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

Sources/MarkBaseServer/RDMADistributionService.swift

@@ -0,0 +1,101 @@
+import Foundation
+import MarkBase
+import Metal
+import RDMAKit
+
+// ─────────────────────────────────────────────────────────────
+// RDMA Distribution Service — Pipeline Parallelism
+// ─────────────────────────────────────────────────────────────
+
+public actor RDMADistributionService {
+    public enum Role: Sendable {
+        case primary(splitLayer: Int)
+        case secondary
+    }
+
+    public struct Config: Sendable {
+        public let role: Role
+        public let peerAddress: String?
+        public let peerPort: UInt16
+
+        public init(role: Role, peerAddress: String? = nil, peerPort: UInt16 = 0) {
+            self.role = role
+            self.peerAddress = peerAddress
+            self.peerPort = peerPort
+        }
+    }
+
+    // RDMA state
+    private var context: RDMAContext?
+    private var pd: RDMAProtectionDomain?
+    private var sendCQ: RDMACompletionQueue?
+    private var recvCQ: RDMACompletionQueue?
+    private var qp: RDMAQueuePair?
+
+    // Registered GPU buffers
+    private var registeredHiddenState: RegisteredMemory?
+
+    public let config: Config
+    public let discovery: RDMADiscovery
+
+    public init(config: Config) {
+        self.config = config
+        self.discovery = RDMADiscovery()
+    }
+
+    /// Initialize RDMA device and resources. Returns false if no RDMA device is available.
+    @discardableResult
+    public func initialize() throws -> Bool {
+        let devices = discovery.listDevices()
+        guard let device = devices.first else {
+            print("  RDMA: No devices found — running in local-only mode")
+            return false
+        }
+
+        let tbInfo = device.isThunderbolt ? " (Thunderbolt)" : ""
+        print("  RDMA: Found device '\(device.name)'\(tbInfo)")
+        context = try discovery.openDevice(device)
+        let attrs = try context!.queryDeviceAttributes()
+        print("  RDMA: Firmware \(attrs.fwVer), max MR size \(attrs.maxMRSize) bytes")
+
+        pd = try RDMAProtectionDomain(context: context!)
+        sendCQ = try RDMACompletionQueue(context: context!, cqe: 128)
+        recvCQ = try RDMACompletionQueue(context: context!, cqe: 128)
+        qp = try RDMAQueuePair(pd: pd!, sendCQ: sendCQ!, recvCQ: recvCQ!)
+        try qp!.modifyToInit()
+        print("  RDMA: QP initialized in RC mode")
+
+        if case .primary = config.role {
+            try qp!.modifyToRTR(destQPN: 0)
+            try qp!.modifyToRTS()
+        }
+
+        print("  RDMA: Ready")
+        return true
+    }
+
+    /// Register a Metal buffer for remote RDMA access
+    public func registerBuffer(_ buffer: MTLBuffer) throws -> RegisteredMemory? {
+        guard let pd else { return nil }
+        let reg = try pd.registerMTLBuffer(
+            UnsafeMutableRawPointer(buffer.contents()),
+            length: buffer.length
+        )
+        print("  RDMA: Registered buffer \(buffer.length) bytes (lkey=\(reg.lkey), rkey=\(reg.rkey))")
+        return reg
+    }
+
+    /// Register the model's hidden state buffer for pipeline-split transfer
+    public func registerHiddenState(_ model: E4BModel) throws {
+        let buf = model.temps.io
+        let reg = try registerBuffer(buf)
+        registeredHiddenState = reg
+        print("  RDMA: Hidden state buffer registered (\(buf.length) bytes)")
+    }
+
+    /// Split point for pipeline parallelism
+    public var splitLayer: Int? {
+        if case .primary(let layer) = config.role { return layer }
+        return nil
+    }
+}