import XCTest
@testable import MarkBase

final class RouterMatmulInputDebugTest: XCTestCase {
    
    func testRouterMatmulInputDebug() throws {
        print("\n═══════════════════════════════════════")
        print("      Router Matmul Input Debug Test")
        print("═══════════════════════════════════════\n")
        
        let modelDir = "/Users/accusys/MarkBaseEngine/models/gemma-4-26b-a4b-it-4bit"
        
        print("Loading model...")
        fflush(stdout)
        let engine = try MarkBaseEngine(autoCompile: true)
        let model = try E4BModel(modelDir: modelDir, engine: engine, maxContextLength: 32)
        
        print("\nStep 1: Get layer 0...")
        fflush(stdout)
        let layer0 = model.layers[0]
        
        print("\nStep 2: Create input buffer (simulating pre_feedforward_norm output)...")
        fflush(stdout)
        
        // Create input buffer with realistic values (from hidden state)
        let hs = model.hiddenSize
        let inputBuffer = engine.device.makeBuffer(length: hs * MemoryLayout<Float>.size)!
        
        // Fill with realistic values (similar to hidden state)
        var inputData = [Float](repeating: 0, count: hs)
        for i in 0..<hs {
            inputData[i] = Float.random(in: -2.0...2.0)  // Realistic hidden state range
        }
        memcpy(inputBuffer.contents(), &inputData, inputData.count * MemoryLayout<Float>.size)
        
        print("  Input buffer created with random values (range -2.0 to 2.0)")
        let inputSample = engine.readFloats(from: inputBuffer, count: 10)
        print("  Input sample: \(inputSample)")
        fflush(stdout)
        
        print("\nStep 3: Test router matmul...")
        fflush(stdout)
        
        guard let router = layer0.routerProj else {
            print("  ERROR: routerProj is nil!")
            XCTFail("routerProj is nil")
            return
        }
        
        print("  Router weights: bits=\(router.bits), inDim=\(router.inDim), outDim=\(router.outDim)")
        fflush(stdout)
        
        // Create output buffer
        let outputBuffer = engine.device.makeBuffer(length: router.outDim * MemoryLayout<Float>.size)!
        
        // Run router matmul
        let cmdBuf = engine.commandQueue.makeCommandBuffer()!
        try layer0.quantizedMatmul(engine: engine, cmdBuf: cmdBuf,
                                   input: inputBuffer, weights: router,
                                   output: outputBuffer)
        cmdBuf.commit()
        cmdBuf.waitUntilCompleted()
        
        print("  Router matmul completed")
        fflush(stdout)
        
        // Read output
        let routerOutput = engine.readFloats(from: outputBuffer, count: router.outDim)
        print("  Router output first 10: \(routerOutput[0..<min(10, router.outDim)])")
        print("  Router output max/min: \(routerOutput.max() ?? 0), \(routerOutput.min() ?? 0)")
        print("  Router output mean: \(routerOutput.reduce(0, +) / Float(router.outDim))")
        fflush(stdout)
        
        // Check if output is all zeros
        let hasNonzero = routerOutput.contains { $0 != 0 }
        if hasNonzero {
            print("  ✓ Router matmul works (has non-zero output)")
        } else {
            print("  ✗ Router matmul returns all zeros!")
            XCTFail("Router matmul returns all zeros")
        }
        
        print("\n═══════════════════════════════════════")
        print("✓ Router matmul input debug completed")
        print("═══════════════════════════════════════\n")
        fflush(stdout)
    }
}