v2: EmbeddingGemma - single cmdBuf fix, needs Metal kernel compilation

2026-07-06 11:37:10 +08:00
parent dbec6b20ea
commit 5e060c7aea
2 changed files with 157 additions and 147 deletions
@@ -0,0 +1,29 @@
+import XCTest
+@testable import MarkBase
+
+final class EmbeddingGemmaTest: XCTestCase {
+    func testLoadAndEmbed() throws {
+        let modelDir = "/Users/accusys/MarkBaseEngine/models/embeddinggemma-300m"
+        guard FileManager.default.fileExists(atPath: modelDir + "/model.safetensors") else {
+            XCTFail("Model not found"); return
+        }
+        let engine = try MarkBaseEngine(autoCompile: true)
+        let model = try EmbeddingGemmaModel(modelDir: modelDir, engine: engine)
+        XCTAssertEqual(model.config.numHiddenLayers, 24)
+        XCTAssertEqual(model.config.hiddenSize, 768)
+        
+        let emb = try model.embed(text: "Hello world")
+        XCTAssertEqual(emb.count, 768)
+        
+        // Check L2 norm is 1.0
+        var norm: Float = 0
+        for v in emb { norm += v * v }
+        norm = sqrt(norm)
+        XCTAssertEqual(norm, 1.0, accuracy: 0.001, "L2 norm should be 1.0")
+        
+        // Check no NaN
+        XCTAssertFalse(emb.contains { $0.isNaN }, "No NaN values")
+        
+        print("✓ Embedding dim=\(emb.count), norm=\(String(format: "%.4f", norm))")
+    }
+}