mlx-swift-examples/Libraries/LLM/Llama.swift

// Copyright © 2024 Apple Inc.

import Foundation
import MLX
import MLXNN

// port of https://github.com/ml-explore/mlx-examples/blob/main/llms/mlx_lm/models/llama.py

private class Attention: Module {

    let args: LlamaConfiguration
    let repeats: Int
    let scale: Float

    @ModuleInfo(key: "q_proj") var wq: Linear
    @ModuleInfo(key: "k_proj") var wk: Linear
    @ModuleInfo(key: "v_proj") var wv: Linear
    @ModuleInfo(key: "o_proj") var wo: Linear

    let rope: RoPE

    public init(_ args: LlamaConfiguration) {
        self.args = args

        let dim = args.hiddenSize
        let heads = args.attentionHeads
        let kvHeads = args.kvHeads

        self.repeats = heads / kvHeads

        let headDim = args.hiddenSize / heads
        self.scale = pow(Float(headDim), -0.5)

        self._wq.wrappedValue = Linear(dim, heads * headDim, bias: false)
        self._wk.wrappedValue = Linear(dim, kvHeads * headDim, bias: false)
        self._wv.wrappedValue = Linear(dim, kvHeads * headDim, bias: false)
        self._wo.wrappedValue = Linear(heads * headDim, dim, bias: false)

        let ropeScale: Float
        if let ropeScaling = args.ropeScaling, ropeScaling["type"] == .string("linear"),
            let factor = ropeScaling["factor"]
        {
            switch factor {
            case .string:
                fatalError("ropeScaling.factor must be a float")
            case .float(let v):
                ropeScale = 1 / v
            }
        } else {
            ropeScale = 1
        }

        self.rope = RoPE(
            dimensions: headDim, traditional: args.ropeTraditional, base: args.ropeTheta,
            scale: ropeScale)
    }

    public func callAsFunction(
        _ x: MLXArray, mask: MLXArray? = nil, cache: (MLXArray, MLXArray)? = nil
    ) -> (MLXArray, (MLXArray, MLXArray)) {
        let (B, L) = (x.dim(0), x.dim(1))

        var queries = wq(x)
        var keys = wk(x)
        var values = wv(x)

        // prepare the queries, keys and values for the attention computation
        queries = queries.reshaped(B, L, args.attentionHeads, -1).transposed(0, 2, 1, 3)
        keys = keys.reshaped(B, L, args.kvHeads, -1).transposed(0, 2, 1, 3)
        values = values.reshaped(B, L, args.kvHeads, -1).transposed(0, 2, 1, 3)

        if repeats > 1 {
            keys = MLXArray.repeat(keys, count: repeats, axis: 1)
            values = MLXArray.repeat(values, count: repeats, axis: 1)
        }

        if let (keyCache, valueCache) = cache {
            queries = rope(queries, offset: keyCache.dim(2))
            keys = rope(keys, offset: keyCache.dim(2))
            keys = concatenated([keyCache, keys], axis: 2)
            values = concatenated([valueCache, values], axis: 2)
        } else {
            queries = rope(queries)
            keys = rope(keys)
        }

        var scores = (queries * self.scale).matmul(keys.transposed(0, 1, 3, 2))
        if let mask {
            scores = scores + mask
        }

        scores = softMax(scores.asType(.float32), axis: -1).asType(scores.dtype)

        let output = matmul(scores, values).transposed(0, 2, 1, 3).reshaped(B, L, -1)

        return (wo(output), (keys, values))
    }
}

private class MLP: Module, UnaryLayer {

    @ModuleInfo(key: "gate_proj") var gate: Linear
    @ModuleInfo(key: "down_proj") var down: Linear
    @ModuleInfo(key: "up_proj") var up: Linear

    public init(dimensions: Int, hiddenDimensions: Int) {
        self._gate.wrappedValue = Linear(dimensions, hiddenDimensions, bias: false)
        self._down.wrappedValue = Linear(hiddenDimensions, dimensions, bias: false)
        self._up.wrappedValue = Linear(dimensions, hiddenDimensions, bias: false)
    }

    public func callAsFunction(_ x: MLXArray) -> MLXArray {
        down(silu(gate(x)) * up(x))
    }
}

private class TransformerBlock: Module {

    @ModuleInfo(key: "self_attn") var attention: Attention
    let mlp: MLP

    @ModuleInfo(key: "input_layernorm") var inputLayerNorm: RMSNorm
    @ModuleInfo(key: "post_attention_layernorm") var postAttentionLayerNorm: RMSNorm

    public init(_ args: LlamaConfiguration) {
        self._attention.wrappedValue = Attention(args)
        self.mlp = MLP(dimensions: args.hiddenSize, hiddenDimensions: args.intermediateSize)
        self._inputLayerNorm.wrappedValue = RMSNorm(
            dimensions: args.hiddenSize, eps: args.rmsNormEps)
        self._postAttentionLayerNorm.wrappedValue = RMSNorm(
            dimensions: args.hiddenSize, eps: args.rmsNormEps)
    }

    public func callAsFunction(
        _ x: MLXArray, mask: MLXArray? = nil, cache: (MLXArray, MLXArray)? = nil
    ) -> (MLXArray, (MLXArray, MLXArray)) {
        var (r, cache) = attention(inputLayerNorm(x), mask: mask, cache: cache)
        let h = x + r
        r = mlp(postAttentionLayerNorm(h))
        let out = h + r
        return (out, cache)
    }
}

public class LlamaModelInner: Module {

    @ModuleInfo(key: "embed_tokens") var embedTokens: Embedding

    fileprivate let layers: [TransformerBlock]
    let norm: RMSNorm

    public init(_ args: LlamaConfiguration) {
        precondition(args.vocabularySize > 0)

        self._embedTokens.wrappedValue = Embedding(
            embeddingCount: args.vocabularySize, dimensions: args.hiddenSize)

        self.layers = (0 ..< args.hiddenLayers)
            .map { _ in
                TransformerBlock(args)
            }
        self.norm = RMSNorm(dimensions: args.hiddenSize, eps: args.rmsNormEps)
    }

    public func callAsFunction(_ inputs: MLXArray, cache: [(MLXArray, MLXArray)]? = nil) -> (
        MLXArray, [(MLXArray, MLXArray)]
    ) {
        var h = embedTokens(inputs)

        var mask: MLXArray? = nil
        if h.dim(1) > 1 {
            mask = MultiHeadAttention.createAdditiveCausalMask(h.dim(1))
            mask = mask?.asType(h.dtype)
        }

        var newCache = [(MLXArray, MLXArray)]()

        for (i, layer) in layers.enumerated() {
            var cacheUpdate: (MLXArray, MLXArray)
            (h, cacheUpdate) = layer(h, mask: mask, cache: cache?[i])
            newCache.append(cacheUpdate)
        }

        return (norm(h), newCache)
    }
}

public class LlamaModel: Module, LLMModel {

    public let vocabularySize: Int
    let model: LlamaModelInner

    @ModuleInfo(key: "lm_head") var lmHead: Linear

    public init(_ args: LlamaConfiguration) {
        self.vocabularySize = args.vocabularySize
        self.model = LlamaModelInner(args)
        self._lmHead.wrappedValue = Linear(args.hiddenSize, args.vocabularySize, bias: false)
    }

    public func callAsFunction(_ inputs: MLXArray, cache: [(MLXArray, MLXArray)]?) -> (
        MLXArray, [(MLXArray, MLXArray)]
    ) {
        let (out, cache) = model(inputs, cache: cache)
        return (lmHead(out), cache)
    }
}

public struct LlamaConfiguration: Codable {

    var hiddenSize: Int
    var hiddenLayers: Int
    var intermediateSize: Int
    var attentionHeads: Int
    var rmsNormEps: Float
    var vocabularySize: Int
    var kvHeads: Int
    var ropeTheta: Float = 10_000
    var ropeTraditional: Bool = false
    var ropeScaling: [String: StringOrNumber]? = nil

    enum CodingKeys: String, CodingKey {
        case hiddenSize = "hidden_size"
        case hiddenLayers = "num_hidden_layers"
        case intermediateSize = "intermediate_size"
        case attentionHeads = "num_attention_heads"
        case rmsNormEps = "rms_norm_eps"
        case vocabularySize = "vocab_size"
        case kvHeads = "num_key_value_heads"
        case ropeTheta = "rope_theta"
        case ropeTraditional = "rope_traditional"
        case ropeScaling = "rope_scaling"
    }

    public init(from decoder: Decoder) throws {
        // custom implementation to handle optional keys with required values
        let container: KeyedDecodingContainer<LlamaConfiguration.CodingKeys> =
            try decoder.container(
                keyedBy: LlamaConfiguration.CodingKeys.self)

        self.hiddenSize = try container.decode(
            Int.self, forKey: LlamaConfiguration.CodingKeys.hiddenSize)
        self.hiddenLayers = try container.decode(
            Int.self, forKey: LlamaConfiguration.CodingKeys.hiddenLayers)
        self.intermediateSize = try container.decode(
            Int.self, forKey: LlamaConfiguration.CodingKeys.intermediateSize)
        self.attentionHeads = try container.decode(
            Int.self, forKey: LlamaConfiguration.CodingKeys.attentionHeads)
        self.rmsNormEps = try container.decode(
            Float.self, forKey: LlamaConfiguration.CodingKeys.rmsNormEps)
        self.vocabularySize = try container.decode(
            Int.self, forKey: LlamaConfiguration.CodingKeys.vocabularySize)
        self.kvHeads = try container.decode(Int.self, forKey: LlamaConfiguration.CodingKeys.kvHeads)
        self.ropeTheta =
            try container.decodeIfPresent(
                Float.self, forKey: LlamaConfiguration.CodingKeys.ropeTheta)
            ?? 10_000
        self.ropeTraditional =
            try container.decodeIfPresent(
                Bool.self, forKey: LlamaConfiguration.CodingKeys.ropeTraditional) ?? false
        self.ropeScaling = try container.decodeIfPresent(
            [String: StringOrNumber].self, forKey: LlamaConfiguration.CodingKeys.ropeScaling)

    }
}