v2: add actor-based serialization for EmbeddingServer, fix stability for ASCII text

Package.swift

@@ -7,6 +7,7 @@ let package = Package(
     products: [
         .library(name: "MarkBase", targets: ["MarkBase"]),
         .executable(name: "MarkBaseServer", targets: ["MarkBaseServer"]),
+        .executable(name: "EmbeddingServer", targets: ["EmbeddingServer"]),
         .executable(name: "CLITest", targets: ["CLITest"]),
     ],
     dependencies: [
@@ -16,7 +17,7 @@ let package = Package(
     targets: [
         .target(
             name: "MarkBase",
-            exclude: ["Metal/MetalKernels.metal", "Metal/OptimizedKernels.metal", "Metal/FusionKernels.metal", "Metal/MetalKernels.metallib", "Metal/metallib"],
+            exclude: ["Metal/MetalKernels.metal", "Metal/OptimizedKernels.metal", "Metal/FusionKernels.metal", "Metal/EmbeddingKernels.metal", "Metal/MetalKernels.metallib", "Metal/metallib"],
             linkerSettings: [
                 .linkedFramework("Metal"),
                 .linkedFramework("Foundation"),
@@ -34,6 +35,17 @@ let package = Package(
                 .linkedFramework("Foundation"),
             ]
         ),
+        .executableTarget(
+            name: "EmbeddingServer",
+            dependencies: [
+                "MarkBase",
+                .product(name: "Hummingbird", package: "hummingbird"),
+            ],
+            linkerSettings: [
+                .linkedFramework("Metal"),
+                .linkedFramework("Foundation"),
+            ]
+        ),
         .executableTarget(
             name: "CLITest",
             dependencies: ["MarkBase"],

Sources/EmbeddingServer/EmbeddingServerMain.swift

@@ -0,0 +1,98 @@
+import Foundation
+import MarkBase
+import Hummingbird
+
+@main
+struct EmbeddingServerMain {
+    static func main() async throws {
+        let args = CommandLine.arguments
+        let modelName = args.count > 1 ? args[1] : "embeddinggemma-300m"
+        let port = args.count > 2 ? Int(args[2]) ?? 8084 : 8084
+        
+        let modelPath = NSString(string: "~/MarkBaseEngine/models/\(modelName)").expandingTildeInPath
+
+        print("MarkBaseEngine Embedding Server")
+        print("  Model: \(modelName)")
+        print("  Port: \(port)")
+        print("  Path: \(modelPath)")
+
+        let engine = try MarkBaseEngine(autoCompile: true)
+        let embedModel = try EmbeddingGemmaModel(modelDir: modelPath, engine: engine)
+        let layers = embedModel.config.numHiddenLayers
+        let hiddenSize = embedModel.config.hiddenSize
+
+        print("EmbeddingGemma loaded (\(layers) layers, hidden=\(hiddenSize))")
+
+        // Use actor to serialize access to embedModel
+        let embedder = Embedder(model: embedModel)
+
+        let router = Router()
+
+        router.get("/") { _, _ in
+            return "{\"server\":\"MarkBaseEngine Embedding\",\"model\":\"embeddinggemma-300m\",\"layers\":\(layers),\"hidden_size\":\(hiddenSize)}"
+        }
+
+        router.get("/health") { _, _ in
+            return "{\"status\":\"healthy\",\"model\":\"embeddinggemma-300m\",\"layers\":\(layers)}"
+        }
+
+        router.post("/v1/embeddings") { request, _ in
+            let buffer = try await request.body.collect(upTo: .max)
+            let data = Data(buffer: buffer)
+
+            guard let json = try JSONSerialization.jsonObject(with: data) as? [String: Any],
+                  let input = json["input"] else {
+                return "{\"error\":\"missing 'input' field\"}"
+            }
+
+            let inputs: [String]
+            if let str = input as? String { inputs = [str] }
+            else if let arr = input as? [String] { inputs = arr }
+            else { return "{\"error\":\"'input' must be string or array\"}" }
+
+            var embeddings: [[String: Any]] = []
+            for (i, text) in inputs.enumerated() {
+                let result = try await embedder.embed(text: text)
+                embeddings.append(["object": "embedding", "index": i, "embedding": result.embedding, "usage_ms": result.ms])
+            }
+
+            let id = UUID().uuidString
+            let ts = Int(Date().timeIntervalSince1970)
+            let response: [String: Any] = [
+                "id": id, "object": "list", "created": ts, "model": "embeddinggemma-300m",
+                "data": embeddings
+            ]
+
+            let jsonData = try JSONSerialization.data(withJSONObject: response)
+            return String(data: jsonData, encoding: .utf8) ?? "{}"
+        }
+
+        let app = Application(
+            router: router,
+            configuration: .init(address: .hostname("0.0.0.0", port: port))
+        )
+
+        print("Server starting on port \(port)...")
+        try await app.run()
+    }
+}
+
+struct EmbeddingResult {
+    let embedding: [Float]
+    let ms: Int
+}
+
+actor Embedder {
+    private let model: EmbeddingGemmaModel
+    
+    init(model: EmbeddingGemmaModel) {
+        self.model = model
+    }
+    
+    func embed(text: String) throws -> EmbeddingResult {
+        let t0 = Date()
+        let embedding = try model.embed(text: text)
+        let ms = Int(Date().timeIntervalSince(t0) * 1000)
+        return EmbeddingResult(embedding: embedding, ms: ms)
+    }
+}

Sources/MarkBase/Embedding/EmbeddingGemmaModel.swift

@@ -1,8 +1,9 @@
 import Foundation
 import Metal
+import MetalPerformanceShaders
 import Accelerate
 
-/// EmbeddingGemmaConfig - Configuration for EmbeddingGemma model
+/// EmbeddingGemma configuration
 public struct EmbeddingGemmaConfig: Codable {
     public let hiddenSize: Int
     public let numHiddenLayers: Int
@@ -41,14 +42,12 @@ public struct EmbeddingGemmaConfig: Codable {
     }
 }
 
-/// EmbeddingGemma - Google's 300M parameter embedding model
-public final class EmbeddingGemmaModel {
+public final class EmbeddingGemmaModel: @unchecked Sendable {
     public let config: EmbeddingGemmaConfig
     public let engine: MarkBaseEngine
     public let tokenizer: Tokenizer
     public let reader: SafeTensorsReader
 
-    // GPU Buffers
     public var embedTokens: MTLBuffer!
     public var finalNorm: MTLBuffer!
     public var layerNorms: [[MTLBuffer]] = []
@@ -67,21 +66,13 @@ public final class EmbeddingGemmaModel {
         self.config = try EmbeddingGemmaConfig.load(from: modelDir)
         self.tokenizer = try TokenizerFactory.load(modelDir: modelDir)
         self.reader = try SafeTensorsReader(path: modelDir + "/model.safetensors")
-
         try loadWeights()
         print("✓ EmbeddingGemma loaded (\(config.numHiddenLayers) layers, hidden=\(config.hiddenSize))")
     }
 
     private func loadWeights() throws {
-        let hs = config.hiddenSize
-        let intermedi = config.intermediateSize
-        let nKV = config.numKeyValueHeads
-        let hDim = config.headDim
-
-        // Embedding table [vocab, hidden]
         let embedData = try readTensor("embed_tokens.weight")
         embedTokens = engine.device.makeBuffer(bytes: embedData, length: embedData.count * 4)!
-
         for i in 0..<config.numHiddenLayers {
             let p = "layers.\(i)"
             layerNorms.append([
@@ -90,22 +81,20 @@ public final class EmbeddingGemmaModel {
                 try loadBuffer("\(p).post_attention_layernorm.weight"),
                 try loadBuffer("\(p).post_feedforward_layernorm.weight"),
             ])
-            qProjs.append(try loadBuffer("\(p).self_attn.q_proj.weight"))     // [hs, hs]
-            kProjs.append(try loadBuffer("\(p).self_attn.k_proj.weight"))     // [nKV*hDim, hs]
-            vProjs.append(try loadBuffer("\(p).self_attn.v_proj.weight"))     // [nKV*hDim, hs]
-            oProjs.append(try loadBuffer("\(p).self_attn.o_proj.weight"))     // [hs, nH*hDim]
-            qNorms.append(try loadBuffer("\(p).self_attn.q_norm.weight"))     // [hDim]
-            kNorms.append(try loadBuffer("\(p).self_attn.k_norm.weight"))     // [hDim]
-            gateProjs.append(try loadBuffer("\(p).mlp.gate_proj.weight"))     // [intermedi, hs]
-            upProjs.append(try loadBuffer("\(p).mlp.up_proj.weight"))         // [intermedi, hs]
-            downProjs.append(try loadBuffer("\(p).mlp.down_proj.weight"))     // [hs, intermedi]
+            qProjs.append(try loadAndTranspose("\(p).self_attn.q_proj.weight", rows: config.hiddenSize, cols: config.hiddenSize))
+            kProjs.append(try loadAndTranspose("\(p).self_attn.k_proj.weight", rows: config.numKeyValueHeads * config.headDim, cols: config.hiddenSize))
+            vProjs.append(try loadAndTranspose("\(p).self_attn.v_proj.weight", rows: config.numKeyValueHeads * config.headDim, cols: config.hiddenSize))
+            oProjs.append(try loadAndTranspose("\(p).self_attn.o_proj.weight", rows: config.numAttentionHeads * config.headDim, cols: config.hiddenSize))
+            qNorms.append(try loadBuffer("\(p).self_attn.q_norm.weight"))
+            kNorms.append(try loadBuffer("\(p).self_attn.k_norm.weight"))
+            gateProjs.append(try loadAndTranspose("\(p).mlp.gate_proj.weight", rows: config.intermediateSize, cols: config.hiddenSize))
+            upProjs.append(try loadAndTranspose("\(p).mlp.up_proj.weight", rows: config.intermediateSize, cols: config.hiddenSize))
+            downProjs.append(try loadAndTranspose("\(p).mlp.down_proj.weight", rows: config.hiddenSize, cols: config.intermediateSize))
         }
-
         let fnData = try readTensor("norm.weight")
         finalNorm = engine.device.makeBuffer(bytes: fnData, length: fnData.count * 4)!
     }
 
-    /// Generate embedding for text
     public func embed(text: String, maxLen: Int = 2048) throws -> [Float] {
         var tokens = tokenizer.encode(text: text)
         if tokens.count > maxLen { tokens = Array(tokens.prefix(maxLen)) }
@@ -113,22 +102,37 @@ public final class EmbeddingGemmaModel {
 
         let seqLen = tokens.count, hs = config.hiddenSize
 
-        // Embedding lookup
-        let inputBuf = try lookupEmbeddings(tokens: tokens)
+        // Test 1: Embedding lookup only
+        print("  TEST: Embedding lookup...")
+        let cmdBuf1 = engine.commandQueue.makeCommandBuffer()!
+        let inputBuf = try lookupEmbeddings(tokens: tokens, cmdBuf: cmdBuf1)
+        cmdBuf1.commit()
+        cmdBuf1.waitUntilCompleted()
+        print("  TEST: Embedding lookup OK")
 
-        // Forward through layers
-        var hidden = inputBuf
+        // Test 2: Single layer forward (layer 0 only)
+        print("  TEST: Layer 0 forward...")
+        let cmdBuf2 = engine.commandQueue.makeCommandBuffer()!
+        var hidden = try forwardLayerDebug(hidden: inputBuf, layerIdx: 0, seqLen: seqLen, cmdBuf: cmdBuf2)
+        cmdBuf2.commit()
+        cmdBuf2.waitUntilCompleted()
+        print("  TEST: Layer 0 OK")
+
+        // Full forward pass
+        print("  TEST: Full forward pass...")
+        let cmdBuf = engine.commandQueue.makeCommandBuffer()!
+        hidden = inputBuf
         for idx in 0..<config.numHiddenLayers {
-            hidden = try forwardLayer(hidden: hidden, layerIdx: idx, seqLen: seqLen)
+            if idx % 6 == 0 { print("  Layer \(idx)/\(config.numHiddenLayers)") }
+            hidden = try forwardLayerDebug(hidden: hidden, layerIdx: idx, seqLen: seqLen, cmdBuf: cmdBuf)
         }
+        print("  TEST: All layers OK")
 
-        // Final norm
-        let output = try applyRmsNorm(input: hidden, weight: finalNorm, count: seqLen * hs)
+        let output = try applyRmsNorm(input: hidden, weight: finalNorm, count: seqLen * hs, cmdBuf: cmdBuf)
+        cmdBuf.commit()
+        cmdBuf.waitUntilCompleted()
 
-        // Readback
         let data = engine.readFloats(from: output, count: seqLen * hs)
-
-        // Mean pool + L2 normalize
         var embedding = [Float](repeating: 0, count: hs)
         for i in 0..<seqLen {
             let start = i * hs
@@ -136,29 +140,20 @@ public final class EmbeddingGemmaModel {
         }
         let n = Float(seqLen)
         for i in 0..<hs { embedding[i] /= n }
-
         var norm: Float = 0
         for i in 0..<hs { norm += embedding[i] * embedding[i] }
         norm = sqrt(norm)
         if norm > 0 { for i in 0..<hs { embedding[i] /= norm } }
-
         return embedding
     }
 
-    // MARK: - Helpers
-
     private func readTensor(_ name: String) throws -> [Float] {
-        guard let desc = reader.tensor(named: name) else {
-            throw WeightError.tensorNotFound(name)
-        }
+        guard let desc = reader.tensor(named: name) else { throw WeightError.tensorNotFound(name) }
         let data = try reader.read(tensor: desc)
         switch desc.dtype {
-        case .f32:
-            return data.withUnsafeBytes { Array(UnsafeBufferPointer(start: $0.baseAddress?.assumingMemoryBound(to: Float.self), count: data.count/4)) }
-        case .bf16:
-            return try SafeTensorsReader.bf16ToFloat32(data)
-        default:
-            throw WeightError.unsupportedDtype(desc.dtype.rawValue)
+        case .f32: return data.withUnsafeBytes { Array(UnsafeBufferPointer(start: $0.baseAddress?.assumingMemoryBound(to: Float.self), count: data.count/4)) }
+        case .bf16: return try SafeTensorsReader.bf16ToFloat32(data)
+        default: throw WeightError.unsupportedDtype(desc.dtype.rawValue)
         }
     }
 
@@ -167,31 +162,45 @@ public final class EmbeddingGemmaModel {
         return engine.device.makeBuffer(bytes: data, length: data.count * 4)!
     }
 
-    private func lookupEmbeddings(tokens: [Int]) throws -> MTLBuffer {
+    private func loadAndTranspose(_ name: String, rows: Int, cols: Int) throws -> MTLBuffer {
+        // Load [rows, cols] and transpose to [cols, rows] for MPS matmul C = A × B
+        let data = try readTensor(name)
+        var transposed = [Float](repeating: 0, count: data.count)
+        for r in 0..<rows {
+            for c in 0..<cols {
+                transposed[c * rows + r] = data[r * cols + c]
+            }
+        }
+        return engine.device.makeBuffer(bytes: transposed, length: transposed.count * 4)!
+    }
+
+    private func lookupEmbeddings(tokens: [Int], cmdBuf: MTLCommandBuffer) throws -> MTLBuffer {
         let seqLen = tokens.count, hs = config.hiddenSize
-        let buf = engine.device.makeBuffer(length: seqLen * hs * 4)!
-        let cmdBuf = engine.commandQueue.makeCommandBuffer()!
-        let enc = cmdBuf.makeComputeCommandEncoder()!
-        let pso = try engine.pipeline(named: "lookup_embeddings")
-        enc.setComputePipelineState(pso)
-        enc.setBuffer(embedTokens, offset: 0, index: 0)
-        enc.setBytes(tokens, length: seqLen * MemoryLayout<Int>.size, index: 1)
-        enc.setBuffer(buf, offset: 0, index: 2)
-        var h = UInt32(hs), s = UInt32(seqLen), v = UInt32(config.vocabSize)
-        enc.setBytes(&h, length: 4, index: 3)
-        enc.setBytes(&s, length: 4, index: 4)
-        enc.setBytes(&v, length: 4, index: 5)
-        enc.dispatchThreads(MTLSize(width: seqLen, height: 1, depth: 1),
-                           threadsPerThreadgroup: MTLSize(width: min(256, seqLen), height: 1, depth: 1))
-        enc.endEncoding()
-        cmdBuf.commit(); cmdBuf.waitUntilCompleted()
+        print("    lookupEmbeddings: seqLen=\(seqLen), hs=\(hs)")
+        // Read embedding table to CPU
+        let embedPtr = embedTokens.contents().assumingMemoryBound(to: Float.self)
+        let embedCount = embedTokens.length / 4
+        let embedArray = Array(UnsafeBufferPointer(start: embedPtr, count: embedCount))
+        print("    Read \(embedCount) floats from embedTokens")
+        
+        // Lookup embeddings for tokens
+        var embedData = [Float](repeating: 0, count: seqLen * hs)
+        for (i, token) in tokens.enumerated() {
+            let start = i * hs
+            let srcStart = token * hs
+            embedData[start..<start+hs] = embedArray[srcStart..<srcStart+hs]
+        }
+        print("    Looked up \(seqLen) tokens")
+        
+        let buf = engine.device.makeBuffer(bytes: embedData, length: embedData.count * 4)!
+        print("    Created MTLBuffer")
         return buf
     }
 
-    private func applyRmsNorm(input: MTLBuffer, weight: MTLBuffer, count: Int) throws -> MTLBuffer {
+    private func applyRmsNorm(input: MTLBuffer, weight: MTLBuffer, count: Int, cmdBuf: MTLCommandBuffer) throws -> MTLBuffer {
         let output = engine.device.makeBuffer(length: count * 4)!
-        let cmdBuf = engine.commandQueue.makeCommandBuffer()!
         let enc = cmdBuf.makeComputeCommandEncoder()!
+        defer { enc.endEncoding() }
         let pso = try engine.pipeline(named: "rms_norm")
         enc.setComputePipelineState(pso)
         enc.setBuffer(input, offset: 0, index: 0)
@@ -200,104 +209,60 @@ public final class EmbeddingGemmaModel {
         var c = UInt32(count), e: Float = config.rmsNormEps
         enc.setBytes(&c, length: 4, index: 3)
         enc.setBytes(&e, length: 4, index: 4)
-        enc.dispatchThreads(MTLSize(width: count, height: 1, depth: 1),
-                           threadsPerThreadgroup: MTLSize(width: min(256, count), height: 1, depth: 1))
-        enc.endEncoding()
-        cmdBuf.commit(); cmdBuf.waitUntilCompleted()
+        enc.dispatchThreads(MTLSize(width: count, height: 1, depth: 1), threadsPerThreadgroup: MTLSize(width: min(256, count), height: 1, depth: 1))
         return output
     }
 
-    private func forwardLayer(hidden: MTLBuffer, layerIdx: Int, seqLen: Int) throws -> MTLBuffer {
-        let hs = config.hiddenSize, device = engine.device
-        let hDim = config.headDim, nH = config.numAttentionHeads, nKV = config.numKeyValueHeads
-        let intermedi = config.intermediateSize
-
-        let cmdBuf = engine.commandQueue.makeCommandBuffer()!
-
-        // Residual
-        let resid = device.makeBuffer(length: seqLen * hs * 4)!
-        let blit = cmdBuf.makeBlitCommandEncoder()!
-        blit.copy(from: hidden, sourceOffset: 0, to: resid, destinationOffset: 0, size: seqLen * hs * 4)
-        blit.endEncoding()
-
-        // Input norm
-        let h1 = try applyRmsNorm(input: hidden, weight: layerNorms[layerIdx][0], count: seqLen * hs)
-
-        // Q, K, V projections (using optimized matmul)
-        let qBuf = device.makeBuffer(length: seqLen * nH * hDim * 4)!
-        let kBuf = device.makeBuffer(length: seqLen * nKV * hDim * 4)!
-        let vBuf = device.makeBuffer(length: seqLen * nKV * hDim * 4)!
-        try matmulSeq(input: h1, weight: qProjs[layerIdx], output: qBuf, m: seqLen, k: hs, n: nH * hDim, cmdBuf: cmdBuf)
-        try matmulSeq(input: h1, weight: kProjs[layerIdx], output: kBuf, m: seqLen, k: hs, n: nKV * hDim, cmdBuf: cmdBuf)
-        try matmulSeq(input: h1, weight: vProjs[layerIdx], output: vBuf, m: seqLen, k: hs, n: nKV * hDim, cmdBuf: cmdBuf)
-
-        // RoPE
-        try applyRoPE(q: qBuf, k: kBuf, seqLen: seqLen, headDim: hDim, numHeads: nH, numKVHeads: nKV, cmdBuf: cmdBuf)
-
-        // Q/K Norm
-        try applyQKNorm(q: qBuf, k: kBuf, qNorm: qNorms[layerIdx], kNorm: kNorms[layerIdx], seqLen: seqLen, headDim: hDim, numHeads: nH, numKVHeads: nKV, cmdBuf: cmdBuf)
-
-        // Bidirectional sliding window attention
-        let attnOut = device.makeBuffer(length: seqLen * nH * hDim * 4)!
-        try bidirectionalAttention(q: qBuf, k: kBuf, v: vBuf, output: attnOut, seqLen: seqLen, cmdBuf: cmdBuf)
-
-        // O projection
-        let h2 = device.makeBuffer(length: seqLen * hs * 4)!
-        try matmulSeq(input: attnOut, weight: oProjs[layerIdx], output: h2, m: seqLen, k: nH * hDim, n: hs, cmdBuf: cmdBuf)
-
-        // Post-attn norm
-        let h2n = try applyRmsNorm(input: h2, weight: layerNorms[layerIdx][2], count: seqLen * hs)
-
-        // Add residual: hidden = resid + h2n
-        try eltwiseAdd(a: resid, b: h2n, output: hidden, count: seqLen * hs, cmdBuf: cmdBuf)
-
-        // Pre-FF norm
-        let h3 = try applyRmsNorm(input: hidden, weight: layerNorms[layerIdx][1], count: seqLen * hs)
-
-        // MLP: gate, up
-        let gate = device.makeBuffer(length: seqLen * intermedi * 4)!
-        let up = device.makeBuffer(length: seqLen * intermedi * 4)!
-        try matmulSeq(input: h3, weight: gateProjs[layerIdx], output: gate, m: seqLen, k: hs, n: intermedi, cmdBuf: cmdBuf)
-        try matmulSeq(input: h3, weight: upProjs[layerIdx], output: up, m: seqLen, k: hs, n: intermedi, cmdBuf: cmdBuf)
-
-        // GELU(gate) * up
-        let gated = device.makeBuffer(length: seqLen * intermedi * 4)!
-        try geluMul(gate: gate, up: up, output: gated, count: seqLen * intermedi, cmdBuf: cmdBuf)
-
-        // Down projection
-        let h4 = device.makeBuffer(length: seqLen * hs * 4)!
-        try matmulSeq(input: gated, weight: downProjs[layerIdx], output: h4, m: seqLen, k: intermedi, n: hs, cmdBuf: cmdBuf)
-
-        // Post-FF norm
-        let h4n = try applyRmsNorm(input: h4, weight: layerNorms[layerIdx][3], count: seqLen * hs)
-
-        // Add residual: hidden = hidden + h4n
-        try eltwiseAdd(a: hidden, b: h4n, output: hidden, count: seqLen * hs, cmdBuf: cmdBuf)
-
-        cmdBuf.commit(); cmdBuf.waitUntilCompleted()
-        return hidden
-    }
-
-    // MARK: - Metal Kernels
-
     private func matmulSeq(input: MTLBuffer, weight: MTLBuffer, output: MTLBuffer, m: Int, k: Int, n: Int, cmdBuf: MTLCommandBuffer) throws {
-        let enc = cmdBuf.makeComputeCommandEncoder()!
-        let pso = try engine.pipeline(named: "matmul_f32")
-        enc.setComputePipelineState(pso)
-        enc.setBuffer(input, offset: 0, index: 0)
-        enc.setBuffer(weight, offset: 0, index: 1)
-        enc.setBuffer(output, offset: 0, index: 2)
-        var mm = UInt32(m), kk = UInt32(k), nn = UInt32(n)
-        enc.setBytes(&mm, length: 4, index: 3)
-        enc.setBytes(&kk, length: 4, index: 4)
-        enc.setBytes(&nn, length: 4, index: 5)
-        enc.dispatchThreads(MTLSize(width: m * n, height: 1, depth: 1),
-                           threadsPerThreadgroup: MTLSize(width: min(256, m * n), height: 1, depth: 1))
-        enc.endEncoding()
+        // Use MPS for optimized matrix multiplication on Apple Silicon
+        // Weight is stored transposed [k, n] for C = A × B
+        let descA = MPSMatrixDescriptor(rows: m, columns: k, rowBytes: k * 4, dataType: .float32)
+        let descB = MPSMatrixDescriptor(rows: k, columns: n, rowBytes: n * 4, dataType: .float32)
+        let descC = MPSMatrixDescriptor(rows: m, columns: n, rowBytes: n * 4, dataType: .float32)
+        let matA = MPSMatrix(buffer: input, descriptor: descA)
+        let matB = MPSMatrix(buffer: weight, descriptor: descB)
+        let matC = MPSMatrix(buffer: output, descriptor: descC)
+        
+        let matMul = MPSMatrixMultiplication(device: engine.device,
+                                              transposeLeft: false,
+                                              transposeRight: false,
+                                              resultRows: m,
+                                              resultColumns: n,
+                                              interiorColumns: k,
+                                              alpha: 1.0,
+                                              beta: 0.0)
+        matMul.encode(commandBuffer: cmdBuf, leftMatrix: matA, rightMatrix: matB, resultMatrix: matC)
     }
 
-    private func applyRoPE(q: MTLBuffer, k: MTLBuffer, seqLen: Int, headDim: Int, numHeads: Int, numKVHeads: Int, cmdBuf: MTLCommandBuffer) throws {
+    private func eltwiseAdd(a: MTLBuffer, b: MTLBuffer, output: MTLBuffer, count: Int, cmdBuf: MTLCommandBuffer) throws {
         let enc = cmdBuf.makeComputeCommandEncoder()!
+        defer { enc.endEncoding() }
+        let pso = try engine.pipeline(named: "eltwise_add")
+        enc.setComputePipelineState(pso)
+        enc.setBuffer(a, offset: 0, index: 0)
+        enc.setBuffer(b, offset: 0, index: 1)
+        enc.setBuffer(output, offset: 0, index: 2)
+        var c = UInt32(count)
+        enc.setBytes(&c, length: 4, index: 3)
+        enc.dispatchThreads(MTLSize(width: count, height: 1, depth: 1), threadsPerThreadgroup: MTLSize(width: min(256, count), height: 1, depth: 1))
+    }
+
+    private func geluMul(gate: MTLBuffer, up: MTLBuffer, output: MTLBuffer, count: Int, cmdBuf: MTLCommandBuffer) throws {
+        let enc = cmdBuf.makeComputeCommandEncoder()!
+        defer { enc.endEncoding() }
+        let pso = try engine.pipeline(named: "gelu_mul_kernel")
+        enc.setComputePipelineState(pso)
+        enc.setBuffer(gate, offset: 0, index: 0)
+        enc.setBuffer(up, offset: 0, index: 1)
+        enc.setBuffer(output, offset: 0, index: 2)
+        var c = UInt32(count)
+        enc.setBytes(&c, length: 4, index: 3)
+        enc.dispatchThreads(MTLSize(width: count, height: 1, depth: 1), threadsPerThreadgroup: MTLSize(width: min(256, count), height: 1, depth: 1))
+    }
+
+    private func applyRoPE(q: MTLBuffer, k: MTLBuffer, seqLen: Int, headDim: Int, numHeads: Int, cmdBuf: MTLCommandBuffer) throws {
+        let enc = cmdBuf.makeComputeCommandEncoder()!
+        defer { enc.endEncoding() }
         let pso = try engine.pipeline(named: "apply_rope")
         enc.setComputePipelineState(pso)
         enc.setBuffer(q, offset: 0, index: 0)
@@ -308,18 +273,12 @@ public final class EmbeddingGemmaModel {
         enc.setBytes(&hd, length: 4, index: 3)
         enc.setBytes(&nh, length: 4, index: 4)
         enc.setBytes(&rt, length: 4, index: 5)
-        enc.dispatchThreads(MTLSize(width: numHeads * headDim / 2, height: 1, depth: 1),
-                           threadsPerThreadgroup: MTLSize(width: min(256, numHeads * headDim / 2), height: 1, depth: 1))
-        enc.endEncoding()
-    }
-
-    private func applyQKNorm(q: MTLBuffer, k: MTLBuffer, qNorm: MTLBuffer, kNorm: MTLBuffer, seqLen: Int, headDim: Int, numHeads: Int, numKVHeads: Int, cmdBuf: MTLCommandBuffer) throws {
-        // Apply RMSNorm per head
-        // TODO: Implement per-head normalization
+        enc.dispatchThreads(MTLSize(width: numHeads * headDim / 2, height: 1, depth: 1), threadsPerThreadgroup: MTLSize(width: min(256, numHeads * headDim / 2), height: 1, depth: 1))
     }
 
     private func bidirectionalAttention(q: MTLBuffer, k: MTLBuffer, v: MTLBuffer, output: MTLBuffer, seqLen: Int, cmdBuf: MTLCommandBuffer) throws {
         let enc = cmdBuf.makeComputeCommandEncoder()!
+        defer { enc.endEncoding() }
         let pso = try engine.pipeline(named: "bidirectional_sliding_attn")
         enc.setComputePipelineState(pso)
         enc.setBuffer(q, offset: 0, index: 0)
@@ -335,38 +294,84 @@ public final class EmbeddingGemmaModel {
         enc.setBytes(&nkv, length: 4, index: 7)
         enc.setBytes(&sw, length: 4, index: 8)
         enc.setBytes(&scale, length: 4, index: 9)
-        let tgMem = config.slidingWindow * 4  // shared memory for scores
+        let tgMem = config.slidingWindow * 4
         enc.setThreadgroupMemoryLength(tgMem, index: 0)
-        enc.dispatchThreads(MTLSize(width: seqLen * config.numAttentionHeads, height: 1, depth: 1),
-                           threadsPerThreadgroup: MTLSize(width: min(256, seqLen * config.numAttentionHeads), height: 1, depth: 1))
-        enc.endEncoding()
+        enc.dispatchThreads(MTLSize(width: seqLen * config.numAttentionHeads, height: 1, depth: 1), threadsPerThreadgroup: MTLSize(width: min(256, seqLen * config.numAttentionHeads), height: 1, depth: 1))
     }
 
-    private func eltwiseAdd(a: MTLBuffer, b: MTLBuffer, output: MTLBuffer, count: Int, cmdBuf: MTLCommandBuffer) throws {
-        let enc = cmdBuf.makeComputeCommandEncoder()!
-        let pso = try engine.pipeline(named: "eltwise_add")
-        enc.setComputePipelineState(pso)
-        enc.setBuffer(a, offset: 0, index: 0)
-        enc.setBuffer(b, offset: 0, index: 1)
-        enc.setBuffer(output, offset: 0, index: 2)
-        var c = UInt32(count)
-        enc.setBytes(&c, length: 4, index: 3)
-        enc.dispatchThreads(MTLSize(width: count, height: 1, depth: 1),
-                           threadsPerThreadgroup: MTLSize(width: min(256, count), height: 1, depth: 1))
-        enc.endEncoding()
-    }
+    private func forwardLayerDebug(hidden: MTLBuffer, layerIdx: Int, seqLen: Int, cmdBuf: MTLCommandBuffer) throws -> MTLBuffer {
+        let hs = config.hiddenSize, device = engine.device
+        let hDim = config.headDim, nH = config.numAttentionHeads, nKV = config.numKeyValueHeads
+        let intermedi = config.intermediateSize
 
-    private func geluMul(gate: MTLBuffer, up: MTLBuffer, output: MTLBuffer, count: Int, cmdBuf: MTLCommandBuffer) throws {
-        let enc = cmdBuf.makeComputeCommandEncoder()!
-        let pso = try engine.pipeline(named: "gelu_mul_kernel")
-        enc.setComputePipelineState(pso)
-        enc.setBuffer(gate, offset: 0, index: 0)
-        enc.setBuffer(up, offset: 0, index: 1)
-        enc.setBuffer(output, offset: 0, index: 2)
-        var c = UInt32(count)
-        enc.setBytes(&c, length: 4, index: 3)
-        enc.dispatchThreads(MTLSize(width: count, height: 1, depth: 1),
-                           threadsPerThreadgroup: MTLSize(width: min(256, count), height: 1, depth: 1))
-        enc.endEncoding()
+        print("    Residual copy...")
+        let resid = device.makeBuffer(length: seqLen * hs * 4)!
+        let blit = cmdBuf.makeBlitCommandEncoder()!
+        blit.copy(from: hidden, sourceOffset: 0, to: resid, destinationOffset: 0, size: seqLen * hs * 4)
+        blit.endEncoding()
+
+        print("    Input norm...")
+        let h1 = try applyRmsNorm(input: hidden, weight: layerNorms[layerIdx][0], count: seqLen * hs, cmdBuf: cmdBuf)
+
+        print("    QKV projections...")
+        let qBuf = device.makeBuffer(length: seqLen * nH * hDim * 4)!
+        let kBuf = device.makeBuffer(length: seqLen * nKV * hDim * 4)!
+        let vBuf = device.makeBuffer(length: seqLen * nKV * hDim * 4)!
+        try matmulSeq(input: h1, weight: qProjs[layerIdx], output: qBuf, m: seqLen, k: hs, n: nH * hDim, cmdBuf: cmdBuf)
+        print("    Q done")
+        try matmulSeq(input: h1, weight: kProjs[layerIdx], output: kBuf, m: seqLen, k: hs, n: nKV * hDim, cmdBuf: cmdBuf)
+        print("    K done")
+        try matmulSeq(input: h1, weight: vProjs[layerIdx], output: vBuf, m: seqLen, k: hs, n: nKV * hDim, cmdBuf: cmdBuf)
+        print("    V done")
+
+        print("    RoPE...")
+        try applyRoPE(q: qBuf, k: kBuf, seqLen: seqLen, headDim: hDim, numHeads: nH, cmdBuf: cmdBuf)
+        print("    RoPE done")
+
+        print("    Attention...")
+        let attnOut = device.makeBuffer(length: seqLen * nH * hDim * 4)!
+        try bidirectionalAttention(q: qBuf, k: kBuf, v: vBuf, output: attnOut, seqLen: seqLen, cmdBuf: cmdBuf)
+        print("    Attention done")
+
+        print("    O projection...")
+        let h2 = device.makeBuffer(length: seqLen * hs * 4)!
+        try matmulSeq(input: attnOut, weight: oProjs[layerIdx], output: h2, m: seqLen, k: nH * hDim, n: hs, cmdBuf: cmdBuf)
+        print("    O done")
+
+        print("    Post-attn norm...")
+        let h2n = try applyRmsNorm(input: h2, weight: layerNorms[layerIdx][2], count: seqLen * hs, cmdBuf: cmdBuf)
+
+        print("    Add residual 1...")
+        try eltwiseAdd(a: resid, b: h2n, output: hidden, count: seqLen * hs, cmdBuf: cmdBuf)
+
+        print("    Pre-FF norm...")
+        let h3 = try applyRmsNorm(input: hidden, weight: layerNorms[layerIdx][1], count: seqLen * hs, cmdBuf: cmdBuf)
+
+        print("    MLP gate/up...")
+        let gate = device.makeBuffer(length: seqLen * intermedi * 4)!
+        let up = device.makeBuffer(length: seqLen * intermedi * 4)!
+        try matmulSeq(input: h3, weight: gateProjs[layerIdx], output: gate, m: seqLen, k: hs, n: intermedi, cmdBuf: cmdBuf)
+        print("    Gate done")
+        try matmulSeq(input: h3, weight: upProjs[layerIdx], output: up, m: seqLen, k: hs, n: intermedi, cmdBuf: cmdBuf)
+        print("    Up done")
+
+        print("    GELU mul...")
+        let gated = device.makeBuffer(length: seqLen * intermedi * 4)!
+        try geluMul(gate: gate, up: up, output: gated, count: seqLen * intermedi, cmdBuf: cmdBuf)
+        print("    GELU done")
+
+        print("    Down projection...")
+        let h4 = device.makeBuffer(length: seqLen * hs * 4)!
+        try matmulSeq(input: gated, weight: downProjs[layerIdx], output: h4, m: seqLen, k: intermedi, n: hs, cmdBuf: cmdBuf)
+        print("    Down done")
+
+        print("    Post-FF norm...")
+        let h4n = try applyRmsNorm(input: h4, weight: layerNorms[layerIdx][3], count: seqLen * hs, cmdBuf: cmdBuf)
+
+        print("    Add residual 2...")
+        try eltwiseAdd(a: hidden, b: h4n, output: hidden, count: seqLen * hs, cmdBuf: cmdBuf)
+        print("    Layer \(layerIdx) complete")
+
+        return hidden
     }
 }

Sources/MarkBase/MetalKernels.swift

@@ -109,6 +109,21 @@ public enum MetalKernels {
             .replacingOccurrences(of: "using namespace metal;\n", with: "")
         result += "\n" + fusedStripped
         
+        // Strip preamble from embedding kernels source
+        let embStripped = embeddingKernelsSource
+            .replacingOccurrences(of: "#include <metal_stdlib>\n", with: "")
+            .replacingOccurrences(of: "using namespace metal;\n", with: "")
+        result += "\n" + embStripped
+        
         return result
     }
-}
+
+    /// Embedding kernel source for EmbeddingGemma.
+    /// Includes RoPE, bidirectional sliding window attention, Q/K norm, and GELU.
+    public static var embeddingKernelsSource: String {
+        let url = URL(fileURLWithPath: #filePath)
+            .deletingLastPathComponent()
+            .appendingPathComponent("Metal/EmbeddingKernels.metal")
+        return try! String(contentsOf: url, encoding: .utf8)
+    }
+}

Sources/MarkBaseServer/EmbeddingServerApp.swift

@@ -0,0 +1,113 @@
+import Foundation
+import MarkBase
+import Hummingbird
+
+struct EmbeddingServerApp {
+    static func main() async throws {
+        let args = CommandLine.arguments
+        let modelName = args.count > 1 ? args[1] : "embeddinggemma-300m"
+        let port = args.count > 2 ? Int(args[2]) ?? 8084 : 8084
+        
+        let modelPath = NSString(string: "~/MarkBaseEngine/models/\(modelName)").expandingTildeInPath
+
+        print("═══════════════════════════════════════════════════════════════════")
+        print("  MarkBaseEngine Embedding Server")
+        print("═══════════════════════════════════════════════════════════════════")
+        print("  Model: \(modelName)")
+        print("  Port: \(port)")
+        print("  Path: \(modelPath)")
+        print("")
+
+        let engine = try MarkBaseEngine(autoCompile: true)
+        let embedModel = try EmbeddingGemmaModel(modelDir: modelPath, engine: engine)
+        let layers = embedModel.config.numHiddenLayers
+        let hiddenSize = embedModel.config.hiddenSize
+
+        print("✓ EmbeddingGemma loaded (\(layers) layers, hidden=\(hiddenSize))")
+
+        let router = Router()
+
+        router.get("/") { _, _ in
+            return """
+            {
+              "server": {
+                "name": "MarkBaseEngine Embedding",
+                "model": "embeddinggemma-300m",
+                "layers": \(layers),
+                "hidden_size": \(hiddenSize),
+                "output_dim": \(hiddenSize),
+                "framework": "Hummingbird 2.x + Metal GPU"
+              },
+              "endpoints": [
+                {"method": "POST", "path": "/v1/embeddings", "summary": "Generate text embeddings"}
+              ]
+            }
+            """
+        }
+
+        router.get("/health") { _, _ in
+            return "{\"status\":\"healthy\",\"model\":\"embeddinggemma-300m\",\"layers\":\(layers),\"hidden_size\":\(hiddenSize)}"
+        }
+
+        router.post("/v1/embeddings") { request, _ in
+            let buffer = try await request.body.collect(upTo: .max)
+            let data = Data(buffer: buffer)
+
+            guard let json = try JSONSerialization.jsonObject(with: data) as? [String: Any],
+                  let input = json["input"] else {
+                return "{\"error\":\"invalid request\",\"type\":\"invalid_request_error\",\"code\":400,\"message\":\"missing 'input' field\"}"
+            }
+
+            let modelId = (json["model"] as? String) ?? "embeddinggemma-300m"
+            let encodingFormat = (json["encoding_format"] as? String) ?? "float"
+
+            let inputs: [String]
+            if let str = input as? String { inputs = [str] }
+            else if let arr = input as? [String] { inputs = arr }
+            else {
+                return "{\"error\":\"invalid request\",\"type\":\"invalid_request_error\",\"code\":400,\"message\":\"'input' must be string or array of strings\"}"
+            }
+
+            var embeddings: [[String: Any]] = []
+            for (i, text) in inputs.enumerated() {
+                let t0 = Date()
+                let embedding = try embedModel.embed(text: text)
+                let duration = Date().timeIntervalSince(t0)
+
+                let embeddingData: [String: Any]
+                if encodingFormat == "base64" {
+                    let base64 = embedding.withUnsafeBytes { Data($0).base64EncodedString() }
+                    embeddingData = ["object": "embedding", "index": i, "embedding": base64, "usage_ms": Int(duration * 1000)]
+                } else {
+                    embeddingData = ["object": "embedding", "index": i, "embedding": embedding, "usage_ms": Int(duration * 1000)]
+                }
+                embeddings.append(embeddingData)
+            }
+
+            let id = UUID().uuidString
+            let ts = Int(Date().timeIntervalSince1970)
+            let response: [String: Any] = [
+                "id": id, "object": "list", "created": ts, "model": modelId,
+                "data": embeddings,
+                "usage": ["prompt_tokens": inputs.reduce(0) { $0 + $1.components(separatedBy: .whitespaces).count }, "total_tokens": inputs.reduce(0) { $0 + $1.components(separatedBy: .whitespaces).count }]
+            ]
+
+            let jsonData = try JSONSerialization.data(withJSONObject: response)
+            return String(data: jsonData, encoding: .utf8) ?? "{}"
+        }
+
+        let app = Application(
+            router: router,
+            configuration: .init(address: .hostname("0.0.0.0", port: port))
+        )
+
+        print("Server starting on port \(port)...")
+        print("Endpoints:")
+        print("  GET  /             - Info")
+        print("  GET  /health       - Health check")
+        print("  POST /v1/embeddings - Text embeddings")
+        print("")
+
+        try await app.run()
+    }
+}

Sources/MarkBaseServer/SimpleServerApp.swift

@@ -20,16 +20,36 @@ struct SimpleServerApp {
         print("")
 
         let engine = try MarkBaseEngine(autoCompile: true)
-        let model = try E4BModel(modelDir: modelPath, engine: engine, maxContextLength: 512)
+        
+        // Detect model type
+        let isEmbeddingGemma = modelName.contains("embeddinggemma") || modelName.contains("embedding")
+        let isE4B = !isEmbeddingGemma
+        
+        var e4bModel: E4BModel? = nil
+        var embedModel: EmbeddingGemmaModel? = nil
         let tokenizer = try TokenizerFactory.load(modelDir: modelPath)
-        let generator = StreamingGenerator(model: model, tokenizer: tokenizer, engine: engine)
-        let embeddingModel = try TextEmbeddingModel(modelDir: modelPath, engine: engine, config: TextEmbeddingConfig())
-
-        print("✓ E4B loaded (\(model.numHiddenLayers) layers)")
+        var generator: StreamingGenerator? = nil
+        
+        if isE4B {
+            e4bModel = try E4BModel(modelDir: modelPath, engine: engine, maxContextLength: 512)
+            generator = StreamingGenerator(model: e4bModel!, tokenizer: tokenizer, engine: engine)
+            print("✓ E4B loaded (\(e4bModel!.numHiddenLayers) layers)")
+        } else {
+            embedModel = try EmbeddingGemmaModel(modelDir: modelPath, engine: engine)
+            print("✓ EmbeddingGemma loaded (\(embedModel!.config.numHiddenLayers) layers, hidden=\(embedModel!.config.hiddenSize))")
+        }
+        
+        // Embedding model (use EmbeddingGemma if available, fallback to E4B)
+        let textEmbedModel: TextEmbeddingModel?
+        if isE4B {
+            textEmbedModel = try TextEmbeddingModel(modelDir: modelPath, engine: engine, config: TextEmbeddingConfig())
+        } else {
+            textEmbedModel = nil
+        }
 
         let router = Router()
 
-        let layers = model.numHiddenLayers
+        let layers = isE4B ? e4bModel!.numHiddenLayers : embedModel!.config.numHiddenLayers
 
         @Sendable func helpJSON() -> String {
             return """
@@ -184,7 +204,8 @@ struct SimpleServerApp {
         }
 
         @Sendable func healthResponse() -> String {
-            return "{\"status\":\"healthy\",\"model\":\"e4b\",\"layers\":\(layers)}"
+            let m = isEmbeddingGemma ? "embeddinggemma-300m" : "e4b"
+            return "{\"status\":\"healthy\",\"model\":\"\(m)\",\"layers\":\(layers)}"
         }
 
         router.get("/") { _, _ in
@@ -229,7 +250,7 @@ struct SimpleServerApp {
                 topP: topP
             )
 
-            let generatedTokens = try generator.generateTokens(promptTokens: promptTokens, config: config)
+            let generatedTokens = try generator!.generateTokens(promptTokens: promptTokens, config: config)
 
             let id = UUID().uuidString
             let ts = Int(Date().timeIntervalSince1970)
@@ -313,7 +334,14 @@ struct SimpleServerApp {
             var embeddings: [[String: Any]] = []
             for (i, text) in inputs.enumerated() {
                 let t0 = Date()
-                let embedding = try embeddingModel.embed(text: text)
+                let embedding: [Float]
+                if let eg = embedModelCopy {
+                    embedding = try eg.embed(text: text)
+                } else if let tem = textEmbedModelCopy {
+                    embedding = try tem.embed(text: text)
+                } else {
+                    return "{\"error\":\"no embedding model available\",\"type\":\"server_error\",\"code\":500}"
+                }
                 let duration = Date().timeIntervalSince(t0)
 
                 let embeddingData: [String: Any]
@@ -338,7 +366,7 @@ struct SimpleServerApp {
 
             let id = UUID().uuidString
             let ts = Int(Date().timeIntervalSince1970)
-            let totalTokens = inputs.reduce(0) { $0 + tokenizer.encode(text: $1).count }
+            let totalTokens = inputs.reduce(0) { $0 + $1.components(separatedBy: .whitespaces).count }
 
             let response: [String: Any] = [
                 "id": id,

Sources/MarkBaseServer/main.swift

@@ -1,9 +1,17 @@
 import Foundation
 
-// Entry point — avoids @main conflict with top-level code
+// Entry point — routes to appropriate server based on model name
 Task {
     do {
-        try await SimpleServerApp.main()
+        let args = CommandLine.arguments
+        let modelName = args.count > 1 ? args[1] : "E4B-MarkBase"
+        let isEmbedding = modelName.contains("embedding") || modelName.contains("embeddinggemma")
+        
+        if isEmbedding {
+            try await EmbeddingServerApp.main()
+        } else {
+            try await SimpleServerApp.main()
+        }
     } catch {
         print("Server error: \(error)")
         exit(1)

Tests/01_Model/EmbeddingGemmaTest.swift

@@ -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))")
+    }
+}