swift-cli/Sources/main.swift

import ArgumentParser
import CoreML
import Foundation
import Tokenizers

// MARK: - CLI Arguments

@main
struct PlapreCLI: AsyncParsableCommand {
    static var configuration = CommandConfiguration(
        commandName: "plapre-cli",
        abstract: "Plapre Pico CoreML TTS Pipeline",
        discussion: "Danish text-to-speech using CoreML models."
    )

    @Argument(help: "Text to synthesize")
    var text: String = "Hej, mit navn er Daniel."

    @Option(name: .shortAndLong, help: "Speaker voice (available: tor, ida, liv, ask, kaj)")
    var speaker: String = "tor"

    @Option(name: .shortAndLong, help: "Output WAV file path")
    var output: String = "output.wav"

    @Flag(name: .long, help: "Use int8 quantized model (smaller)")
    var int8 = false

    // MARK: - Run Pipeline

    mutating func run() async throws {
        print("Plapre Pico CoreML TTS Pipeline")
        print("================================\n")

        print("Text: \(text)")
        print("Speaker: \(speaker)")
        print("Output: \(output)\n")

        let pipelineStart = CFAbsoluteTimeGetCurrent()

        // Load speaker and tokenize
        let speakerEmb = try loadSpeaker(speaker)
        print("Loaded speaker embedding (\(speakerEmb.count) dims)")

        let tokenizer = try await measureAsync("Tokenizer load") {
            try await AutoTokenizer.from(modelFolder: PlapreConfig.repoRoot)
        }
        let textTokens = tokenizer.encode(text: text, addSpecialTokens: false).map { Int32($0) }
        print("Tokenized: \(textTokens.count) tokens: \(textTokens)")

        // Build input sequence: [EOS, text_marker, tokens..., audio_marker]
        var inputSeq: [Int32] =
            [PlapreConfig.eosToken, PlapreConfig.textMarkerToken] + textTokens + [
                PlapreConfig.audioMarkerToken
            ]
        let inputLen = inputSeq.count
        print("Input sequence: \(inputLen) tokens")

        while inputSeq.count < PlapreConfig.prefillSequenceLength {
            inputSeq.append(PlapreConfig.eosToken)
        }

        // Load RoPE tables
        print("\nLoading RoPE tables...")
        let ropeCosF32 = try loadRopeTable("rope_cos.npy")
        let ropeSinF32 = try loadRopeTable("rope_sin.npy")
        let ropeCos16: [Float16] = ropeCosF32.map { Float16($0) }
        let ropeSin16: [Float16] = ropeSinF32.map { Float16($0) }
        print("RoPE cos: \(ropeCos16.count) values, sin: \(ropeSin16.count) values")

        // Compile models
        print("\nCompiling models...")
        let decodeModel = try measure("Compile PlaprePico") {
            try compileModel(
                at: PlapreConfig.modelURL(for: "PlaprePico", useInt8: int8))
        }
        let kanadeModel = try measure("Compile KanadeDecoder") {
            try compileModel(
                at: PlapreConfig.modelURL(for: "KanadeDecoder", useInt8: false))
        }
        let vocoderModel = try measure("Compile Vocoder") {
            try compileModel(
                at: PlapreConfig.modelURL(for: "Vocoder", useInt8: false))
        }

        // Pre-allocate MLMultiArrays (performance-critical: single allocation, reused for all steps)
        let pInputIds = try! MLMultiArray(shape: [1, 1], dataType: .int32)
        let pCausalMask = try! MLMultiArray(
            shape: [1, 1, 1, NSNumber(value: PlapreConfig.maxContextLength)], dataType: .float16)
        let pCos = try! MLMultiArray(
            shape: [1, 1, 1, NSNumber(value: PlapreConfig.headDimension)], dataType: .float16)
        let pSin = try! MLMultiArray(
            shape: [1, 1, 1, NSNumber(value: PlapreConfig.headDimension)], dataType: .float16)
        let pUpdateMask = try! MLMultiArray(
            shape: [1, 1, NSNumber(value: PlapreConfig.maxContextLength), 1], dataType: .float16)
        let pSpeakerEmb = makeFloat16Array(
            speakerEmb, shape: [1, PlapreConfig.speakerEmbeddingDimension])
        let pIsSpeaker = try! MLMultiArray(shape: [1], dataType: .float16)

        // Initialize causal mask to all -inf
        pCausalMask.withUnsafeMutableBufferPointer(ofType: Float16.self) { ptr, _ in
            for i in 0..<PlapreConfig.maxContextLength { ptr[i] = Float16(-65504.0) }
        }

        let inputProvider = DecodeInputProvider([
            "input_ids": pInputIds,
            "causal_mask": pCausalMask,
            "cos": pCos,
            "sin": pSin,
            "update_mask": pUpdateMask,
            "speaker_embedding": pSpeakerEmb,
            "is_speaker_step": pIsSpeaker,
        ])

        // State for stateful model
        var state = decodeModel.makeState()
        var lastLogits = [Float16](repeating: 0, count: PlapreConfig.vocabSize)

        func runDecodeStep(token: Int32, pos: Int, isSpeaker: Bool = false) throws {
            pInputIds.withUnsafeMutableBufferPointer(ofType: Int32.self) { ptr, _ in ptr[0] = token
            }
            pCausalMask.withUnsafeMutableBufferPointer(ofType: Float16.self) { ptr, _ in
                ptr[pos] = Float16(0.0)
            }

            let ropeOffset = pos * PlapreConfig.headDimension
            pCos.withUnsafeMutableBufferPointer(ofType: Float16.self) { ptr, _ in
                for i in 0..<PlapreConfig.headDimension { ptr[i] = ropeCos16[ropeOffset + i] }
            }
            pSin.withUnsafeMutableBufferPointer(ofType: Float16.self) { ptr, _ in
                for i in 0..<PlapreConfig.headDimension { ptr[i] = ropeSin16[ropeOffset + i] }
            }
            pUpdateMask.withUnsafeMutableBufferPointer(ofType: Float16.self) { ptr, _ in
                if pos > 0 { ptr[pos - 1] = Float16(0.0) }
                ptr[pos] = Float16(1.0)
            }
            pIsSpeaker.withUnsafeMutableBufferPointer(ofType: Float16.self) { ptr, _ in
                ptr[0] = isSpeaker ? Float16(1.0) : Float16(0.0)
            }

            let output = try decodeModel.prediction(from: inputProvider, using: state)
            let arr = output.featureValue(for: "logits")!.multiArrayValue!
            arr.withUnsafeBufferPointer(ofType: Float16.self) { ptr in
                for i in 0..<PlapreConfig.vocabSize { lastLogits[i] = ptr[i] }
            }
        }

        // === Prefill ===
        print("\n--- Prefill (token-by-token) ---")
        let inputTokens = Array(inputSeq.prefix(inputLen))
        print("Processing \(inputTokens.count) input tokens...")

        let prefillStart = CFAbsoluteTimeGetCurrent()
        state = decodeModel.makeState()
        pCausalMask.withUnsafeMutableBufferPointer(ofType: Float16.self) { ptr, _ in
            for i in 0..<PlapreConfig.maxContextLength { ptr[i] = Float16(-65504.0) }
        }
        pUpdateMask.withUnsafeMutableBufferPointer(ofType: Float16.self) { ptr, _ in
            for i in 0..<PlapreConfig.maxContextLength { ptr[i] = Float16(0.0) }
        }
        try runDecodeStep(token: inputTokens[0], pos: 0, isSpeaker: true)

        for i in 1..<inputTokens.count {
            try runDecodeStep(token: inputTokens[i], pos: i, isSpeaker: false)
        }

        let prefillElapsed = CFAbsoluteTimeGetCurrent() - prefillStart
        let prefillTokPerSec = Double(inputTokens.count) / prefillElapsed
        print(
            "  ⏱ Prefill: \(formatTime(prefillElapsed)) (\(inputTokens.count) tokens, \(String(format: "%.1f", prefillTokPerSec)) tok/s)"
        )

        let firstToken = sampleFromLogits(
            lastLogits, temperature: PlapreConfig.defaultTemperature, topK: PlapreConfig.defaultTopK
        )
        print("First generated token: \(firstToken)")

        // === Autoregressive decode ===
        print("\n--- Decode ---")
        var generatedTokens: [Int32] = [firstToken]
        let maxTokens = min(
            PlapreConfig.maxGenerationTokens, PlapreConfig.maxContextLength - inputLen - 1)
        print("Generating up to \(maxTokens) tokens...")

        let decodeStart = CFAbsoluteTimeGetCurrent()
        var nextToken = firstToken
        var consecutiveNonAudio = 0

        for step in 1..<maxTokens {
            let pos = inputLen + step - 1
            try runDecodeStep(token: nextToken, pos: pos)
            nextToken = sampleFromLogits(
                lastLogits, temperature: PlapreConfig.defaultTemperature,
                topK: PlapreConfig.defaultTopK)
            generatedTokens.append(nextToken)

            if nextToken == PlapreConfig.eosToken {
                print("  EOS at step \(step)")
                break
            }

            if nextToken >= PlapreConfig.audioTokenOffset && nextToken <= PlapreConfig.audioTokenMax
            {
                consecutiveNonAudio = 0
            } else {
                consecutiveNonAudio += 1
                if consecutiveNonAudio >= PlapreConfig.nonAudioStopThreshold {
                    print(
                        "  Stopping: \(PlapreConfig.nonAudioStopThreshold) consecutive non-audio tokens at step \(step)"
                    )
                    break
                }
            }

            if step % 25 == 0 {
                let elapsed = CFAbsoluteTimeGetCurrent() - decodeStart
                let tokPerSec = Double(step) / elapsed
                print(
                    "  Step \(step) (\(Float(step) / Float(PlapreConfig.audioTokensPerSecond))s audio) — \(formatTime(elapsed)) elapsed, \(String(format: "%.1f", tokPerSec)) tok/s"
                )
            }
        }

        let decodeElapsed = CFAbsoluteTimeGetCurrent() - decodeStart
        let decodeSteps = generatedTokens.count - 1
        let decodeTokPerSec = Double(decodeSteps) / decodeElapsed
        let audioSeconds =
            Float(
                generatedTokens.filter {
                    $0 >= PlapreConfig.audioTokenOffset && $0 <= PlapreConfig.audioTokenMax
                }.count) / Float(PlapreConfig.audioTokensPerSecond)
        let rtf = Float(decodeElapsed) / audioSeconds

        print(
            "  ⏱ Decode: \(formatTime(decodeElapsed)) (\(decodeSteps) steps, \(String(format: "%.1f", decodeTokPerSec)) tok/s)"
        )
        print(
            "  ⏱ Audio generated: \(String(format: "%.1f", audioSeconds))s — RTF \(String(format: "%.2f", rtf))x (1.0 = realtime)"
        )

        // === Audio synthesis ===
        let audioTokens = generatedTokens.filter {
            $0 >= PlapreConfig.audioTokenOffset && $0 <= PlapreConfig.audioTokenMax
        }
        print(
            "\nGenerated \(generatedTokens.count) tokens, \(audioTokens.count) audio (\(Float(audioTokens.count) / Float(PlapreConfig.audioTokensPerSecond))s)"
        )

        guard !audioTokens.isEmpty else {
            throw PipelineError.noAudioTokensGenerated
        }

        // === Kanade + Vocoder in chunks ===
        let numChunks =
            (audioTokens.count + PlapreConfig.kanadeChunkSize - 1) / PlapreConfig.kanadeChunkSize
        print("\n--- Kanade + Vocoder (\(numChunks) chunk\(numChunks == 1 ? "" : "s")) ---")

        var waveform: [Float] = []
        let audioDecodeStart = CFAbsoluteTimeGetCurrent()

        for chunkIdx in 0..<numChunks {
            let chunkStart = CFAbsoluteTimeGetCurrent()
            let start = chunkIdx * PlapreConfig.kanadeChunkSize
            let end = min(start + PlapreConfig.kanadeChunkSize, audioTokens.count)
            let chunkTokens = Array(audioTokens[start..<end])

            var kanadeIndices = chunkTokens.map { $0 - Int32(PlapreConfig.audioTokenOffset) }
            let actualCount = kanadeIndices.count
            while kanadeIndices.count < PlapreConfig.kanadeChunkSize {
                kanadeIndices.append(kanadeIndices.last ?? 0)
            }

            // Kanade: tokens → mel
            let kanadeStart = CFAbsoluteTimeGetCurrent()
            let kanadeInput: [String: MLFeatureValue] = [
                "token_indices": .init(
                    multiArray: makeInt32Array(kanadeIndices, shape: [PlapreConfig.kanadeChunkSize])
                ),
                "speaker_embedding": .init(
                    multiArray: makeFloat32Array(
                        speakerEmb, shape: [1, PlapreConfig.speakerEmbeddingDimension])),
            ]
            let kanadeProvider = try MLDictionaryFeatureProvider(dictionary: kanadeInput)
            let kanadeOutput = try await kanadeModel.prediction(from: kanadeProvider)
            let mel = kanadeOutput.featureValue(for: "mel")!.multiArrayValue!
            let kanadeElapsed = CFAbsoluteTimeGetCurrent() - kanadeStart

            // Vocoder: mel → waveform
            let vocoderStart = CFAbsoluteTimeGetCurrent()
            let vocoderInput: [String: MLFeatureValue] = [
                "mel": .init(multiArray: mel)
            ]
            let vocoderProvider = try MLDictionaryFeatureProvider(dictionary: vocoderInput)
            let vocoderOutput = try await vocoderModel.prediction(from: vocoderProvider)
            let chunkWaveform = readFloat32Array(
                vocoderOutput.featureValue(for: "waveform")!.multiArrayValue!)
            let vocoderElapsed = CFAbsoluteTimeGetCurrent() - vocoderStart

            let samplesPerToken = chunkWaveform.count / PlapreConfig.kanadeChunkSize
            let usableSamples = actualCount * samplesPerToken
            waveform.append(contentsOf: chunkWaveform.prefix(usableSamples))

            let chunkElapsed = CFAbsoluteTimeGetCurrent() - chunkStart
            let chunkDuration = String(
                format: "%.1f", Float(usableSamples) / Float(PlapreConfig.sampleRate))
            print(
                "  Chunk \(chunkIdx + 1)/\(numChunks): \(actualCount) tokens → \(chunkDuration)s audio — Kanade \(formatTime(kanadeElapsed)), Vocoder \(formatTime(vocoderElapsed)), total \(formatTime(chunkElapsed))"
            )
        }

        let audioDecodeElapsed = CFAbsoluteTimeGetCurrent() - audioDecodeStart
        print(
            "  ⏱ Audio decode total: \(formatTime(audioDecodeElapsed)) (\(numChunks) chunk\(numChunks == 1 ? "" : "s"))"
        )

        print(
            "Total waveform: \(waveform.count) samples (\(String(format: "%.1f", Float(waveform.count) / Float(PlapreConfig.sampleRate)))s)"
        )

        // === Write WAV ===
        let outputURL = URL(fileURLWithPath: output)
        try writeWAV(waveform, to: outputURL)
        print("\nSaved to \(output)")

        // === Timing Summary ===
        let pipelineElapsed = CFAbsoluteTimeGetCurrent() - pipelineStart
        let totalAudioDuration = Float(waveform.count) / Float(PlapreConfig.sampleRate)
        print("\n========== Timing Summary ==========")
        print("  Total pipeline:    \(formatTime(pipelineElapsed))")
        print("  Audio output:      \(String(format: "%.1f", totalAudioDuration))s")
        print(
            "  Overall RTF:       \(String(format: "%.2f", Float(pipelineElapsed) / totalAudioDuration))x"
        )
        print("====================================")
        print("Done!")
    }
}

enum PipelineError: LocalizedError {
    case noAudioTokensGenerated

    var errorDescription: String? {
        switch self {
        case .noAudioTokensGenerated:
            return "No audio tokens were generated. The model may have failed to produce speech."
        }
    }
}