Add dual ACE-Step music studio interfaces

2026-09-08 18:46:54 +02:00
parent f58d61140e
commit a4e894fe70
12 changed files with 404 additions and 23 deletions
@@ -1,9 +1,9 @@
 # ACE-Step 1.5 XL SFT: isolierter Athena-Test

-Dieser Compose-Dienst ist absichtlich noch nicht Teil des produktiven Stacks.
-Er startet nicht automatisch, bindet die Testoberflaeche nur an localhost und
-kann mitsamt seinem Container entfernt werden, ohne Router oder Hermes zu
-veraendern. Gewichte und Ergebnisse liegen getrennt unter `/data`.
+Der GPU-Worker startet nur im Musikmodus und bindet seine rohe
+Gradio-Oberflaeche nur an localhost. Die separate `fspecii/ace-step-ui`
+bleibt als leichte React/Express-Oberflaeche aktiv; ihre SQLite-Datenbank,
+Bibliothek und Uploads liegen persistent unter `/data/music/ace-step-ui`.

 Das offizielle Image ist auf den am 8. September 2026 geladenen Digest
 `sha256:95652cd780c78a1b1a7f6f0335530430f0ae53d96c7c12d59f9f39fa23d38567`
@@ -30,6 +30,23 @@ bleiben bewusst vorgenommene Aenderungen wieder im jeweiligen Browser erhalten.
 Beim Wechsel des gepinnten Image-Digests muessen die Overrides gegen die neue
 Upstream-Fassung geprueft werden.

+## Zwei Musikoberflaechen
+
+Die originale Gradio-Oberflaeche aus demselben ACE-Step-Image ist der stabile
+Produktionspfad fuer Simple, Custom, Cover, Remix und Repaint. Sie ist im
+WireGuard-Netz unter `http://192.168.1.212:7862` erreichbar.
+
+`music-ui` baut [fspecii/ace-step-ui](https://github.com/fspecii/ace-step-ui)
+reproduzierbar von Commit `a1fdf91829ec6f7b98844f80e323529cd155dbf2`.
+Ein lokaler Backport des noch offenen Upstream-PR
+[#109](https://github.com/fspecii/ace-step-ui/pull/109) gleicht das Frontend an
+das aktuelle 72-Felder-Gradio-Schema an. Athena-spezifisch sind nur die
+persistente Ablage, die XL-SFT-Anzeige und die gemeldeten Laufzeitlimits.
+Schlaegt bei einer spaeteren Upstream-Fassung ein Patch-Anker fehl, bricht der
+Image-Build ab. Die Community-UI unter `http://192.168.1.212:7861` ist bis zu
+vollstaendigen Ende-zu-Ende-Tests von Cover und Remix als experimentell
+gekennzeichnet.
+
 ## Start

 Vor dem Start muessen das aktive llama.cpp-Profil und Qwen3-TTS beendet sein.
@@ -37,24 +54,28 @@ Die RTX 5080 wird ueber ihre UUID exklusiv an den Container uebergeben.

 ```bash
 export ACESTEP_GPU_UUID="GPU-..."
+export ACESTEP_UI_JWT_SECRET="$(openssl rand -hex 32)"
+docker compose up -d music-ui
 docker compose --profile music-test up -d music-worker
 docker compose logs -f music-worker
 ```

-Die lokale Oberflaeche kann vom Mac aus ueber einen SSH-Tunnel erreicht werden:
+Alternativ zum WireGuard-Zugriff lassen sich beide Oberflaechen per SSH-Tunnel
+erreichen:

 ```bash
-ssh -L 7861:127.0.0.1:7861 root@athena.scc.kit.edu
+ssh -L 7861:127.0.0.1:7861 -L 7862:127.0.0.1:7862 root@athena.scc.kit.edu
 ```

-Danach im Browser `http://127.0.0.1:7861` aufrufen.
+`ace-step-ui` spricht den Worker ausschliesslich ueber das interne Docker-Netz
+an. Dafuer startet ACE-Step mit aktivierten, benannten API-Endpunkten
+(`--enable-api`).

 ## Beenden

 ```bash
-docker compose --profile music-test down
+docker compose --profile music-test stop music-worker
 ```

-Der Befehl entfernt nur den Testcontainer und das Testnetz. Caches,
-Modellgewichte und Ausgaben bleiben fuer reproduzierbare Messungen erhalten.
-Sie werden erst nach dokumentierter Ablehnung gezielt geloescht.
+Der Befehl stoppt nur den GPU-Worker. Die Musikoberflaeche, ihre Bibliothek,
+Caches, Modellgewichte und Ausgaben bleiben erhalten.
@@ -0,0 +1,38 @@
+FROM node:22-bookworm AS build
+
+ARG ACE_STEP_UI_COMMIT
+RUN test -n "$ACE_STEP_UI_COMMIT"
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends git python3 make g++ \
+    && rm -rf /var/lib/apt/lists/*
+RUN git clone https://github.com/fspecii/ace-step-ui.git /src \
+    && cd /src \
+    && git checkout --detach "$ACE_STEP_UI_COMMIT"
+COPY patch-source.mjs /tmp/patch-source.mjs
+RUN node /tmp/patch-source.mjs /src
+RUN cd /src \
+    && npm ci \
+    && npm run build
+RUN cd /src/server \
+    && npm ci \
+    && npm run build \
+    && npm prune --omit=dev
+
+FROM node:22-bookworm-slim AS runtime
+
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends nginx curl ca-certificates ffmpeg \
+    && rm -rf /var/lib/apt/lists/*
+COPY --from=build /src/dist /usr/share/nginx/html
+COPY --from=build /src/server/dist /app/server/dist
+COPY --from=build /src/server/node_modules /app/server/node_modules
+COPY --from=build /src/server/package.json /app/server/package.json
+COPY --from=build /src/server/public /app/server/public
+COPY --from=build /src/server/audio-editor /app/server/audio-editor
+COPY nginx.conf /etc/nginx/nginx.conf
+COPY entrypoint.sh /usr/local/bin/ace-step-ui-entrypoint
+RUN chmod 0755 /usr/local/bin/ace-step-ui-entrypoint \
+    && mkdir -p /data/audio /data/datasets/uploads
+
+EXPOSE 3000 3001
+ENTRYPOINT ["/usr/local/bin/ace-step-ui-entrypoint"]
@@ -0,0 +1,18 @@
+#!/bin/sh
+set -eu
+
+node /app/server/dist/index.js &
+server_pid=$!
+
+trap 'kill "$server_pid" 2>/dev/null || true' INT TERM EXIT
+nginx -g 'daemon off;' &
+nginx_pid=$!
+
+while kill -0 "$server_pid" 2>/dev/null && kill -0 "$nginx_pid" 2>/dev/null; do
+  sleep 1
+done
+
+kill "$server_pid" "$nginx_pid" 2>/dev/null || true
+wait "$server_pid" 2>/dev/null || true
+wait "$nginx_pid" 2>/dev/null || true
+exit 1
@@ -0,0 +1,35 @@
+worker_processes auto;
+pid /tmp/nginx.pid;
+
+events {
+  worker_connections 1024;
+}
+
+http {
+  include /etc/nginx/mime.types;
+  default_type application/octet-stream;
+  sendfile on;
+  client_max_body_size 512m;
+
+  server {
+    listen 3000;
+    server_name _;
+    root /usr/share/nginx/html;
+    index index.html;
+
+    location ~ ^/(api|audio|editor|blog|demucs-web)(/|$) {
+      proxy_pass http://127.0.0.1:3001;
+      proxy_http_version 1.1;
+      proxy_set_header Host $host;
+      proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+      proxy_set_header X-Forwarded-Proto $scheme;
+      proxy_buffering off;
+      proxy_read_timeout 1800s;
+      proxy_send_timeout 1800s;
+    }
+
+    location / {
+      try_files $uri $uri/ /index.html;
+    }
+  }
+}
@@ -0,0 +1,187 @@
+import fs from 'node:fs';
+import path from 'node:path';
+
+const root = process.argv[2];
+if (!root) throw new Error('source root argument is required');
+
+function patch(relativePath, transform) {
+  const filename = path.join(root, relativePath);
+  const before = fs.readFileSync(filename, 'utf8');
+  const after = transform(before);
+  if (after === before) throw new Error(`patch made no change: ${relativePath}`);
+  fs.writeFileSync(filename, after);
+}
+
+function replaceOnce(text, before, after, label) {
+  const first = text.indexOf(before);
+  if (first < 0) throw new Error(`patch anchor missing: ${label}`);
+  if (text.indexOf(before, first + 1) >= 0) throw new Error(`patch anchor repeated: ${label}`);
+  return text.slice(0, first) + after + text.slice(first + before.length);
+}
+
+patch('server/src/services/acestep.ts', (text) => {
+  text = replaceOnce(
+    text,
+    "const AUDIO_DIR = path.join(__dirname, '../../public/audio');",
+    'const AUDIO_DIR = config.storage.audioDir;',
+    'persistent generated audio',
+  );
+
+  const oldArgs = `    params.instruction || 'Fill the audio semantic mask with the style described in the text prompt.', // 17: Instruction
+    params.audioCoverStrength ?? 1.0,                             // 18: Audio Cover Strength
+    0.0,                                                          // 19: Cover Noise Strength (ACE-Step v1.5 new param, default 0.0)
+    (params.taskType === 'audio2audio' ? 'cover' : params.taskType) || 'text2music', // 20: Task Type
+    params.useAdg ?? false,                                       // 21: Use ADG
+    params.cfgIntervalStart ?? 0.0,                               // 22: CFG Interval Start
+    params.cfgIntervalEnd ?? 1.0,                                 // 23: CFG Interval End
+    params.shift ?? 3.0,                                          // 24: Shift
+    params.inferMethod || 'ode',                                  // 25: Inference Method
+    params.customTimesteps || '',                                 // 26: Custom Timesteps
+    params.audioFormat || 'mp3',                                  // 27: Audio Format
+    params.lmTemperature ?? 0.85,                                 // 28: LM Temperature
+    isThinking,                                                   // 29: Think
+    params.lmCfgScale ?? 2.0,                                    // 30: LM CFG Scale
+    params.lmTopK ?? 0,                                           // 31: LM Top-K
+    params.lmTopP ?? 0.9,                                         // 32: LM Top-P
+    params.lmNegativePrompt || 'NO USER INPUT',                   // 33: LM Negative Prompt
+    useCot ? (params.useCotMetas ?? true) : false,                // 34: CoT Metas
+    useCot ? (params.useCotCaption ?? true) : false,              // 35: CaptionRewrite
+    useCot ? (params.useCotLanguage ?? true) : false,             // 36: CoT Language
+    params.isFormatCaption ?? false,                              // 37: Is Format Caption State
+    params.constrainedDecodingDebug ?? false,                     // 38: Constrained Decoding Debug
+    params.allowLmBatch ?? true,                                  // 39: ParallelThinking
+    params.getScores ?? false,                                    // 40: Auto Score
+    params.getLrc ?? false,                                       // 41: Auto LRC (timestamped lyrics)
+    params.scoreScale ?? 0.5,                                     // 42: Quality Score Sensitivity (0.01-1.0)
+    params.lmBatchChunkSize ?? 8,                                 // 43: LM Batch Chunk Size
+    params.trackName || null,                                     // 44: Track Name
+    params.completeTrackClasses || [],                            // 45: Track Names
+    true,                                                         // 46: Enable Normalization (ACE-Step v1.5, default true)
+    -1.0,                                                         // 47: Normalization DB (ACE-Step v1.5, default -1.0)
+    0.0,                                                          // 48: Latent Shift (ACE-Step v1.5, default 0.0)
+    1.0,                                                          // 49: Latent Rescale (ACE-Step v1.5, default 1.0)
+    params.autogen ?? false,                                      // 50: AutoGen`;
+
+  const newArgs = `    params.instruction || 'Fill the audio semantic mask based on the given conditions:', // 17: Instruction
+    params.audioCoverStrength ?? 1.0,                             // 18: LM Codes Strength
+    0.0,                                                          // 19: Cover Strength
+    false,                                                        // 20: no_fsq
+    params.useAdg ?? false,                                       // 21: Use ADG
+    params.cfgIntervalStart ?? 0.0,                               // 22: CFG Interval Start
+    params.cfgIntervalEnd ?? 1.0,                                 // 23: CFG Interval End
+    params.shift ?? 1.0,                                          // 24: Shift
+    params.inferMethod || 'ode',                                  // 25: Inference Method
+    'euler',                                                      // 26: Sampler Mode
+    0.0,                                                          // 27: Velocity Norm Threshold
+    0.0,                                                          // 28: Velocity EMA Factor
+    false,                                                        // 29: Enable DCW
+    'double',                                                     // 30: DCW Mode
+    0.02,                                                         // 31: DCW Scaler
+    0.06,                                                         // 32: DCW High Scaler
+    'haar',                                                       // 33: DCW Wavelet
+    params.customTimesteps || '',                                 // 34: Custom Timesteps
+    params.audioFormat || 'flac',                                 // 35: Audio Format
+    '320k',                                                       // 36: MP3 Bitrate
+    48000,                                                        // 37: MP3 Sample Rate
+    params.lmTemperature ?? 0.85,                                 // 38: LM Temperature
+    isThinking,                                                   // 39: Think
+    params.lmCfgScale ?? 2.0,                                    // 40: LM CFG Scale
+    params.lmTopK ?? 0,                                           // 41: LM Top-K
+    params.lmTopP ?? 0.9,                                         // 42: LM Top-P
+    params.lmNegativePrompt || 'NO USER INPUT',                   // 43: LM Negative Prompt
+    useCot ? (params.useCotMetas ?? true) : false,                // 44: CoT Metas
+    useCot ? (params.useCotCaption ?? true) : false,              // 45: CaptionRewrite
+    useCot ? (params.useCotLanguage ?? true) : false,             // 46: CoT Language
+    params.constrainedDecodingDebug ?? false,                     // 47: Constrained Decoding Debug
+    params.allowLmBatch ?? true,                                  // 48: ParallelThinking
+    params.getScores ?? false,                                    // 49: Auto Score
+    params.getLrc ?? false,                                       // 50: Auto LRC
+    params.scoreScale ?? 0.5,                                     // 51: Quality Score Sensitivity
+    params.lmBatchChunkSize ?? 8,                                 // 52: LM Batch Chunk Size
+    params.trackName || null,                                     // 53: Track Name
+    params.completeTrackClasses || [],                            // 54: Track Names
+    true,                                                         // 55: Enable Normalization
+    -1.0,                                                         // 56: Target Peak (dB)
+    0.0,                                                          // 57: Fade In
+    0.0,                                                          // 58: Fade Out
+    0.0,                                                          // 59: Latent Shift
+    1.0,                                                          // 60: Latent Rescale
+    'balanced',                                                   // 61: Repaint Mode
+    0.5,                                                          // 62: Repaint Strength
+    0.0,                                                          // 63: variance
+    '',                                                           // 64: repaint seed
+    false,                                                        // 65: Edit
+    null,                                                         // 66: source caption
+    null,                                                         // 67: source lyrics
+    0.0,                                                          // 68: n_min
+    1.0,                                                          // 69: n_max
+    1,                                                            // 70: n_avg
+    params.autogen ?? false,                                      // 71: AutoGen`;
+
+  return replaceOnce(text, oldArgs, newArgs, 'Athena Gradio argument schema');
+});
+
+patch('server/src/services/storage/local.ts', (text) => {
+  text = replaceOnce(
+    text,
+    "import type { StorageProvider } from './index.js';",
+    "import type { StorageProvider } from './index.js';\nimport { config } from '../../config/index.js';",
+    'storage config import',
+  );
+  return replaceOnce(
+    text,
+    "const AUDIO_DIR = path.join(__dirname, '../../../public/audio');",
+    'const AUDIO_DIR = config.storage.audioDir;',
+    'persistent uploaded audio',
+  );
+});
+
+patch('server/src/index.ts', (text) => replaceOnce(
+  text,
+  "app.use('/audio', express.static(path.join(__dirname, '../public/audio')));",
+  "app.use('/audio', express.static(config.storage.audioDir));",
+  'persistent audio static route',
+));
+
+patch('server/src/routes/referenceTrack.ts', (text) => {
+  text = replaceOnce(
+    text,
+    "import { spawn } from 'child_process';",
+    "import { spawn } from 'child_process';\nimport { config } from '../config/index.js';",
+    'reference audio config import',
+  );
+  return replaceOnce(
+    text,
+    "const AUDIO_DIR = path.join(__dirname, '../../public/audio');",
+    'const AUDIO_DIR = config.storage.audioDir;',
+    'persistent reference audio',
+  );
+});
+
+patch('server/src/routes/generate.ts', (text) => {
+  text = replaceOnce(
+    text,
+    "    const ALL_DIT_MODELS = [\n      'acestep-v15-turbo',",
+    "    const ALL_DIT_MODELS = [\n      'acestep-v15-xl-sft',            // Athena production model\n      'acestep-v15-turbo',",
+    'XL-SFT model list',
+  );
+
+  const start = text.indexOf("router.get('/limits'");
+  const end = text.indexOf("router.get('/debug/", start);
+  if (start < 0 || end < 0) throw new Error('limits route anchors missing');
+  const limits = `router.get('/limits', async (_req, res: Response) => {
+  // The UI container intentionally has no CUDA or ACE-Step Python runtime.
+  // These are the limits reported by Athena's dedicated RTX 5080 worker.
+  res.json({
+    tier: process.env.ACESTEP_TIER || 'tier5',
+    gpu_memory_gb: Number(process.env.ACESTEP_GPU_MEMORY_GB || 15.5),
+    max_duration_with_lm: Number(process.env.ACESTEP_MAX_DURATION_WITH_LM || 480),
+    max_duration_without_lm: Number(process.env.ACESTEP_MAX_DURATION_WITHOUT_LM || 600),
+    max_batch_size_with_lm: Number(process.env.ACESTEP_MAX_BATCH_WITH_LM || 4),
+    max_batch_size_without_lm: Number(process.env.ACESTEP_MAX_BATCH_WITHOUT_LM || 4),
+  });
+});
+
+`;
+  return text.slice(0, start) + limits + text.slice(end);
+});
@@ -15,7 +15,8 @@ services:
      ACESTEP_DEVICE: cuda
      # The image entrypoint forwards only ACESTEP_EXTRA_ARGS to the UI CLI.
      # One result per run avoids the batch=2 VRAM/time penalty.
-      ACESTEP_EXTRA_ARGS: "--batch_size 1"
+      # Named Gradio endpoints are consumed by the separate ace-step-ui service.
+      ACESTEP_EXTRA_ARGS: "--batch_size 1 --enable-api"
      TOKENIZERS_PARALLELISM: "false"
      NVIDIA_VISIBLE_DEVICES: ${ACESTEP_GPU_UUID:?set ACESTEP_GPU_UUID to the RTX 5080 UUID}
    deploy:
@@ -26,7 +27,8 @@ services:
              device_ids: ["${ACESTEP_GPU_UUID:?set ACESTEP_GPU_UUID to the RTX 5080 UUID}"]
              capabilities: [gpu]
    ports:
-      - "127.0.0.1:${ACESTEP_UI_PORT:-7861}:7860"
+      # Raw Gradio remains available for diagnostics; users open music-ui below.
+      - "127.0.0.1:${ACESTEP_GRADIO_PORT:-7862}:7860"
    volumes:
      - ${ACESTEP_CHECKPOINTS_DIR:-/data/models/acestep/checkpoints}:/app/checkpoints
      - ${ACESTEP_HF_CACHE_DIR:-/data/models/acestep/hf-cache}:/root/.cache/huggingface
@@ -45,3 +47,45 @@ services:
      timeout: 10s
      start_period: 300s
      retries: 3
+    networks:
+      - music
+      - frontend
+
+  music-ui:
+    build:
+      context: ./ace-step-ui
+      args:
+        ACE_STEP_UI_COMMIT: a1fdf91829ec6f7b98844f80e323529cd155dbf2
+    image: mike-ai/ace-step-ui:a1fdf918
+    container_name: mike-ai-music-ui
+    environment:
+      NODE_ENV: production
+      PORT: "3001"
+      FRONTEND_URL: ${ACESTEP_UI_PUBLIC_URL:-http://192.168.1.212:7861}
+      ACESTEP_API_URL: http://music-worker:7860
+      DATABASE_PATH: /data/acestep.db
+      AUDIO_DIR: /data/audio
+      DATASETS_DIR: /data/datasets
+      DATASETS_UPLOADS_DIR: /data/datasets/uploads
+      JWT_SECRET: ${ACESTEP_UI_JWT_SECRET:?set ACESTEP_UI_JWT_SECRET}
+    ports:
+      - "127.0.0.1:${ACESTEP_UI_PORT:-7861}:3000"
+    volumes:
+      - ${ACESTEP_UI_DATA_DIR:-/data/music/ace-step-ui}:/data
+    restart: unless-stopped
+    healthcheck:
+      test: ["CMD-SHELL", "curl -fsS http://127.0.0.1:3000/ >/dev/null && curl -fsS http://127.0.0.1:3001/health >/dev/null"]
+      interval: 30s
+      timeout: 10s
+      start_period: 30s
+      retries: 3
+    networks:
+      - music
+      - frontend
+
+networks:
+  music:
+    name: mike-ai-music
+  frontend:
+    external: true
+    name: mike-ai_frontend