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 +1,58 @@
+import Metal
+
+// ══════════════════════════════════════════════════════════════════
+// Batch Forward Temps - Extended buffers for batch processing
+// ══════════════════════════════════════════════════════════════════
+
+extension ForwardTemps {
+    
+    /// Create batch-specific temporary buffers
+    /// These are separate from single-token buffers to avoid interference
+    public func createBatchBuffers(
+        device: MTLDevice,
+        batchSize: Int,
+        hiddenSize: Int,
+        nHeads: Int,
+        headDim: Int,
+        intermediateSize: Int
+    ) throws -> BatchTemps {
+        return try BatchTemps(
+            device: device,
+            batchSize: batchSize,
+            hiddenSize: hiddenSize,
+            nHeads: nHeads,
+            headDim: headDim,
+            intermediateSize: intermediateSize
+        )
+    }
+}
+
+/// Batch-specific temporary buffers for parallel layer processing
+public struct BatchTemps {
+    public let hBatch: MTLBuffer        // [batchSize, hiddenSize] - hidden state batch
+    public let qBatch: MTLBuffer        // [batchSize, nHeads * headDim] - query batch
+    public let nsBatch: MTLBuffer       // [batchSize, nHeads * headDim] - norm scratch batch
+    public let interBatch: MTLBuffer    // [batchSize, intermediateSize] - intermediate batch
+    
+    public init(
+        device: MTLDevice,
+        batchSize: Int,
+        hiddenSize: Int,
+        nHeads: Int,
+        headDim: Int,
+        intermediateSize: Int
+    ) throws {
+        func buf(_ n: Int) throws -> MTLBuffer {
+            guard let b = device.makeBuffer(length: n * MemoryLayout<Float>.stride,
+                                            options: .storageModeShared)
+            else { throw NSError(domain: "BatchTemps", code: -1, 
+                                 userInfo: [NSLocalizedDescriptionKey: "Buffer creation failed"]) }
+            return b
+        }
+        
+        hBatch = try buf(batchSize * hiddenSize)
+        qBatch = try buf(batchSize * nHeads * headDim)
+        nsBatch = try buf(batchSize * nHeads * headDim)
+        interBatch = try buf(batchSize * intermediateSize)
+    }
+}