✨ feat: Implement 'temperature' setting for model creativity control

Author: wrapss

package.json

@@ -51,6 +51,11 @@
             "default": 256,
             "description": "Max number of new tokens to be generated."
           },
+          "inference.temperature": {
+            "type": "number",
+            "default": 0.2,
+            "description": "Temperature of the model. Increasing the temperature will make the model answer more creatively."
+          },
           "inference.model": {
             "type": "string",
             "enum": [

src/prompts/autocomplete.ts

@@ -10,6 +10,7 @@ export async function autocomplete(args: {
     suffix: string,
     maxLines: number,
     maxTokens: number,
+    temperature: number,
     canceled?: () => boolean,
 }): Promise<string> {
 
@@ -19,7 +20,8 @@ export async function autocomplete(args: {
         prompt: adaptPrompt({ prefix: args.prefix, suffix: args.suffix, model: args.model }),
         raw: true,
         options: {
-            num_predict: args.maxTokens
+            num_predict: args.maxTokens,
+            temperature: args.temperature
         }
     };
 

src/prompts/provider.ts

@@ -67,6 +67,7 @@ export class PromptProvider implements vscode.InlineCompletionItemProvider {
                     let model = config.get('model') as string;
                     let maxLines = config.get('maxLines') as number;
                     let maxTokens = config.get('maxTokens') as number;
+                    let temperature = config.get('temperature') as number;
                     if (endpoint.endsWith('/')) {
                         endpoint = endpoint.slice(0, endpoint.length - 1);
                     }
@@ -102,6 +103,7 @@ export class PromptProvider implements vscode.InlineCompletionItemProvider {
                             model: model,
                             maxLines: maxLines,
                             maxTokens: maxTokens,
+                            temperature,
                             canceled: () => token.isCancellationRequested,
                         });
                         info(`AI completion completed: ${res}`);

src/test/suite/extension.test.ts

@@ -19,14 +19,16 @@ suite('Extension Test Suite', () => {
 		let model = 'codellama:7b-code-q4_K_S'; // Lightweight llm for tests
 		let maxLines = 16;
 		let maxTokens = 256;
+		let temperature = 0.2;
 		let prompt = 'fun main(): ';
 		let result = await autocomplete({
 			endpoint,
 			model,
 			prefix: prompt,
 			suffix: '',
 			maxLines,
-			maxTokens
+			maxTokens,
+			temperature
 		});
 		console.warn(result);
 	});