2b-studio / studio /daw /api.py
2Btainment's picture
Round 3: True-Peak, Stems-Jobs (Queue/ETA/Abbruch), Cloud-Autosave privat, Smart-Prompt mit Takes, Mixer + FX-Presets, Undo komplett
2cdb2ce verified
Raw History Blame Contribute Delete
19.9 kB
"""DAW-Backend als Gradio-API (gr.api): Generieren je Element, Übergänge, FX, Export, Projekte.
Der Browser ruft die Endpunkte über /gradio_api/call/<name> auf. Dateien liegen in DAW_DIR und
werden über /gradio_api/file= ausgeliefert; Eingabepfade werden auf erlaubte Ordner geprüft.
"""
from __future__ import annotations
import json
import os
import re
import tempfile
import threading
import time
import zipfile
from pathlib import Path
from typing import Any
from uuid import uuid4
import gradio as gr
import numpy as np
from studio.adapters.base import GenerateRequest
from studio.adapters.health import set_visitor_ip_token
from studio.adapters.router import AdapterRouter
from studio.daw import analyze, engine, jobs, stems
from studio.model_status import is_unusable, model_state
from studio.registry import load_registry
DAW_DIR = Path(os.environ.get("DAW_DIR", Path(tempfile.gettempdir()) / "2bs-daw")).resolve()
DAW_DIR.mkdir(parents=True, exist_ok=True)
GEN_DIR = DAW_DIR / "gen"
REGISTRY = load_registry()
MOCK = os.environ.get("DAW_MOCK") == "1"
if threading.current_thread() is threading.main_thread():
engine.preload()
ELEMENTS: dict[str, dict[str, Any]] = {
"full_song": {"label": "Full Song", "models": ["minimax-music3", "ace-step", "yue2"],
"tpl": "{p}", "lyrics": True, "dur": 60, "loop": False},
"voice": {"label": "Voice A-cappella", "models": ["ace-step", "yue2", "minimax-music3"],
"tpl": "a cappella, solo lead vocals only, dry studio vocal recording, no instruments, no drums, {p}",
"neg": "instruments, drums, bass, guitar, piano, synth", "lyrics": True, "dur": 30, "loop": False},
"drums": {"label": "Drums", "models": ["ace-step", "2b-musicgen-medium", "audioldm2-pub"],
"tpl": "isolated drum kit loop only, {p}, {bpm} BPM, no melody, no bass, no vocals",
"neg": "melody, bass, vocals, pads", "dur": 8, "loop": True},
"bass": {"label": "Bass", "models": ["ace-step", "2b-musicgen-medium"],
"tpl": "isolated bass line only, {p}, {bpm} BPM, key {key}, no drums, no vocals",
"neg": "drums, vocals, pads", "dur": 8, "loop": True},
"instrument": {"label": "Instrument", "models": ["ace-step", "2b-musicgen-medium", "audioldm2-pub"],
"tpl": "solo {p}, isolated instrument recording, {bpm} BPM, key {key}, no drums",
"neg": "drums, vocals", "dur": 8, "loop": True},
"instrumental_song": {"label": "Instrumental Song", "models": ["ace-step", "2b-musicgen-medium"],
"tpl": "instrumental, {p}, {bpm} BPM, key {key}, no vocals", "neg": "vocals, singing",
"dur": 30, "loop": False},
"melody_solo": {"label": "Melody (solo)", "models": ["ace-step", "2b-musicgen-medium"],
"tpl": "single monophonic lead melody, {p}, {bpm} BPM, key {key}, solo instrument, no chords, no drums",
"neg": "drums, chords, vocals", "dur": 8, "loop": True},
"melody_ensemble": {"label": "Melody Ensemble", "models": ["ace-step", "2b-musicgen-medium"],
"tpl": "melodic ensemble section playing together, {p}, {bpm} BPM, key {key}, layered harmonies, no drums",
"neg": "drums, vocals", "dur": 8, "loop": True},
"sfx": {"label": "SFX", "models": ["tangoflux", "audiogen", "audioldm2-pub", "procedural"],
"tpl": "{p}", "dur": 4, "loop": False},
"transition": {"label": "Übergang", "models": ["procedural", "tangoflux", "audiogen"],
"tpl": "{p}, transition sound effect", "dur": 4, "loop": False},
}
def _model_info(mid: str) -> dict[str, Any]:
if mid == "procedural":
return {"id": mid, "label": "2B Synth (prozedural)", "badge": "free", "badge_label": "frei", "state": "live"}
m = REGISTRY.get(mid)
try:
state = model_state(mid, REGISTRY)
except Exception: # noqa: BLE001
state = "unknown"
return {"id": mid, "label": m.label, "badge": m.license_badge, "badge_label": m.license_badge_label,
"state": state, "license": m.license, "max_s": m.max_duration_s}
def daw_catalog() -> str:
"""Elemente, passende Modelle (mit Lizenz-Badge/Status) und Plugin-Status."""
out = {"elements": [], "plugins": engine.plugin_status(), "mock": MOCK,
"stems": {"available": stems.available(), "max_s": stems.MAX_SECONDS, "warm": stems._MODEL is not None}}
for key, el in ELEMENTS.items():
out["elements"].append({"id": key, "label": el["label"], "lyrics": bool(el.get("lyrics")),
"dur": el["dur"], "loop": el["loop"], "models": [_model_info(m) for m in el["models"]]})
return json.dumps(out)
def _allowed(path: str) -> Path:
p = Path(path).resolve()
roots = [DAW_DIR, Path(os.environ.get("GRADIO_TEMP_DIR", Path(tempfile.gettempdir()) / "gradio")).resolve()]
if not any(str(p).startswith(str(r) + os.sep) for r in roots) or not p.is_file():
raise gr.Error("Datei nicht erlaubt oder nicht gefunden.")
return p
def _out(data: np.ndarray, sr: int, stem: str, **meta: Any) -> str:
dest = DAW_DIR / f"{stem}-{uuid4().hex[:10]}.wav"
engine.save(dest, data, sr)
return json.dumps({"path": str(dest), "url": f"/gradio_api/file={dest}", **engine.stats(data, sr), **meta})
def _visitor(request: gr.Request | None) -> None:
try:
token = request.headers.get("x-ip-token") if request is not None else None
except Exception: # noqa: BLE001
token = None
set_visitor_ip_token(token)
def _run_model(mid: str, prompt: str, negative: str, lyrics: str, seconds: float, seed: int) -> tuple[np.ndarray, int, str]:
if mid == "procedural" or MOCK:
kind = engine.procedural_kind(prompt)
return engine.procedural(kind, max(0.5, seconds), seed=seed), engine.SR, f"2B Synth · {kind}"
m = REGISTRY.get(mid)
dur = int(max(1, min(m.max_duration_s, round(seconds))))
req = GenerateRequest(model_id=mid, genre_style_instruments=prompt[:900], lyrics=lyrics, negative=negative,
duration_s=max(1, min(300, dur)),
advanced={"seed": seed, "instrumental": not lyrics.strip()})
track = AdapterRouter(REGISTRY, GEN_DIR).generate(req)
data, sr = engine.load(track.path)
return data, sr, m.label
def daw_generate(element: str, model: str, prompt: str, negative: str, lyrics: str, seconds: float, bpm: float,
key: str, seed: float, request: gr.Request) -> str:
"""Element craften. model='auto' probiert die passenden Modelle der Reihe nach."""
_visitor(request)
el = ELEMENTS.get(element)
if el is None:
raise gr.Error("Unbekanntes Element.")
bpm = float(bpm or 120)
p = el["tpl"].format(p=(prompt or el["label"]).strip(), bpm=int(bpm), key=(key or "C minor").strip())
neg = ", ".join(x for x in (negative.strip(), el.get("neg", "")) if x)
lyr = lyrics if el.get("lyrics") else ""
seconds = float(seconds or el["dur"])
candidates = [model] if model and model != "auto" else [
m for m in el["models"] if m == "procedural" or not is_unusable(model_state(m, REGISTRY))]
errors: list[str] = []
started = time.monotonic()
for mid in candidates[:3]:
try:
data, sr, label = _run_model(mid, p, neg, lyr, seconds, int(seed or 0))
break
except Exception as error: # noqa: BLE001
errors.append(f"{mid}: {str(error)[:160]}")
else:
raise gr.Error("Kein Modell lieferte Audio. " + " | ".join(errors)[:600])
if el["loop"]:
data = engine.stretch_to_bpm(data, sr, _bpm_guess(data, sr), bpm)
data = engine.fit(data, sr, seconds, loop=el["loop"])
if engine.peak_db(data) > -0.3: # Clipping-Schutz für Worker, die über 0 dBFS liefern
data, _ = engine.fx_normalize(data, sr, peak=-1.0)
info = _model_info(mid)
return _out(data, sr, element, model=mid, model_label=label, badge=info["badge"], badge_label=info["badge_label"],
prompt=p, elapsed_s=round(time.monotonic() - started, 1), fallback_errors=errors)
def _bpm_guess(data: np.ndarray, sr: int) -> float | None:
try:
from studio.studio_room.dsp import _cheap_bpm
return _cheap_bpm(data.mean(axis=0), sr)
except Exception: # noqa: BLE001
return None
def daw_transition(path_a: str, path_b: str, prompt: str, model: str, seconds: float, request: gr.Request) -> str:
"""Übergang zwischen zwei Clips: SFX (Riser/Impact) per Prompt + Crossfade-Vorschau."""
_visitor(request)
seconds = float(seconds or 4)
res = json.loads(daw_generate("transition", model or "auto", prompt or "riser", "", "", seconds, 120, "", 0, request))
if path_a and path_b:
a, sr = engine.load(_allowed(path_a))
b, srb = engine.load(_allowed(path_b))
if srb == sr:
xf = engine.mix_crossfade(a[:, -int(seconds * sr):], b[:, :int(seconds * sr)], sr, seconds / 2)
res["crossfade"] = json.loads(_out(xf, sr, "xfade"))
return json.dumps(res)
def daw_fx(path: str, effect: str, params: str, start: float, end: float) -> str:
"""Destruktiver Effekt auf einen Clip-Bereich; liefert neue Datei (Original bleibt)."""
data, sr = engine.load(_allowed(path))
s = max(0, int(float(start or 0) * sr))
e = int(float(end) * sr) if end and float(end) > 0 else data.shape[1]
seg = data[:, s:e]
if seg.shape[1] < 64:
raise gr.Error("Bereich zu kurz.")
before = engine.stats(seg, sr)
out, eng = engine.apply_effect(seg, sr, effect, json.loads(params or "{}"))
return _out(out, sr, f"fx-{effect}", engine=eng, before=before)
def daw_encode(path: str, fmt: str) -> str:
"""Mixdown (WAV aus dem Browser) → MP3 320 oder WAV 24-bit."""
data, sr = engine.load(_allowed(path))
if fmt == "mp3":
dest = DAW_DIR / f"2b-studio-mix-{uuid4().hex[:8]}.mp3"
engine.save(dest, data, sr, quality=320)
else:
dest = DAW_DIR / f"2b-studio-mix-{uuid4().hex[:8]}.wav"
engine.save(dest, data, sr)
return json.dumps({"path": str(dest), "url": f"/gradio_api/file={dest}", **engine.stats(data, sr)})
def _safe_name(name: str) -> str:
return re.sub(r"[^A-Za-z0-9_-]+", "-", name or "projekt").strip("-")[:48] or "projekt"
def daw_save(project: str, title: str, oauth_token: gr.OAuthToken | None = None,
profile: gr.OAuthProfile | None = None) -> str:
"""Projekt (JSON + Audio) als .2bproj.zip; mit HF-Login zusätzlich ins private Dataset des Nutzers."""
proj = json.loads(project)
name = _safe_name(title)
dest = DAW_DIR / f"{name}-{time.strftime('%Y%m%d-%H%M%S')}.2bproj.zip"
with zipfile.ZipFile(dest, "w", zipfile.ZIP_DEFLATED) as z:
for asset in proj.get("assets", []):
src = _allowed(asset["path"])
arc = f"audio/{asset['id']}{src.suffix}"
z.write(src, arc)
asset["path"] = arc
asset.pop("url", None)
z.writestr("project.json", json.dumps(proj, ensure_ascii=False, indent=1))
result: dict[str, Any] = {"path": str(dest), "url": f"/gradio_api/file={dest}", "hf": None}
if oauth_token is not None and profile is not None and os.environ.get("DAW_HF_SAVE", "1") == "1":
try:
from huggingface_hub import HfApi
user = str(profile.username or "")
if not USER_RX.fullmatch(user):
raise ValueError("Ungültiger HF-Benutzername.")
api = HfApi(token=oauth_token.token)
repo = _private_repo(api, user)
api.upload_file(path_or_fileobj=str(dest), path_in_repo=f"projects/{dest.name}", repo_id=repo,
repo_type="dataset", commit_message=f"2B-Studio: {name}")
result["hf"] = f"https://huggingface.co/datasets/{repo}"
except Exception as error: # noqa: BLE001
result["hf_error"] = str(error)[:200]
return json.dumps(result)
def daw_load(path: str) -> str:
""".2bproj.zip entpacken → Projekt-JSON mit abspielbaren URLs."""
src = _allowed(path)
target = DAW_DIR / f"proj-{uuid4().hex[:10]}"
target.mkdir()
with zipfile.ZipFile(src) as z:
for member in z.namelist():
if member.startswith("/") or ".." in member:
raise gr.Error("Ungültige Projektdatei.")
z.extractall(target)
proj = json.loads((target / "project.json").read_text(encoding="utf-8"))
for asset in proj.get("assets", []):
p = (target / asset["path"]).resolve()
asset["path"] = str(p)
asset["url"] = f"/gradio_api/file={p}"
return json.dumps(proj)
STEM_LABELS = {"vocals": "Vocals", "drums": "Drums", "bass": "Bass", "other": "Other"}
def daw_stems(path: str, start: float, end: float) -> str:
"""Startet Demucs-Stem-Trennung (htdemucs, CPU) als Job → {'job': id}."""
if not stems.available():
raise gr.Error("Stem-Trennung ist auf diesem Server nicht installiert.")
data, sr = engine.load(_allowed(path))
s = max(0, int(float(start or 0) * sr))
e = int(float(end) * sr) if end and float(end) > 0 else data.shape[1]
seg = data[:, s:e]
if seg.shape[1] / sr > stems.MAX_SECONDS + 0.5:
raise gr.Error(f"Clip ist {seg.shape[1] / sr:.0f} s lang – auf CPU max. {int(stems.MAX_SECONDS)} s. Bitte vorher schneiden.")
def work(report):
parts, msr = stems.separate(seg, sr, report)
out = []
for name, audio in parts.items():
info = json.loads(_out(audio, msr, f"stem-{name}"))
out.append({"stem": name, "label": STEM_LABELS.get(name, name), **info})
return out
secs = seg.shape[1] / sr
try:
job = jobs.start("stems", work, est_s=stems.estimate_s(secs), timeout_s=max(240.0, secs * 6 + 120))
except RuntimeError as error:
raise gr.Error(str(error)) from None
return json.dumps({"job": job, "seconds": round(secs, 1), "est_s": round(stems.estimate_s(secs))})
def daw_job(job_id: str) -> str:
job = jobs.get(job_id)
if job is None:
raise gr.Error("Job unbekannt oder abgelaufen.")
return json.dumps({k: job[k] for k in ("id", "state", "progress", "msg", "result", "error", "queue", "wait_s", "eta_s")})
def daw_job_cancel(job_id: str) -> str:
return json.dumps({"cancelled": jobs.cancel(job_id)})
_ANALYSIS: dict[str, dict] = {}
def daw_analyze(path: str, start: float, end: float) -> str:
"""Tempo und Tonart eines Clip-Bereichs (CPU, max. 60 s)."""
src = _allowed(path)
key = f"{src}|{float(start or 0):.2f}|{float(end or 0):.2f}"
if key not in _ANALYSIS:
data, sr = engine.load(src)
s = max(0, int(float(start or 0) * sr))
e = int(float(end) * sr) if end and float(end) > 0 else data.shape[1]
_ANALYSIS[key] = analyze.analyze(data[:, s:e], sr)
if len(_ANALYSIS) > 512:
_ANALYSIS.pop(next(iter(_ANALYSIS)))
return json.dumps(_ANALYSIS[key])
def daw_whoami(profile: gr.OAuthProfile | None = None) -> str:
user = str(getattr(profile, "username", "") or "") if profile is not None else ""
ok = bool(user) and os.environ.get("DAW_HF_SAVE", "1") == "1"
return json.dumps({"user": user or None, "cloud": ok})
USER_RX = re.compile(r"[A-Za-z0-9][A-Za-z0-9._-]{0,95}")
SESSION_RX = re.compile(r"[a-z0-9]{6,32}")
_UPLOADED: dict[str, set[str]] = {}
def _private_repo(api: Any, user: str) -> str:
"""Privates Dataset <user>/2b-studio-projects anlegen/prüfen – niemals öffentlich schreiben."""
if not USER_RX.fullmatch(user):
raise ValueError("Ungültiger HF-Benutzername.")
repo = f"{user}/2b-studio-projects"
api.create_repo(repo, repo_type="dataset", private=True, exist_ok=True)
info = api.repo_info(repo, repo_type="dataset")
if not getattr(info, "private", False):
raise PermissionError(f"{repo} ist öffentlich – aus Sicherheitsgründen wird dort nichts gespeichert.")
return repo
def daw_autosave(project: str, title: str, session: str, oauth_token: gr.OAuthToken | None = None,
profile: gr.OAuthProfile | None = None) -> str:
"""Cloud-Autosave: Projekt-JSON + noch nicht hochgeladene Audios (inhaltsadressiert) in EINEM Commit
nach autosave/<session>/ im privaten Dataset. Jede Browser-Sitzung schreibt nur ihren eigenen Ordner."""
if oauth_token is None or profile is None or os.environ.get("DAW_HF_SAVE", "1") != "1":
return json.dumps({"ok": False, "reason": "login"})
if not SESSION_RX.fullmatch(session or ""):
raise gr.Error("Ungültige Sitzungs-ID.")
import hashlib
from huggingface_hub import CommitOperationAdd, HfApi
user = str(profile.username or "")
api = HfApi(token=oauth_token.token)
try:
repo = _private_repo(api, user)
except Exception as error: # noqa: BLE001
return json.dumps({"ok": False, "reason": "repo", "error": str(error)[:200]})
proj = json.loads(project)
base = f"autosave/{session}"
known = _UPLOADED.setdefault(repo, set())
if not known:
try:
known.update(f for f in api.list_repo_files(repo, repo_type="dataset") if f.startswith("autosave/"))
except Exception: # noqa: BLE001
pass
ops = []
for asset in proj.get("assets", []):
src = _allowed(asset["path"])
digest = hashlib.sha1(src.read_bytes()).hexdigest()[:20]
arc = f"{base}/audio/{digest}{src.suffix}"
asset["path"], asset["sha1"] = arc, digest
asset.pop("url", None)
if arc not in known:
ops.append(CommitOperationAdd(path_in_repo=arc, path_or_fileobj=str(src)))
proj["saved_at"] = time.time()
ops.append(CommitOperationAdd(path_in_repo=f"{base}/project.json",
path_or_fileobj=json.dumps(proj, ensure_ascii=False).encode("utf-8")))
try:
info = api.create_commit(repo, operations=ops, repo_type="dataset",
commit_message=f"2B-Studio Autosave: {_safe_name(title)} ({session})")
except Exception as error: # noqa: BLE001
return json.dumps({"ok": False, "reason": "commit", "error": str(error)[:200]})
known.update(op.path_in_repo for op in ops)
return json.dumps({"ok": True, "repo": f"https://huggingface.co/datasets/{repo}", "path": base,
"files": len(ops), "commit": getattr(info, "oid", None), "at": proj["saved_at"]})
def _prewarm_stems() -> None:
if stems.available() and os.environ.get("DAW_STEMS_PREWARM", "1") == "1":
def warm():
time.sleep(3)
try:
stems.warm()
except Exception: # noqa: BLE001
pass
threading.Thread(target=warm, daemon=True, name="stems-prewarm").start()
def register(app: gr.Blocks) -> None:
"""Alle Endpunkte registrieren (innerhalb von `with gr.Blocks()`)."""
gr.api(daw_catalog, api_name="daw_catalog")
gr.api(daw_generate, api_name="daw_generate", concurrency_limit=6)
gr.api(daw_transition, api_name="daw_transition", concurrency_limit=6)
gr.api(daw_fx, api_name="daw_fx", concurrency_limit=2)
gr.api(daw_encode, api_name="daw_encode", concurrency_limit=2)
gr.api(daw_save, api_name="daw_save", concurrency_limit=2)
gr.api(daw_load, api_name="daw_load", concurrency_limit=2)
gr.api(daw_stems, api_name="daw_stems", concurrency_limit=2)
gr.api(daw_job, api_name="daw_job", concurrency_limit=None)
gr.api(daw_job_cancel, api_name="daw_job_cancel", concurrency_limit=None)
gr.api(daw_analyze, api_name="daw_analyze", concurrency_limit=2)
gr.api(daw_whoami, api_name="daw_whoami", concurrency_limit=None)
gr.api(daw_autosave, api_name="daw_autosave", concurrency_limit=2)
_prewarm_stems()