phi3 (#54)

awni · davidkoski · web-flow · commit b951b78eb235 · 2024-04-24T09:31:01.000-07:00
* phi3

Co-authored-by: David Koski &lt;dkoski@apple.com&gt;
diff --git a/Applications/LLMEval/ContentView.swift b/Applications/LLMEval/ContentView.swift
@@ -157,7 +157,7 @@ class LLMEvaluator {
 
     /// this controls which model loads -- phi4bit is one of the smaller ones so this will fit on
     /// more devices
-    let modelConfiguration = ModelConfiguration.phi4bit
+    let modelConfiguration = ModelConfiguration.phi34bit
 
     /// parameters controlling the output
     let generateParameters = GenerateParameters(temperature: 0.6)
diff --git a/Libraries/LLM/Configuration.swift b/Libraries/LLM/Configuration.swift
@@ -30,6 +30,7 @@ public enum ModelType: String, Codable {
     case mistral
     case llama
     case phi
+    case phi3
     case gemma
     case qwen2
     case starcoder2
@@ -45,6 +46,10 @@ public enum ModelType: String, Codable {
             let configuration = try JSONDecoder().decode(
                 PhiConfiguration.self, from: Data(contentsOf: configuration))
             return PhiModel(configuration)
+        case .phi3:
+            let configuration = try JSONDecoder().decode(
+                Phi3Configuration.self, from: Data(contentsOf: configuration))
+            return Phi3Model(configuration)
         case .gemma:
             let configuration = try JSONDecoder().decode(
                 GemmaConfiguration.self, from: Data(contentsOf: configuration))
diff --git a/Libraries/LLM/Evaluate.swift b/Libraries/LLM/Evaluate.swift
@@ -60,16 +60,16 @@ public struct GenerateParameters {
     public var temperature: Float = 0.6
 
     /// top p sampling
-    public var topP: Float = 0.9
+    public var topP: Float = 1.0
 
     /// penalty factor for repeating tokens
-    public var repetitionPenalty: Float = 1.0
+    public var repetitionPenalty: Float?
 
     /// number of tokens to consider for repetition penalty
     public var repetitionContextSize: Int = 20
 
     public init(
-        temperature: Float = 0.6, topP: Float = 0.9, repetitionPenalty: Float = 1.0,
+        temperature: Float = 0.6, topP: Float = 1.0, repetitionPenalty: Float? = nil,
         repetitionContextSize: Int = 20
     ) {
         self.temperature = temperature
@@ -111,11 +111,11 @@ public struct TokenIterator: Sequence, IteratorProtocol {
         var logits: MLXArray
         (logits, cache) = model(expandedDimensions(y, axis: 0), cache: cache.isEmpty ? nil : cache)
         logits = logits[0..., -1, 0...]
-        if parameters.repetitionPenalty > 1.0 {
+        if let repetitionPenalty = parameters.repetitionPenalty {
             // apply repetition penalty
             logits = applyRepetitionPenalty(
                 logits: logits, repetitionContext: repetitionContext,
-                penalty: parameters.repetitionPenalty)
+                penalty: repetitionPenalty)
         }
         y = sample(logits: logits, temp: parameters.temperature, topP: parameters.topP)
         // append the current token to the context and check repetitionPenalty context see if need to remove the first token
diff --git a/Libraries/LLM/Models.swift b/Libraries/LLM/Models.swift
@@ -116,6 +116,13 @@ extension ModelConfiguration {
         "Instruct: \(prompt)\nOutput: "
     }
 
+    public static let phi34bit = ModelConfiguration(
+        id: "mlx-community/Phi-3-mini-4k-instruct-4bit-no-q-embed"
+    ) {
+        prompt in
+        "<s><|user|>\n\(prompt)<|end|>\n<|assistant|>\n"
+    }
+
     public static let gemma2bQuantized = ModelConfiguration(
         id: "mlx-community/quantized-gemma-2b-it",
         overrideTokenizer: "PreTrainedTokenizer"
@@ -146,6 +153,7 @@ extension ModelConfiguration {
                 mistral7B4bit,
                 codeLlama13b4bit,
                 phi4bit,
+                phi34bit,
                 gemma2bQuantized,
                 qwen205b4bit,
             ])
diff --git a/Libraries/LLM/Phi3.swift b/Libraries/LLM/Phi3.swift
@@ -0,0 +1,257 @@
+// Copyright © 2024 Apple Inc.
+
+import Foundation
+import MLX
+import MLXFast
+import MLXNN
+
+private class Attention: Module {
+
+    let args: Phi3Configuration
+    let scale: Float
+
+    @ModuleInfo(key: "qkv_proj") var wqkv: Linear
+    @ModuleInfo(key: "o_proj") var wo: Linear
+
+    let rope: RoPE
+
+    public init(_ args: Phi3Configuration) {
+        self.args = args
+
+        let dim = args.hiddenSize
+        let heads = args.attentionHeads
+        let kvHeads = args.kvHeads
+
+        let headDim = args.hiddenSize / heads
+        self.scale = pow(Float(headDim), -0.5)
+
+        self._wqkv.wrappedValue = Linear(dim, (heads + 2 * 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))
+
+        let qkv = split(wqkv(x), parts: 3, axis: -1)
+        var queries = qkv[0]
+        var keys = qkv[1]
+        var values = qkv[2]
+
+        // 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 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)
+        }
+
+        let output = MLXFast.scaledDotProductAttention(
+            queries: queries, keys: keys, values: values, scale: scale, mask: mask
+        )
+        .transposed(0, 2, 1, 3)
+        .reshaped(B, L, -1)
+
+        return (wo(output), (keys, values))
+    }
+}
+
+private class MLP: Module, UnaryLayer {
+
+    @ModuleInfo(key: "gate_up_proj") var gate_up: Linear
+    @ModuleInfo(key: "down_proj") var down: Linear
+
+    public init(dimensions: Int, hiddenDimensions: Int) {
+        self._gate_up.wrappedValue = Linear(dimensions, 2 * hiddenDimensions, bias: false)
+        self._down.wrappedValue = Linear(hiddenDimensions, dimensions, bias: false)
+    }
+
+    public func callAsFunction(_ x: MLXArray) -> MLXArray {
+        let gu = split(gate_up(x), parts: 2, axis: -1)
+        return down(silu(gu[0]) * gu[1])
+    }
+}
+
+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: Phi3Configuration) {
+        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 Phi3ModelInner: Module {
+
+    @ModuleInfo(key: "embed_tokens") var embedTokens: Embedding
+
+    fileprivate let layers: [TransformerBlock]
+    let norm: RMSNorm
+
+    public init(_ args: Phi3Configuration) {
+        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 Phi3Model: Module, LLMModel {
+
+    public let vocabularySize: Int
+    let model: Phi3ModelInner
+
+    @ModuleInfo(key: "lm_head") var lmHead: Linear
+
+    public init(_ args: Phi3Configuration) {
+        self.vocabularySize = args.vocabularySize
+        self.model = Phi3ModelInner(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 Phi3Configuration: 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<Phi3Configuration.CodingKeys> =
+            try decoder.container(
+                keyedBy: Phi3Configuration.CodingKeys.self)
+
+        self.hiddenSize = try container.decode(
+            Int.self, forKey: Phi3Configuration.CodingKeys.hiddenSize)
+        self.hiddenLayers = try container.decode(
+            Int.self, forKey: Phi3Configuration.CodingKeys.hiddenLayers)
+        self.intermediateSize = try container.decode(
+            Int.self, forKey: Phi3Configuration.CodingKeys.intermediateSize)
+        self.attentionHeads = try container.decode(
+            Int.self, forKey: Phi3Configuration.CodingKeys.attentionHeads)
+        self.rmsNormEps = try container.decode(
+            Float.self, forKey: Phi3Configuration.CodingKeys.rmsNormEps)
+        self.vocabularySize = try container.decode(
+            Int.self, forKey: Phi3Configuration.CodingKeys.vocabularySize)
+        self.kvHeads = try container.decode(Int.self, forKey: Phi3Configuration.CodingKeys.kvHeads)
+        self.ropeTheta =
+            try container.decodeIfPresent(
+                Float.self, forKey: Phi3Configuration.CodingKeys.ropeTheta)
+            ?? 10_000
+        self.ropeTraditional =
+            try container.decodeIfPresent(
+                Bool.self, forKey: Phi3Configuration.CodingKeys.ropeTraditional) ?? false
+        self.ropeScaling = try container.decodeIfPresent(
+            [String: StringOrNumber].self, forKey: Phi3Configuration.CodingKeys.ropeScaling)
+
+    }
+}
+
+// MARK: - LoRA
+
+extension Phi3Model: LoRAModel {
+    public func loraLinearLayers() -> LoRALinearLayers {
+        model.layers.map { ($0.attention, ["qkv_proj"]) }
+    }
+}
diff --git a/Tools/llm-tool/LLMTool.swift b/Tools/llm-tool/LLMTool.swift
@@ -53,10 +53,10 @@ struct GenerateArguments: ParsableArguments {
     var temperature: Float = 0.6
 
     @Option(name: .long, help: "The top p sampling")
-    var topP: Float = 0.9
+    var topP: Float = 1.0
 
     @Option(name: .long, help: "The penalty factor for repeating tokens")
-    var repetitionPenalty: Float = 1.0
+    var repetitionPenalty: Float?
 
     @Option(name: .long, help: "The number of tokens to consider for repetition penalty")
     var repetitionContextSize: Int = 20
diff --git a/mlx-swift-examples.xcodeproj/project.pbxproj b/mlx-swift-examples.xcodeproj/project.pbxproj
