Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: adopt the WAV save_from_media already transcoded instead of decoding it again (#15916)

Daniel Hane5acea16e8db
diff --git a/studio/backend/core/inference/audio_inputs.py b/studio/backend/core/inference/audio_inputs.py
index 57e2e36cb17..0c55421ae53 100644
--- a/studio/backend/core/inference/audio_inputs.py
+++ b/studio/backend/core/inference/audio_inputs.py
@@ -351,6 +351,7 @@ def save_from_media(
rate = info["sample_rate"]
with wave.open(str(full), "rb") as reader:
total = reader.getnframes()
+ channels = reader.getnchannels()
first = int(start_s * rate)
if first >= total:
raise AudioInputError(
@@ -359,14 +360,31 @@ def save_from_media(
"Unsloth Studio can only cut inside the track.",
)
last = total if end_s is None else min(total, max(first + 1, int(end_s * rate)))
- reader.setpos(first)
- data = reader.readframes(last - first)
- with wave.open(str(cut), "wb") as writer:
- writer.setparams(reader.getparams())
- writer.writeframes(data)
- payload = cut.read_bytes()
+ whole = not first and last == total
+ if not whole:
+ reader.setpos(first)
+ with wave.open(str(cut), "wb") as writer:
+ writer.setparams(reader.getparams())
+ writer.writeframes(reader.readframes(last - first))
+ if whole:
+ os.replace(full, cut)
+ digest = hashlib.sha256()
+ with open(cut, "rb") as f:
+ for chunk in iter(lambda: f.read(1 << 20), b""):
+ digest.update(chunk)
+ # transcode() wrote this WAV, so transcoding it again would only rewrite the same bytes.
+ prepared = {
+ "duration_s": round((last - first) / rate, 3),
+ "sample_rate": rate,
+ "channels": channels,
+ }
return _finish_upload(
- directory, cut, hashlib.sha256(payload).hexdigest(), _clean_name(name), len(payload)
+ directory,
+ cut,
+ digest.hexdigest(),
+ _clean_name(name),
+ os.stat(cut).st_size,
+ prepared = prepared,
)
finally:
full.unlink(missing_ok = True)
@@ -385,8 +403,15 @@ def _find_by_sha(directory: Path, sha: str) -> Optional[tuple[str, dict[str, Any
def _finish_upload(
- directory: Path, tmp: Path, sha: str, name: str, size: int
+ directory: Path,
+ tmp: Path,
+ sha: str,
+ name: str,
+ size: int,
+ *,
+ prepared: Optional[dict[str, Any]] = None,
) -> tuple[dict[str, Any], bool]:
+ """``prepared``: ``tmp`` is already a ``transcode`` output with this info, so it is moved in as is."""
existing = _find_by_sha(directory, sha)
if existing is not None:
input_id, meta = existing
@@ -396,7 +421,11 @@ def _finish_upload(
input_id = uuid.uuid4().hex
wav = directory / f"{input_id}.wav"
try:
- info = transcode(tmp, wav)
+ if prepared is None:
+ info = transcode(tmp, wav)
+ else:
+ os.replace(tmp, wav)
+ info = prepared
now = _now()
meta = {
"name": name,
diff --git a/studio/backend/tests/test_audio_inputs.py b/studio/backend/tests/test_audio_inputs.py
index cb97d9b344c..5672accb6b5 100644
--- a/studio/backend/tests/test_audio_inputs.py
+++ b/studio/backend/tests/test_audio_inputs.py
@@ -4,7 +4,9 @@
from __future__ import annotations
import asyncio
+import hashlib
import io
+import json
import math
import os
import time
@@ -377,3 +379,47 @@ def test_the_sweep_reaps_a_prepared_upload_a_crash_left_behind():
os.utime(path, (time.time() - age, time.time() - age))
audio_inputs.sweep()
assert not stale.exists() and waiting.exists()
+
+
+ "fmt, codec, rate, layout, start_s, end_s",
+ [
+ ("mp4", "aac", 44100, "stereo", 0.25, 1.1),
+ ("wav", "pcm_s16le", 96000, "mono", 0.0, None),
+ ("flac", "flac", 16000, "5.1", 0.5, None),
+ ],
+)
+def test_a_media_cut_is_decoded_once_and_stored_as_cut(
+ tmp_path, monkeypatch, fmt, codec, rate, layout, start_s, end_s
+):
+ src = tmp_path / f"clip.{fmt}"
+ src.write_bytes(encode(fmt, codec, rate, layout, 1.5))
+ whole = tmp_path / "whole.wav"
+ audio_inputs.transcode(src, whole)
+ with wave.open(str(whole)) as r:
+ out_rate, channels, total = r.getframerate(), r.getnchannels(), r.getnframes()
+ first = int(start_s * out_rate)
+ last = total if end_s is None else min(total, int(end_s * out_rate))
+ r.setpos(first)
+ expected = io.BytesIO()
+ with wave.open(expected, "wb") as w:
+ w.setparams(r.getparams())
+ w.writeframes(r.readframes(last - first))
+
+ calls = []
+ real = audio_inputs.transcode
+ monkeypatch.setattr(audio_inputs, "transcode", lambda *a, **k: calls.append(a) or real(*a, **k))
+ record, created = audio_inputs.save_from_media(src, "clip", start_s, end_s)
+ assert created and len(calls) == 1
+ stored = audio_inputs.inputs_dir() / f"{record['id']}.wav"
+ assert stored.read_bytes() == expected.getvalue()
+ assert record["sample_rate"] == out_rate and record["channels"] == channels
+ assert record["duration_s"] == round((last - first) / out_rate, 3)
+ meta = json.loads((audio_inputs.inputs_dir() / f"{record['id']}.json").read_text())
+ assert meta["sha256"] == hashlib.sha256(expected.getvalue()).hexdigest()
+ assert meta["bytes"] == len(expected.getvalue())
+
+ again, created = audio_inputs.save_from_media(src, "clip", start_s, end_s)
+ assert not created and again["id"] == record["id"]
+ leftovers = [p.name for p in audio_inputs.inputs_dir().iterdir() if p.name.startswith(".")]
+ assert leftovers == []