stage changes

package.json

@@ -724,6 +724,16 @@
       {
         "command": "devchat.fix_chinese",
         "title": "Devchat:修复此"
+      },
+			{
+				"command": "DevChat.codecomplete",
+				"title": "Codecomplete",
+				"category": "DevChat"
+			},
+      {
+        "command": "DevChat.codecomplete_callback",
+        "title": "Codecomplete Callback",
+        "category": "DevChat"
       }
     ],
     "keybindings": [
@@ -731,6 +741,11 @@
         "command": "devchat.openChatPanel",
         "key": "ctrl+shift+/",
         "mac": "cmd+shift+/"
+      },
+      {
+				"command": "DevChat.codecomplete",
+				"key": "ctrl+shift+'",
+				"mac": "cmd+shift+'"
 			}
     ],
     "menus": {

src/contributes/codecomplete/README.md

@@ -0,0 +1,3 @@
+
+status.ts: 代码补全状态表达接口。预期有三种状态：未就绪、就绪、代码补全中。
+

src/contributes/codecomplete/cache.ts

@@ -0,0 +1,74 @@
+/*
+ 使用内存存储最近的代码补全结果
+*/
+
+type CacheItem = {
+    value: any;
+    timestamp: number;
+};
+
+class MemoryCacheManager {
+    private static maxCapacity: number = 5;
+    private cache: Map<string, CacheItem>;
+
+    constructor() {
+        this.cache = new Map<string, CacheItem>();
+    }
+
+    /**
+     * 添加或更新缓存
+     */
+    set(key: string, value: any): void {
+        // 首先检查缓存中是否已经有了该键值对，若有，则更新；若没有，则添加
+        if (this.cache.has(key)) {
+            this.cache.set(key, { value, timestamp: Date.now() });
+        } else {
+            // 先确保缓存没有超出最大容量
+            if (this.cache.size >= MemoryCacheManager.maxCapacity) {
+                this.evict();
+            }
+            this.cache.set(key, { value, timestamp: Date.now() });
+        }
+    }
+
+    /**
+     * 获取缓存
+     */
+    get(key: string): any | undefined {
+        const item = this.cache.get(key);
+        if (item) {
+            // 更新timestamp以反映最近一次访问
+            item.timestamp = Date.now();
+            return item.value;
+        }
+        return undefined;
+    }
+
+    /**
+     * 删除指定的缓存项
+     */
+    delete(key: string): boolean {
+        return this.cache.delete(key);
+    }
+
+    /**
+     * 依据时间顺序（最久未使用）删除缓存项
+     */
+    private evict(): void {
+        let oldestKey: string | null = null;
+        let oldestTimestamp: number = Infinity;
+
+        for (const [key, item] of this.cache.entries()) {
+            if (item.timestamp < oldestTimestamp) {
+                oldestTimestamp = item.timestamp;
+                oldestKey = key;
+            }
+        }
+
+        if (oldestKey !== null) {
+            this.cache.delete(oldestKey);
+        }
+    }
+}
+
+export default MemoryCacheManager;

src/contributes/codecomplete/chunkFilter.ts

@@ -0,0 +1,148 @@
+/*
+ 对实时返回的chunk进行过滤，判断当前是否已经满足停止条件，避免无效代码占用补全时间
+*/
+import * as vscode from 'vscode';
+
+import { logger } from '../../util/logger';
+import { CodeCompletionChunk, streamComplete } from './llm';
+
+
+// 代码补全返回结果定义
+export interface CodeCompleteResult {
+    prompt: string;
+    code: string;
+    id: string;
+}
+
+
+export class LLMStreamComplete {
+    private token: vscode.CancellationToken;
+    private curlineIndent: number = 0;
+    constructor(token: vscode.CancellationToken, curlineIndent: number) {
+        this.token = token;
+        this.curlineIndent = curlineIndent;
+    }
+
+    async * chunkStopCanceled(chunks: AsyncIterable<CodeCompletionChunk>) {
+        for await (const chunk of chunks) {
+            if (this.token.isCancellationRequested) {
+                break;
+            }
+            yield chunk;
+        }
+    }
+
+    // 过滤第一个就是换行符的chunk，避免补全时出现空行
+    async * stopWhenFirstCharIsNewLine(chunks: AsyncIterable<CodeCompletionChunk>) {
+        let isFirst = true;
+        for await (const chunk of chunks) {
+            if (chunk.text.length === 0) {
+                yield chunk;
+            }
+
+            if (isFirst && chunk.text[0] === "\n") {
+                break;
+            }
+            isFirst = false;
+            yield chunk;
+        }
+    }
+
+    // 当前chunk中字符串不是以行为单位，需要重新整合为以行为单位。
+    async * toLines(chunks: AsyncIterable<CodeCompletionChunk>) {
+        let line = "";
+        let id = "";
+        for await (const chunk of chunks) {
+            if (chunk.id) {
+                id = chunk.id;
+            }
+
+            line += chunk.text;
+            while (line.indexOf("\n") !== -1) {
+                const index = line.indexOf("\n");
+                yield {
+                    text: line.slice(0, index + 1),
+                    id
+                };
+                line = line.slice(index + 1);
+            }
+        }
+
+        if (line.length > 0) {
+            yield { text: line, id };
+        }
+    }
+
+    // async * stopAtLineEnd(chunks: AsyncIterable<CodeCompletionChunk>) {
+    //     for await (const chunk of chunks) {
+    //         if (chunk.text.indexOf("\n") !== -1) {
+    //             chunk.text = chunk.text.slice(0, chunk.text.indexOf("\n"));
+    //             yield chunk;
+    //             break;
+    //         }
+    //         yield chunk;
+    //     }
+    // }
+
+    async * stopAtSameBlock(chunks: AsyncIterable<CodeCompletionChunk>) {
+        let index = 0;
+        let preIndent = -1;
+        let hasIndentBigger = false;
+        let sameIndentTimes = 0;
+        for await (const chunk of chunks) {
+            let lineIndent = chunk.text.search(/\S/);
+            if (index === 0) {
+                lineIndent = this.curlineIndent;
+            }
+
+            if (index > 0 && chunk.text.trim().length > 0 && lineIndent < this.curlineIndent ) {
+                break;
+            }
+            if (index > 0 && preIndent === 0 && lineIndent === 0) {
+                break;
+            }
+            if (index > 0 && hasIndentBigger && lineIndent === this.curlineIndent && chunk.text.trim().length > 3) {
+                break;
+            }
+            if (index > 0 && preIndent === lineIndent) {
+                sameIndentTimes += 1;
+            } else {
+                sameIndentTimes = 0;
+            }
+
+            if (sameIndentTimes > 1) {
+                break;
+            }
+            if (lineIndent > this.curlineIndent) {
+                hasIndentBigger = true;
+            }
+
+            preIndent = lineIndent;
+
+            index += 1;
+            yield chunk;
+        }
+    }
+
+    async llmStreamComplete(prompt: string) : Promise<CodeCompleteResult | undefined> {
+        // TODO
+        // 对LLM的异常进行捕获，避免中断代码补全
+
+        const chunks = streamComplete(prompt);
+        const chunks2 = this.chunkStopCanceled(chunks);
+        const chunks3 = this.toLines(chunks2);
+        const chunks4 = this.stopAtSameBlock(chunks3);
+
+        let id = "";
+        let completionCode = "";
+        for await (const chunk of chunks4) {
+            completionCode += chunk.text;
+            if (chunk.id) {
+                id = chunk.id;
+            }
+        }
+
+        logger.channel()?.info("code:", completionCode);
+        return { prompt, code: completionCode, id };
+    }
+}

src/contributes/codecomplete/codecomplete.ts

@@ -0,0 +1,154 @@
+import * as vscode from 'vscode';
+
+import { logger } from '../../util/logger';
+import Debouncer from './debouncer';
+import MemoryCacheManager from './cache';
+import { createPrompt } from './promptCreator';
+import { CodeCompleteResult, LLMStreamComplete } from './chunkFilter';
+import { nvidiaStarcoderComplete } from './llm';
+
+
+export function registerCodeCompleteCallbackCommand(context: vscode.ExtensionContext) {
+    let disposable = vscode.commands.registerCommand(
+        "DevChat.codecomplete_callback",
+        async (callback: any) => {
+            callback();
+        }
+    );
+
+    context.subscriptions.push(disposable);
+}
+
+export class InlineCompletionProvider implements vscode.InlineCompletionItemProvider {  
+    private debouncer: Debouncer;
+    private cache: MemoryCacheManager;
+
+    constructor() {
+        // TODO
+        // Read delay time from config
+        this.debouncer = new Debouncer(500);
+        this.cache = new MemoryCacheManager();
+    }
+
+    async codeComplete(document: vscode.TextDocument, position: vscode.Position, context: vscode.InlineCompletionContext, token: vscode.CancellationToken): Promise<CodeCompleteResult | undefined> {
+        // TODO
+        // create prompt
+        const fsPath = document.uri.fsPath;
+        const fileContent = document.getText();
+        const prompt = await createPrompt(fsPath, fileContent, position.line, position.character);
+
+        // check cache
+        const result = await this.cache.get(prompt);
+        if(result) {
+            return result;
+        }
+
+        // TODO
+        // call code_completion
+        const lines = fileContent.split('\n');
+        let curlineIndent = lines[position.line].search(/\S/);
+        if (curlineIndent === -1) {
+            curlineIndent = 0;
+        }
+        const completor = new LLMStreamComplete(token, curlineIndent);
+        const response = await completor.llmStreamComplete(prompt);
+        if (!response || response.code.length === 0) {
+            return undefined;
+        }
+
+        if (token.isCancellationRequested) {
+            return undefined;
+        }
+
+        // cache result
+        this.cache.set(prompt, response);
+        return response;
+    }
+
+    async provideInlineCompletionItems(document: vscode.TextDocument, position: vscode.Position, context: vscode.InlineCompletionContext, token: vscode.CancellationToken): Promise<vscode.InlineCompletionItem[] | null> {
+        const result = await this.debouncer.debounce();
+        if(!result) {
+            return [];
+        }
+
+        const response: CodeCompleteResult | undefined = await this.codeComplete(document, position, context, token);
+        if(!response) {
+            return [];
+        }
+
+        if (token.isCancellationRequested) {
+            return [];
+        }
+
+        // TODO
+        // 代码补全建议是否已经被用户看到，这个需要更加准确的方式来识别。
+        logger.channel()?.info("code complete show.");
+
+        const logRejectionTimeout: NodeJS.Timeout = setTimeout(() => {
+            logger.channel()?.info("code complete not accept.");
+        }, 10_000);
+
+        // 代码补全回调处理
+        const callback = () => {
+            logger.channel()?.info("accept:", response.id);
+            // delete cache
+            this.cache.delete(response.prompt);
+            // delete timer
+            clearTimeout(logRejectionTimeout);
+        };
+
+        return [
+            new vscode.InlineCompletionItem(
+                response.code,
+                new vscode.Range(
+                    position,
+                    position
+                ),
+                {
+                    title: "code complete accept",
+                    command: "DevChat.codecomplete_callback",
+                    arguments: [callback],
+                }
+            ),
+        ];
+
+        // // 等待时间（单位：毫秒），可根据需要调整
+        // const delayTime = 5000;
+
+        // // 创建一个新的Promise，用于实现等待逻辑
+        // await new Promise((resolve) => {
+        //     const timer = setTimeout(resolve, delayTime);
+            
+        //     // 如果请求在等待时间结束前被取消，则清除定时器
+        //     token.onCancellationRequested(() => clearTimeout(timer));
+        // });
+		// logger.channel()?.info("----->");
+
+        // // 如果请求已被取消，则直接返回null
+        // if (token.isCancellationRequested) {
+        //     logger.channel()?.info("request cancelled before completion");
+        //     return [];
+        // }
+
+		// // 根据文档和位置计算补全项（这里仅作示例，实际实现可能会有所不同） 
+		// // 获取position前文本
+		// const documentText = document.getText();
+		// const offsetPos = document.offsetAt(position);
+
+		// // 获取position前文本
+		// const prefix = documentText.substring(0, offsetPos);
+		// const suffix = documentText.substring(offsetPos);
+
+		// const prompt = "<fim_prefix>" + prefix + "<fim_suffix>" + suffix + "<fim_middle>";
+
+		// // call code_completion
+		// const response = await code_completion(prompt);
+		// if (!response) {
+		// 	logger.channel()?.info("finish provideInlineCompletionItems");
+		// 	return [];
+		// }
+
+		// logger.channel()?.info("finish provideInlineCompletionItems");
+		// return [new vscode.InlineCompletionItem(response[0], new vscode.Range(position, position))];  
+    }  
+}

src/contributes/codecomplete/debouncer.ts

@@ -0,0 +1,31 @@
+export class Debouncer {
+    private debouncing = false;
+    private debounceTimeout?: NodeJS.Timeout;
+    private lastTimeStampt?: string;
+
+    constructor(private debounceDelay: number) { }
+
+    async debounce(): Promise<boolean> {
+        const timestampt = Date.now().toString();
+        this.lastTimeStampt = timestampt;
+        
+        if (this.debouncing) {
+            this.debounceTimeout?.refresh();
+            const lastTimestampt = await new Promise<string | undefined>((resolve) =>
+                setTimeout(() => {
+                    resolve(this.lastTimeStampt);
+                }, this.debounceDelay)
+            );
+            return timestampt === lastTimestampt;
+        } else {
+            this.debouncing = true;
+            this.lastTimeStampt = timestampt;
+            this.debounceTimeout = setTimeout(() => {
+                this.debouncing = false;
+            }, this.debounceDelay);
+            return true;
+        }
+    }
+}
+
+export default Debouncer;

src/contributes/codecomplete/llm.ts

@@ -0,0 +1,86 @@
+
+/*
+ 通过LLM模型生成代码补全
+*/
+import axios from 'axios';
+
+import { logger } from "../../util/logger";
+import { Chunk } from 'webpack';
+
+
+// 定义代码补全chunk结构内容
+export interface CodeCompletionChunk {
+    text: string;
+    id: string;
+}
+
+export async function* streamComplete(prompt: string): AsyncGenerator<CodeCompletionChunk> {
+    for await (const chunk of nvidiaStarcoderComplete(prompt)) {
+        yield chunk;
+    }
+}
+
+export async function * nvidiaStarcoderComplete(prompt: string) : AsyncGenerator<CodeCompletionChunk> {
+    const invokeUrl = 'https://api.nvcf.nvidia.com/v2/nvcf/pexec/functions/6acada03-fe2f-4e4d-9e0a-e711b9fd1b59';
+
+    const headers = {
+        "Authorization": "Bearer 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",
+        "Accept": "text/event-stream",
+        "Content-Type": "application/json",
+    };
+
+    const payload = {
+        "prompt": prompt,
+        "temperature": 0.2,
+        "top_p": 0.7,
+        "max_tokens": 1024,
+        "seed": 42,
+        "bad": null,
+        "stop": ["<file_sep>"],
+        "stream": true
+    };
+
+    try {
+        const response = await fetch(invokeUrl, {
+            method: 'POST',
+    		headers,
+    		body: JSON.stringify(payload),
+        });
+
+        if (response.ok && response.body) {
+            const stream = response.body as any;
+            const decoder = new TextDecoder("utf-8");
+
+            for await (const chunk of stream) {
+                const chunkText = decoder.decode(chunk).trim();
+                // data: {"id":"5d3376e0-2abc-4230-b796-c6fc9ae91cd4","choices":[{"index":0,"delta":"-","finish_reason":null}]}\n\n
+                if (!chunkText.startsWith("data:")) {
+                    // log unexpected data
+                    logger.channel()?.info("Unexpected data: " + chunkText);
+                    return;
+                }
+
+                const jsonData = chunkText.substring(5).trim();
+                if (jsonData === "[DONE]") {
+                    return;
+                }
+
+                
+                try {
+                    const data = JSON.parse(chunkText.substring(5).trim());
+                    yield {
+                        text: data.choices[0].delta,
+                        id: data.id
+                    };
+                } catch (e: any) {
+                    logger.channel()?.info("receve:", chunkText);
+                    logger.channel()?.error("JSON Parsing Error:", e.message);
+                }
+            }
+        } else {
+            logger.channel()?.error("Error making request:", response.statusText);
+        }
+    } catch (error: any) {
+        logger.channel()?.error("Error making request:", error.message);
+    }
+}

src/contributes/codecomplete/modelConfig.ts

@@ -0,0 +1,23 @@
+
+
+export interface ModelConfigTemplate {
+    template: string;
+    stop: string[];
+}
+
+const stableCodeTemplate: ModelConfigTemplate = {
+    template: "<fim_prefix>{{{prefix}}}<fim_suffix>{{{suffix}}}<fim_middle>",
+    stop: ["<fim_prefix>", "<fim_suffix>", "<fim_middle>", "<|endoftext|>"],
+};
+
+const MODLE_COMPLETE_CONFIG = {
+    'starcoder': stableCodeTemplate,
+    'starcoder2': stableCodeTemplate,
+};
+
+export function getModelConfigTemplate(modelName: string): ModelConfigTemplate | undefined {
+    if (modelName in MODLE_COMPLETE_CONFIG) {
+        return MODLE_COMPLETE_CONFIG[modelName];
+    }
+    return undefined;
+}

src/contributes/codecomplete/promptCreator.ts

@@ -0,0 +1,69 @@
+/*
+ 针对代码补全功能，构建prompt
+
+ prompt的好坏，取决于提供的上下文信息。
+ 通过AST获取相对完整的信息，可能会增加提示的准确度，但也会增加代码提示的复杂度。
+ */
+
+import { logger } from "../../util/logger";
+import { log } from "console";
+
+
+const PREFIX_MAX_SIZE: number = 600;
+const SUFFIX_MAX_SIZE: number = 400;
+
+// 尽量获取一个完整的代码片段作为代码补全的上下文
+// 解析AST是一个好的方法，但还是会有点偏重计算。先尝试通过缩进来定位合适的块。
+// 整体范围保持在30行代码以内。
+async function curfilePrompt(filePath: string, fileContent: string, line: number, column: number) {
+    // 以line, column为中心，向前后扩展, 按行找出符合PREFIX_MAX_SIZE， SUFFIX_MAX_SIZE长度显示的prefix, suffix
+    // 分割文件内容为行数组
+    const lines = fileContent.split('\n');
+
+    // 初始化prefix和suffix内容及长度
+    let prefix = '';
+    let suffix = '';
+    let prefixSize = 0;
+    let suffixSize = 0;
+
+    // 从光标所在行开始，向上构建前缀
+    for (let i = line; i >= 0; i--) {
+        let lineText: string = lines[i] + '\n';
+        if (i === line) {
+            lineText = lines[i].substring(0, column);
+        }
+
+        prefix = lineText + prefix;
+        prefixSize += lineText.length;
+        if (prefixSize > PREFIX_MAX_SIZE) {
+            break;
+        }
+    }
+
+    // 从光标所在行下一行开始，向下构建后缀
+    for (let i = line; i < lines.length; i++) {
+        let lineText = lines[i] + '\n';
+        if (i === line) {
+            lineText = lines[i].substring(column, lines[i].length) + '\n';
+        }
+
+        suffix += lineText;
+        suffixSize += lineText.length;
+        if (suffixSize > PREFIX_MAX_SIZE) {
+            break;
+        }
+    }
+
+    // 返回前缀和后缀
+    return {
+        prefix,
+        suffix
+    };
+}
+
+export async function createPrompt(filePath: string, fileContent: string, line: number, column: number) {
+    const { prefix, suffix } = await curfilePrompt(filePath, fileContent, line, column);
+    const prompt = "<fim_prefix>" + prefix + "<fim_suffix>" + suffix + "<fim_middle>";
+
+    return prompt;
+}

src/contributes/codecomplete/recentEdits.ts

@@ -0,0 +1,4 @@
+/*
+ 记录最近修改的内容，用于代码补全
+*/
+

src/contributes/codecomplete/symbols.ts

@@ -0,0 +1,5 @@
+/*
+ 获取符号名称、符号类型（函数、变量、类等）、符号所在文件路径、符号所在行数等信息
+ 猜想这些信息会有助于代码补全功能的准确率提升
+*/
+

src/extension.ts

@@ -131,6 +131,12 @@ async function activate(context: vscode.ExtensionContext) {
     regLanguageContext();
     regDevChatView(context);
 	
+  const provider = new InlineCompletionProvider();  
+  const selector = { pattern: "**" }; 
+  context.subscriptions.push(vscode.languages.registerInlineCompletionItemProvider(selector, provider));  
+  registerCodeCompleteCallbackCommand(context);
+
+
   registerOpenChatPanelCommand(context);
   registerAddContextCommand(context);
   registerAskForCodeCommand(context);