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 +1,51 @@
+import Foundation
+
+public struct AudioConfig: Codable {
+    public let hiddenSize: Int
+    public let numAttentionHeads: Int
+    public let numHiddenLayers: Int
+    public let convKernelSize: Int
+    public let attentionChunkSize: Int
+    public let attentionContextLeft: Int
+    public let attentionContextRight: Int
+    public let attentionLogitCap: Float
+    public let hiddenAct: String
+    public let rmsNormEps: Float
+    public let outputProjDims: Int
+    public let subsamplingConvChannels: [Int]
+    public let residualWeight: Float
+    
+    public init(
+        hiddenSize: Int = 1024,
+        numAttentionHeads: Int = 8,
+        numHiddenLayers: Int = 12,
+        convKernelSize: Int = 5,
+        attentionChunkSize: Int = 12,
+        attentionContextLeft: Int = 13,
+        attentionContextRight: Int = 0,
+        attentionLogitCap: Float = 50.0,
+        hiddenAct: String = "silu",
+        rmsNormEps: Float = 1e-6,
+        outputProjDims: Int = 1536,
+        subsamplingConvChannels: [Int] = [128, 32],
+        residualWeight: Float = 0.5
+    ) {
+        self.hiddenSize = hiddenSize
+        self.numAttentionHeads = numAttentionHeads
+        self.numHiddenLayers = numHiddenLayers
+        self.convKernelSize = convKernelSize
+        self.attentionChunkSize = attentionChunkSize
+        self.attentionContextLeft = attentionContextLeft
+        self.attentionContextRight = attentionContextRight
+        self.attentionLogitCap = attentionLogitCap
+        self.hiddenAct = hiddenAct
+        self.rmsNormEps = rmsNormEps
+        self.outputProjDims = outputProjDims
+        self.subsamplingConvChannels = subsamplingConvChannels
+        self.residualWeight = residualWeight
+    }
+    
+    public var headDim: Int {
+        hiddenSize / numAttentionHeads
+    }
+}