markbaseengine/Tests/MarkBaseTests/MultimodalSequentialTest.swift

import XCTest
@testable import MarkBase

final class MultimodalSequentialTest: XCTestCase {
    
    func testE2BFullMultimodal() throws {
        print("\n═══════════════════════════════════════════════════════════════════")
        print("  E2B Full Multimodal Test (TEXT + Audio + Vision)")
        print("═══════════════════════════════════════════════════════════════════\n")
        
        let modelDir = "/Users/accusys/.cache/huggingface/hub/models--mlx-community--gemma-4-e2b-it-4bit/snapshots/2c3e507453b4f218d05fe3cc97bea5c5a654257e"
        
        print("Step 1: Load MultimodalModel (TEXT + Audio + Vision)...")
        let engine = try MarkBaseEngine(autoCompile: true)
        let loadStart = Date()
        let mmModel = try MultimodalModel(modelDir: modelDir, engine: engine, maxContextLength: 256)
        let loadTime = Date().timeIntervalSince(loadStart) * 1000
        print("  ✓ Model loaded in \(loadTime) ms")
        print("  TEXT: hidden=\(mmModel.textModel.hiddenSize), layers=\(mmModel.textModel.numHiddenLayers)")
        print("  Audio: \(mmModel.audioTowerE2B != nil ? "E2B (12 layers)" : "N/A")")
        print("  Vision: \(mmModel.visionTowerE2B != nil ? "E2B (16 layers)" : "N/A")")
        
        // ── Audio Test ───────────────────────────────────────────────────
        print("\nStep 2: Process Audio...")
        let seqLen = 100
        let nMels = 128
        var melFeatures: [[Float]] = []
        for _ in 0..<seqLen {
            var frame: [Float] = []
            for _ in 0..<nMels {
                frame.append(Float.random(in: -0.5...0.5))
            }
            melFeatures.append(frame)
        }
        
        let audioStart = Date()
        let audioEmbeds = try mmModel.processAudio(audioFeatures: melFeatures)
        let audioTime = Date().timeIntervalSince(audioStart) * 1000
        print("  ✓ Audio processed in \(audioTime) ms")
        print("  Output: [\(audioEmbeds.count / 1536), 1536]")
        print("  NaN: \(audioEmbeds.contains { $0.isNaN })")
        
        // ── Vision Test ───────────────────────────────────────────────────
        print("\nStep 3: Process Vision...")
        let numPatches = 256
        let patchDim = 768
        var patches: [Float] = []
        for _ in 0..<numPatches * patchDim {
            patches.append(Float.random(in: -0.5...0.5))
        }
        
        let visionStart = Date()
        // VisionTowerE2B forward pass placeholder (needs full implementation)
        let visionTime = Date().timeIntervalSince(visionStart) * 1000
        print("  ✓ Vision placeholder (forward not fully implemented)")
        print("  Time: \(visionTime) ms")
        
        // ── Text Generation ───────────────────────────────────────────────
        print("\nStep 4: Generate Text tokens...")
        var tokens: [Int] = [2]  // BOS
        let genStart = Date()
        for _ in 0..<10 {
            let logits = try mmModel.textModel.forward(tokenId: tokens.last!, position: tokens.count - 1)
            var maxIdx = 0
            var maxLogit = logits[0]
            for i in 1..<logits.count {
                if logits[i] > maxLogit {
                    maxLogit = logits[i]
                    maxIdx = i
                }
            }
            tokens.append(maxIdx)
        }
        let genTime = Date().timeIntervalSince(genStart) * 1000
        print("  ✓ Generated 10 tokens in \(genTime) ms")
        print("  Tokens: \(tokens.suffix(10))")
        
        print("\n═══════════════════════════════════════════════════════════════════")
        print("✓ E2B Full Multimodal test passed")
        print("  Load: \(loadTime) ms, Audio: \(audioTime) ms, Vision: \(visionTime) ms, Gen: \(genTime) ms")
        print("═══════════════════════════════════════════════════════════════════\n")
        
        XCTAssertFalse(audioEmbeds.contains { $0.isNaN }, "Audio should not have NaN")
    }
    
    func testE4BFullMultimodal() throws {
        print("\n═══════════════════════════════════════════════════════════════════")
        print("  E4B Full Multimodal Test (TEXT + Audio + Vision)")
        print("═══════════════════════════════════════════════════════════════════\n")
        
        let modelDir = "/Users/accusys/MarkBaseEngine/models/E4B-MarkBase"
        
        print("Step 1: Load MultimodalModel (TEXT + Audio + Vision)...")
        let engine = try MarkBaseEngine(autoCompile: true)
        let loadStart = Date()
        let mmModel = try MultimodalModel(modelDir: modelDir, engine: engine, maxContextLength: 256)
        let loadTime = Date().timeIntervalSince(loadStart) * 1000
        print("  ✓ Model loaded in \(loadTime) ms")
        print("  TEXT: hidden=\(mmModel.textModel.hiddenSize), layers=\(mmModel.textModel.numHiddenLayers)")
        print("  Audio: \(mmModel.audioTowerFull != nil ? "Full (12 layers)" : "N/A")")
        print("  Vision: \(mmModel.visionTowerFull != nil ? "Full (16 layers)" : "N/A")")
        
        // ── Audio Test ───────────────────────────────────────────────────
        print("\nStep 2: Process Audio...")
        let seqLen = 100
        let nMels = 128
        var melFeatures: [[Float]] = []
        for _ in 0..<seqLen {
            var frame: [Float] = []
            for _ in 0..<nMels {
                frame.append(Float.random(in: -0.5...0.5))
            }
            melFeatures.append(frame)
        }
        
        let audioStart = Date()
        let audioEmbeds = try mmModel.processAudio(audioFeatures: melFeatures)
        let audioTime = Date().timeIntervalSince(audioStart) * 1000
        print("  ✓ Audio processed in \(audioTime) ms")
        print("  Output: [\(audioEmbeds.count / 1536), 1536]")
        print("  NaN: \(audioEmbeds.contains { $0.isNaN })")
        
        // ── Vision Test ───────────────────────────────────────────────────
        print("\nStep 3: Process Vision...")
        let numPatches = 256
        let patchDim = 768
        var patches: [Float] = []
        for _ in 0..<numPatches * patchDim {
            patches.append(Float.random(in: -0.5...0.5))
        }
        
        let visionStart = Date()
        let visionEmbeds = try mmModel.processVision(patchEmbeddings: patches, numPatches: numPatches)
        let visionTime = Date().timeIntervalSince(visionStart) * 1000
        print("  ✓ Vision processed in \(visionTime) ms")
        print("  Output: [\(numPatches), \(visionEmbeds.count / numPatches)]")
        print("  NaN: \(visionEmbeds.contains { $0.isNaN })")
        
        // ── Text Generation ───────────────────────────────────────────────
        print("\nStep 4: Generate Text tokens...")
        var tokens: [Int] = [2]  // BOS
        let genStart = Date()
        for _ in 0..<10 {
            let logits = try mmModel.textModel.forward(tokenId: tokens.last!, position: tokens.count - 1)
            var maxIdx = 0
            var maxLogit = logits[0]
            for i in 1..<logits.count {
                if logits[i] > maxLogit {
                    maxLogit = logits[i]
                    maxIdx = i
                }
            }
            tokens.append(maxIdx)
        }
        let genTime = Date().timeIntervalSince(genStart) * 1000
        print("  ✓ Generated 10 tokens in \(genTime) ms")
        print("  Tokens: \(tokens.suffix(10))")
        
        print("\n═══════════════════════════════════════════════════════════════════")
        print("✓ E4B Full Multimodal test passed")
        print("  Load: \(loadTime) ms, Audio: \(audioTime) ms, Vision: \(visionTime) ms, Gen: \(genTime) ms")
        print("═══════════════════════════════════════════════════════════════════\n")
        
        XCTAssertFalse(audioEmbeds.contains { $0.isNaN }, "Audio should not have NaN")
        XCTAssertFalse(visionEmbeds.contains { $0.isNaN }, "Vision should not have NaN")
    }
    
    func test12BFullMultimodal() throws {
        print("\n═══════════════════════════════════════════════════════════════════")
        print("  12B Full Multimodal Test (TEXT + Audio + Vision)")
        print("═══════════════════════════════════════════════════════════════════\n")
        
        let modelDir = "/Users/accusys/.cache/huggingface/hub/models--mlx-community--gemma-4-12B-it-4bit/snapshots/73bcf09092aa277861d5a191b989b666f7f32e8f"
        
        print("Step 1: Load MultimodalModel (TEXT + Audio + Vision)...")
        let engine = try MarkBaseEngine(autoCompile: true)
        let loadStart = Date()
        let mmModel = try MultimodalModel(modelDir: modelDir, engine: engine, maxContextLength: 256)
        let loadTime = Date().timeIntervalSince(loadStart) * 1000
        print("  ✓ Model loaded in \(loadTime) ms")
        print("  TEXT: hidden=\(mmModel.textModel.hiddenSize), layers=\(mmModel.textModel.numHiddenLayers)")
        print("  Audio: \(mmModel.audioTower != nil ? "12B (projection)" : "N/A")")
        print("  Vision: \(mmModel.visionTower != nil ? "12B (simplified)" : "N/A")")
        
        // ── Audio Test ───────────────────────────────────────────────────
        print("\nStep 2: Process Audio (640-dim embeddings)...")
        let seqLen = 100
        let audioDim = 640
        var audioEmbedsInput: [Float] = []
        for _ in 0..<seqLen * audioDim {
            audioEmbedsInput.append(Float.random(in: -0.5...0.5))
        }
        
        let inputBuffer = engine.device.makeBuffer(bytes: audioEmbedsInput, length: audioEmbedsInput.count * 4)!
        let outputBuffer = engine.device.makeBuffer(length: seqLen * mmModel.textModel.hiddenSize * 4)!
        
        let audioStart = Date()
        if let tower = mmModel.audioTower {
            try tower.forward(inputBuffer: inputBuffer, seqLen: seqLen, outputBuffer: outputBuffer)
        }
        let audioTime = Date().timeIntervalSince(audioStart) * 1000
        print("  ✓ Audio processed in \(audioTime) ms")
        
        // ── Vision Test ───────────────────────────────────────────────────
        print("\nStep 3: Process Vision...")
        let numPatches = 256
        let patchDim = 768
        var patches: [Float] = []
        for _ in 0..<numPatches * patchDim {
            patches.append(Float.random(in: -0.5...0.5))
        }
        
        let visionStart = Date()
        let visionEmbeds = try mmModel.processVision(patchEmbeddings: patches, numPatches: numPatches)
        let visionTime = Date().timeIntervalSince(visionStart) * 1000
        print("  ✓ Vision processed in \(visionTime) ms")
        print("  Output: [\(numPatches), \(visionEmbeds.count / numPatches)]")
        print("  NaN: \(visionEmbeds.contains { $0.isNaN })")
        
        // ── Text Generation ───────────────────────────────────────────────
        print("\nStep 4: Generate Text tokens...")
        var tokens: [Int] = [2]  // BOS
        let genStart = Date()
        for _ in 0..<10 {
            let logits = try mmModel.textModel.forward(tokenId: tokens.last!, position: tokens.count - 1)
            var maxIdx = 0
            var maxLogit = logits[0]
            for i in 1..<logits.count {
                if logits[i] > maxLogit {
                    maxLogit = logits[i]
                    maxIdx = i
                }
            }
            tokens.append(maxIdx)
        }
        let genTime = Date().timeIntervalSince(genStart) * 1000
        print("  ✓ Generated 10 tokens in \(genTime) ms")
        print("  Tokens: \(tokens.suffix(10))")
        
        print("\n═══════════════════════════════════════════════════════════════════")
        print("✓ 12B Full Multimodal test passed")
        print("  Load: \(loadTime) ms, Audio: \(audioTime) ms, Vision: \(visionTime) ms, Gen: \(genTime) ms")
        print("═══════════════════════════════════════════════════════════════════\n")
        
        XCTAssertFalse(visionEmbeds.contains { $0.isNaN }, "Vision should not have NaN")
    }
}