime = Date().timeIntervalSince(loadStart) * 1000
+        
+        print("  ✓ Vision tower loaded in \(loadTime) ms")
+        print("  Config: hiddenDim=\(visionTower.hiddenDim), patchDim=\(visionTower.patchDim)")
+        
+        print("\nStep 3: Generate synthetic patches...")
+        let numPatches = 256
+        var patches: [Float] = []
+        for _ in 0..<numPatches * visionTower.patchDim {
+            patches.append(Float.random(in: -0.5...0.5))
+        }
+        
+        print("\nStep 4: Process vision forward pass...")
+        let patchBuffer = engine.device.makeBuffer(bytes: patches, length: patches.count * 4)!
+        let outputBuffer = engine.device.makeBuffer(length: numPatches * visionTower.outputDim * 4)!
+        
+        let forwardStart = Date()
+        try visionTower.forward(patchEmbeddings: patchBuffer, numPatches: numPatches, outputBuffer: outputBuffer)
+        let forwardTime = Date().timeIntervalSince(forwardStart) * 1000
+        
+        let ptr = outputBuffer.contents().assumingMemoryBound(to: Float.self)
+        let output = Array(UnsafeBufferPointer(start: ptr, count: numPatches * visionTower.outputDim))
+        
+        print("  ✓ Forward pass completed in \(forwardTime) ms")
+        print("  Output shape: [\(numPatches), \(visionTower.outputDim)]")
+        print("  Range: [\(output.min() ?? 0), \(output.max() ?? 0)]")
+        print("  NaN count: \(output.filter { $0.isNaN }.count)")
+        
+        XCTAssertFalse(output.contains { $0.isNaN }, "Vision output should not have NaN")
+        
+        print("\n═══════════════════════════════════════")
+        print("✓ Full vision test passed - no TEXT model")
+        print("  Load: \(loadTime) ms, Forward: \(forwardTime) ms")
+        print("═══════════════════════════════════════\n")
+    }
+}