feat: 调度器+并发管线+GPU优化+日志分级+前端修复

- scheduler: ffmpeg 异步线程 + GPU 串行调度 + 模型复用(2N→2 次加载)
- pipeline: 阶段拆分(extract/asr/translate),中间数据存 Task 字段
- translate_service: 长度排序批处理,padding 浪费减少 91%
- model_manager: ASR/翻译不共驻,BatchedInferencePipeline 批量解码
- 日志分级: INFO=任务流转里程碑,DEBUG=进度详情;默认 INFO
- 前端: 日志最新在上+滚动感知+退避轮询;24h 时间;上传中状态显示
- /health: 返回完整 Whisper/NLLB 配置
- upload_service: 单事务 complete + 扩展名白名单
- task_router: 合并 UploadSession 虚拟任务到列表
- Dockerfile: CPU/GPU 独立构建链,deps 缓存稳定
- prefetch_models: 安装时预下载模型权重
This commit is contained in:
audio2text dev
2026-07-06 21:59:59 +08:00
parent 00e2a95fb7
commit 73110848f4
30 changed files with 1238 additions and 259 deletions

View File

@@ -202,10 +202,11 @@ a { color: var(--accent); text-decoration: none; }
SHARED_JS = """
// 任务状态中文标签
const STATUS_LABEL = {
queued: "排队中", extracting: "提取音频", transcribing: "语音识别",
segmenting: "断句重算", translating: "翻译中", done: "完成", failed: "失败"
uploading: "上传中", queued: "排队中", extracting: "提取音频",
transcribing: "语音识别", segmenting: "断句重算", translating: "翻译中",
done: "完成", failed: "失败"
};
const ACTIVE_STATES = ["queued","extracting","transcribing","segmenting","translating"];
const ACTIVE_STATES = ["uploading","queued","extracting","transcribing","segmenting","translating"];
// HTML 转义(防 XSS
function escapeHtml(s) {
@@ -222,6 +223,16 @@ function fmtBytes(n) {
return u === 0 ? x + " B" : x.toFixed(1) + " " + units[u];
}
// 24 小时制时间格式化(不受浏览器 locale 影响,避免 am/pm 混淆)
function _pad2(n) { return n < 10 ? "0" + n : "" + n; }
function fmtTime24(d) {
return _pad2(d.getHours()) + ":" + _pad2(d.getMinutes()) + ":" + _pad2(d.getSeconds());
}
function fmtDateTime24(d) {
return d.getFullYear() + "-" + _pad2(d.getMonth()+1) + "-" + _pad2(d.getDate())
+ " " + fmtTime24(d);
}
// 并发池indices 中的每个元素交给 worker最多 concurrency 个并发
async function runPool(indices, concurrency, worker) {
let cursor = 0;

View File

@@ -68,7 +68,11 @@ function renderTable(tasks) {{
const tr = document.createElement("tr");
const label = STATUS_LABEL[task.status] || task.status;
const stateClass = "st-" + (task.status === "done" ? "done" : task.status === "failed" ? "fail" : "running");
const created = new Date(task.created_at).toLocaleString();
let created;
try {{ created = fmtDateTime24(new Date(task.created_at + "Z")); }}
catch (e) {{ created = task.created_at; }}
let idDisplay = task.is_upload ? "" : "#" + task.id;
let action;
if (task.status === "done") {{
@@ -84,10 +88,13 @@ function renderTable(tasks) {{
let progress;
if (task.status === "done") progress = "100%";
else if (task.status === "failed") progress = "";
else progress = `<div class="mini-bar"><div class="mini-fill" style="width:${{task.progress}}%"></div></div>${{task.progress.toFixed(0)}}%`;
else {{
const pct = (task.progress == null) ? 0 : task.progress;
progress = `<div class="mini-bar"><div class="mini-fill" style="width:${{pct}}%"></div></div>${{pct.toFixed(0)}}%`;
}}
tr.innerHTML = `
<td class="muted">#${{task.id}}</td>
<td class="muted">${{idDisplay}}</td>
<td>${{escapeHtml(task.filename)}}</td>
<td><span class="status-tag ${{stateClass}}">${{label}}</span></td>
<td>${{progress}}</td>

View File

@@ -42,12 +42,13 @@ function renderTasks(tasks) {{
for (const task of tasks) {{
tasksEl.appendChild(makeTaskCard(task));
}}
// 对进行中的任务启动轮询
// 对进行中的任务启动轮询(上传会话用 upload_id 去重Task 用 id 去重)
for (const task of tasks) {{
if (ACTIVE_STATES.includes(task.status) && !pollingIds.has(task.id)) {{
pollingIds.add(task.id);
pollTask(task.id);
}}
if (!ACTIVE_STATES.includes(task.status)) continue;
const pollKey = task.is_upload ? task.upload_id : String(task.id);
if (pollingIds.has(pollKey)) continue;
pollingIds.add(pollKey);
pollTask(task);
}}
}}
@@ -61,6 +62,7 @@ function makeTaskCard(task) {{
const el = document.createElement("div");
el.className = "card task server-task";
el.dataset.taskId = task.id;
if (task.is_upload) el.dataset.uploadId = task.upload_id;
el.innerHTML = renderTaskInner(task);
return el;
}}
@@ -69,7 +71,11 @@ function renderTaskInner(task) {{
const label = STATUS_LABEL[task.status] || task.status;
const stateClass = task.status === "done" ? "state-done"
: task.status === "failed" ? "state-fail" : "state-running";
const created = new Date(task.created_at).toLocaleString();
// 上传中created_at 可能是 naive UTC前端按 UTC 解析
let created;
try {{ created = fmtDateTime24(new Date(task.created_at + "Z")); }}
catch (e) {{ created = task.created_at; }}
let body;
if (task.status === "done") {{
@@ -77,35 +83,72 @@ function renderTaskInner(task) {{
}} else if (task.status === "failed") {{
body = `<div class="task-meta fail-msg">${{escapeHtml(task.error || "未知错误")}}</div>`;
}} else {{
body = `<div class="bar"><div class="fill" style="width:${{task.progress}}%"></div><span class="pct">${{task.progress.toFixed(0)}}%</span></div>`;
const pct = (task.progress == null) ? 0 : task.progress;
const sizeInfo = task.size_bytes ? ` · ${{fmtBytes(task.size_bytes)}}` : "";
body = `<div class="bar"><div class="fill" style="width:${{pct}}%"></div><span class="pct">${{pct.toFixed(0)}}%${{sizeInfo}}</span></div>`;
}}
return `
<div class="task-head">
<span class="fname">#${{task.id}} ${{escapeHtml(task.filename)}}</span>
<span class="fname">${{task.is_upload ? "" : "#" + task.id + " "}}${{escapeHtml(task.filename)}}</span>
<span class="fstate ${{stateClass}}">${{label}}</span>
</div>
${{body}}
<div class="task-time">${{created}}</div>`;
}}
async function pollTask(taskId) {{
async function pollTask(task) {{
// 上传会话:轮询 upload status 接口Task轮询 task 接口
const isUpload = task.is_upload === true && task.upload_id;
const url = isUpload
? "/api/tasks/chunk-uploads/" + task.upload_id + "/status"
: "/api/tasks/" + task.id;
let backoff = POLL_INTERVAL;
const tick = async () => {{
try {{
const r = await fetch("/api/tasks/" + taskId);
if (!r.ok) {{ pollingIds.delete(taskId); return; }}
const task = await r.json();
const card = tasksEl.querySelector('.server-task[data-task-id="' + taskId + '"]');
if (!card) {{ pollingIds.delete(taskId); return; }}
const r = await fetch(url);
if (!r.ok) {{ pollingIds.delete(task.id); pollingIds.delete(task.upload_id); return; }}
const data = await r.json();
card.innerHTML = renderTaskInner(task);
if (task.status === "done" || task.status === "failed") {{
pollingIds.delete(taskId);
return;
if (isUpload) {{
// 上传会话complete 后 task_id 出现,切换为 Task 轮询
if (data.completed && data.task_id) {{
pollingIds.delete(task.upload_id);
pollingIds.add(data.task_id);
pollTask({{ id: data.task_id, is_upload: false }});
// 刷新列表让新 Task 卡片出现
refreshList();
return;
}}
// 更新上传进度
const card = tasksEl.querySelector('.server-task[data-upload-id="' + task.upload_id + '"]');
if (!card) {{ pollingIds.delete(task.upload_id); return; }}
const uploaded = (data.uploaded_chunks || []).length;
const total = data.total_chunks || 1;
const pct = Math.min(4, uploaded / total * 4);
const fakeTask = {{
status: "uploading", progress: pct, is_upload: true,
upload_id: task.upload_id, filename: data.filename,
size_bytes: data.size_bytes, created_at: task.created_at,
}};
card.innerHTML = renderTaskInner(fakeTask);
backoff = POLL_INTERVAL;
setTimeout(tick, backoff);
}} else {{
// Task 轮询
const card = tasksEl.querySelector('.server-task[data-task-id="' + task.id + '"]');
if (!card) {{ pollingIds.delete(task.id); return; }}
card.innerHTML = renderTaskInner(data);
if (data.status === "done" || data.status === "failed") {{
pollingIds.delete(task.id);
return;
}}
backoff = POLL_INTERVAL;
setTimeout(tick, backoff);
}}
setTimeout(tick, POLL_INTERVAL);
}} catch (e) {{
setTimeout(tick, POLL_INTERVAL);
// 网络错误:指数退避,上限 30s
backoff = Math.min(backoff * 1.6, 30000);
setTimeout(tick, backoff);
}}
}};
tick();

View File

@@ -2,25 +2,34 @@
共享 _shared.py 的 BASE_CSS / SHARED_JS。
页面专属级别过滤按钮、自动刷新开关、清空、traceback 折叠。
显示策略:
- 默认 INFO仅阶段转换/模型加载卸载/任务流转开始结束),可切 DEBUG 看进度详情。
- 最新日志在顶部;用户向上滚动浏览历史时不会被自动刷新拉走(仅当停在顶部时跟随)。
- 轮询带指数退避,连续失败时拉长间隔,避免服务不可达时打爆。
"""
from __future__ import annotations
from ._shared import render_page
POLL_INTERVAL_LOGS = 2000
POLL_INTERVAL_LOGS = 2000 # 正常轮询间隔ms
POLL_MAX_INTERVAL = 30000 # 退避上限ms
DEFAULT_TAIL = 200
_PAGE_JS = f"""
const POLL_INTERVAL = {POLL_INTERVAL_LOGS};
const POLL_MIN_INTERVAL = {POLL_INTERVAL_LOGS};
const POLL_MAX_INTERVAL = {POLL_MAX_INTERVAL};
const DEFAULT_TAIL = {DEFAULT_TAIL};
const API = "/api/logs";
let currentLevel = "debug";
let currentLevel = "info"; // 默认 INFO
let autoRefresh = true;
let timer = null;
let pollInterval = POLL_MIN_INTERVAL; // 动态退避
let userScrolled = false; // 用户是否主动向下浏览历史
const logsEl = document.getElementById("logs");
const containerEl = document.getElementById("logs");
const emptyEl = document.getElementById("empty");
const statusEl = document.getElementById("status");
@@ -29,26 +38,34 @@ const LEVEL_CLASS = {{
CRITICAL: "st-fail"
}};
// 监听滚动:用户向下(往历史方向)浏览时暂停自动跟随,回到顶部则恢复
containerEl.addEventListener("scroll", () => {{
// scrollTop 越小越靠近顶部(最新)。接近顶部 = 用户在看最新
userScrolled = containerEl.scrollTop > 4;
}});
document.querySelectorAll(".filter").forEach(btn => {{
btn.addEventListener("click", () => {{
document.querySelectorAll(".filter").forEach(b => b.classList.remove("active"));
btn.classList.add("active");
currentLevel = btn.dataset.level;
logsEl.innerHTML = "";
containerEl.innerHTML = "";
pollInterval = POLL_MIN_INTERVAL;
fetchLogs();
}});
}});
document.getElementById("autorefresh").addEventListener("change", e => {{
autoRefresh = e.target.checked;
if (autoRefresh) fetchLogs(); else if (timer) {{ clearTimeout(timer); timer = null; }}
if (autoRefresh) {{ pollInterval = POLL_MIN_INTERVAL; fetchLogs(); }}
else if (timer) {{ clearTimeout(timer); timer = null; }}
}});
document.getElementById("clear-btn").addEventListener("click", async () => {{
if (!confirm("确定清空所有日志缓冲?")) return;
try {{
await fetch(API, {{ method: "DELETE" }});
logsEl.innerHTML = "";
containerEl.innerHTML = "";
statusEl.textContent = "已清空";
}} catch (e) {{ statusEl.textContent = "清空失败"; }}
}});
@@ -56,57 +73,87 @@ document.getElementById("clear-btn").addEventListener("click", async () => {{
async function fetchLogs() {{
try {{
const r = await fetch(`${{API}}?level=${{currentLevel}}&tail=${{DEFAULT_TAIL}}`);
if (!r.ok) {{ statusEl.textContent = "HTTP " + r.status; scheduleNext(); return; }}
if (!r.ok) {{ statusEl.textContent = "HTTP " + r.status; scheduleNext(true); return; }}
const data = await r.json();
renderLogs(data.logs);
statusEl.textContent = `${{data.count}} 条 · 更新 ${{new Date().toLocaleTimeString()}}`;
scheduleNext();
statusEl.textContent = `${{data.count}} 条 · 更新 ${{fmtTime24(new Date())}}`;
pollInterval = POLL_MIN_INTERVAL; // 成功,重置间隔
scheduleNext(false);
}} catch (e) {{
statusEl.textContent = "获取失败";
scheduleNext();
scheduleNext(true); // 失败,退避
}}
}}
function renderLogs(logs) {{
if (!logs || logs.length === 0) {{
if (logsEl.children.length === 0) emptyEl.style.display = "block";
if (containerEl.children.length === 0) emptyEl.style.display = "block";
return;
}}
emptyEl.style.display = "none";
const existing = new Set(Array.from(logsEl.children).map(el => el.dataset.key));
// 服务端已按时间升序返回(旧→新)。倒序渲染:最新在最上面。
// 用 key=ts|level|logger|msg 去重(含 logger 名,避免不同模块同消息被误判重复)
const existing = new Set(Array.from(containerEl.children).map(el => el.dataset.key));
const wasAtTop = containerEl.scrollTop <= 4;
const frag = document.createDocumentFragment();
for (const log of logs) {{
const key = `${{log.ts}}|${{log.level}}|${{log.msg}}`;
// 服务端按时间升序返回(旧→新)。倒序遍历 + appendChild → frag 内最新在前:
// logs[N-1](最新)先 append → frag 第一个logs[0](最旧)最后 → frag 最后
for (let i = logs.length - 1; i >= 0; i--) {{
const log = logs[i];
const key = `${{log.ts}}|${{log.level}}|${{log.logger}}|${{log.msg}}`;
if (existing.has(key)) continue;
const row = document.createElement("div");
row.className = "log-row " + (LEVEL_CLASS[log.level] || "st-running");
row.dataset.key = key;
const hasTrace = !!log.traceback;
row.innerHTML = `
<span class="log-ts">${{escapeHtml(log.ts)}}</span>
<span class="status-tag ${{LEVEL_CLASS[log.level] || 'st-running'}}">${{escapeHtml(log.level)}}</span>
<span class="log-logger">${{escapeHtml(log.logger)}}</span>
<span class="log-msg">${{escapeHtml(log.msg)}}${{hasTrace ? ' <span class="trace-toggle">[traceback]</span>' : ''}}</span>`;
if (hasTrace) {{
const pre = document.createElement("pre");
pre.className = "trace";
pre.textContent = log.traceback;
pre.style.display = "none";
row.appendChild(pre);
row.querySelector(".trace-toggle").addEventListener("click", e => {{
e.stopPropagation();
pre.style.display = pre.style.display === "none" ? "block" : "none";
}});
}}
frag.appendChild(row);
frag.appendChild(buildRow(log, key));
}}
if (frag.children.length > 0) {{
// 整块插入到容器顶部:最新日志在最上面
containerEl.insertBefore(frag, containerEl.firstChild);
}}
// 裁剪:超出上限删最旧(底部)
while (containerEl.children.length > DEFAULT_TAIL) {{
containerEl.removeChild(containerEl.lastChild);
}}
// 仅当用户停在顶部(看最新)时保持滚定在顶,否则不打扰浏览历史
if (wasAtTop && !userScrolled) {{
containerEl.scrollTop = 0;
}}
logsEl.appendChild(frag);
while (logsEl.children.length > DEFAULT_TAIL) logsEl.removeChild(logsEl.firstChild);
logsEl.scrollTop = logsEl.scrollHeight;
}}
function scheduleNext() {{
if (autoRefresh) timer = setTimeout(fetchLogs, POLL_INTERVAL);
function buildRow(log, key) {{
const row = document.createElement("div");
row.className = "log-row " + (LEVEL_CLASS[log.level] || "st-running");
row.dataset.key = key;
const hasTrace = !!log.traceback;
row.innerHTML = `
<span class="log-ts">${{escapeHtml(log.ts)}}</span>
<span class="status-tag ${{LEVEL_CLASS[log.level] || 'st-running'}}">${{escapeHtml(log.level)}}</span>
<span class="log-logger">${{escapeHtml(log.logger)}}</span>
<span class="log-msg">${{escapeHtml(log.msg)}}${{hasTrace ? ' <span class="trace-toggle">[traceback]</span>' : ''}}</span>`;
if (hasTrace) {{
const pre = document.createElement("pre");
pre.className = "trace";
pre.textContent = log.traceback;
pre.style.display = "none";
row.appendChild(pre);
row.querySelector(".trace-toggle").addEventListener("click", e => {{
e.stopPropagation();
pre.style.display = pre.style.display === "none" ? "block" : "none";
}});
}}
return row;
}}
function scheduleNext(failed) {{
if (!autoRefresh) return;
if (failed) {{
// 指数退避:每次失败 ×1.6,上限 30s
pollInterval = Math.min(pollInterval * 1.6, POLL_MAX_INTERVAL);
}}
timer = setTimeout(fetchLogs, pollInterval);
}}
fetchLogs();
@@ -136,12 +183,12 @@ _PAGE_CSS = """
_BODY = """
<h1>日志</h1>
<p class="sub">实时查看服务日志。debug=详细子步骤info=仅阶段转换error=完整错误。自动刷新每 2 秒。</p>
<p class="sub">默认显示 INFO任务流转/模型加载卸载)。切 DEBUG 看进度详情,切 警告+/仅错误 过滤问题。自动刷新每 2 秒,失败自动退避。最新日志在顶部,向下浏览历史时不会被拉走。</p>
<div class="toolbar">
<div class="filters">
<button class="filter active" data-level="debug">全部 (DEBUG)</button>
<button class="filter" data-level="info">简略 (INFO)</button>
<button class="filter" data-level="debug">全部 (DEBUG)</button>
<button class="filter active" data-level="info">简略 (INFO)</button>
<button class="filter" data-level="warning">警告+</button>
<button class="filter" data-level="error">仅错误</button>
</div>