updated applyChatTemplate with lazy memoization

Sources/Tokenizers/Tokenizer.swift

@@ -285,6 +285,9 @@ public class PreTrainedTokenizer: Tokenizer {
 
     private let cleanUpTokenizationSpaces: Bool
 
+    /// Cache for compiled Jinja templates keyed by their literal template string
+    private var compiledChatTemplateCache: [String: Template] = [:]
+
     public required init(tokenizerConfig: Config, tokenizerData: Config) throws {
         var addedTokens: [String: Int] = [:]
         var specialTokens: [String: Int] = [:]
@@ -332,6 +335,15 @@ public class PreTrainedTokenizer: Tokenizer {
         model = try TokenizerModel.from(tokenizerConfig: tokenizerConfig, tokenizerData: tokenizerData, addedTokens: addedTokens)
     }
 
+    private func compiledTemplate(for templateString: String) throws -> Template {
+        if let cached = compiledChatTemplateCache[templateString] {
+            return cached
+        }
+        let compiled = try Template(templateString)
+        compiledChatTemplateCache[templateString] = compiled
+        return compiled
+    }
+
     func preTokenize(_ text: String, options: PreTokenizerOptions) -> [String] {
         guard let preTokenizer else { return [text] }
         return preTokenizer(text: text, options: options)
@@ -530,7 +542,7 @@ public class PreTrainedTokenizer: Tokenizer {
             throw TokenizerError.missingChatTemplate
         }
 
-        let template = try Template(selectedChatTemplate)
+        let template = try compiledTemplate(for: selectedChatTemplate)
         var context: [String: Any] = [
             "messages": messages,
             "add_generation_prompt": addGenerationPrompt,

Tests/TokenizersTests/ChatTemplateTests.swift

@@ -5,6 +5,7 @@
 //  Created by Anthony DePasquale on 2/10/24.
 //
 
+import Foundation
 import Tokenizers
 import XCTest
 
@@ -277,4 +278,32 @@ class ChatTemplateTests: XCTestCase {
             }
         }
     }
+
+    /// Performance: cached vs uncached template application
+    func testApplyChatTemplatePerformanceCached() async throws {
+        let tokenizer = try await Self.sharedPhiTokenizer()
+
+        // Purposely reuse the same template literal to hit the memoized compiled template
+        let mistral7BDefaultTemplate = "{{bos_token}}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ ' [INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ ' ' + message['content'] + ' ' + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}"
+
+        // Prime cache once
+        _ = try tokenizer.applyChatTemplate(messages: messages, chatTemplate: mistral7BDefaultTemplate)
+
+        measure(metrics: [XCTClockMetric()]) {
+            _ = try! tokenizer.applyChatTemplate(messages: messages, chatTemplate: mistral7BDefaultTemplate)
+        }
+    }
+
+    /// Performance: simulate uncached runs by varying the template to bypass memoization
+    func testApplyChatTemplatePerformanceUncached() async throws {
+        let tokenizer = try await Self.sharedPhiTokenizer()
+
+        let baseTemplate = "{{bos_token}}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ ' [INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ ' ' + message['content'] + ' ' + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}"
+
+        measure(metrics: [XCTClockMetric()]) {
+            // Make the template string unique each iteration to force a fresh compilation
+            let uniqueTemplate = baseTemplate + "{# perf \(UUID().uuidString) #}"
+            _ = try! tokenizer.applyChatTemplate(messages: messages, chatTemplate: uniqueTemplate)
+        }
+    }
 }