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: let YuE2 finish a song instead of cutting it at the asked length (#13346)

Daniel Han85ac36d3c182
diff --git a/studio/backend/core/inference/audio_cpp_models.py b/studio/backend/core/inference/audio_cpp_models.py
index 7aff439edb7..b985e2dcb9b 100644
--- a/studio/backend/core/inference/audio_cpp_models.py
+++ b/studio/backend/core/inference/audio_cpp_models.py
@@ -194,6 +194,8 @@ class MusicSpec:
# The runtime's default seed is fixed, so repeat runs would be identical.
fixed_seed: bool = True
rtf: float = 4.0
+ # Token budget as a multiple of the asked length, for a model that ends the song itself.
+ overrun: float = 1.0
def mode(self, mode_id: Optional[str]) -> Optional[MusicMode]:
return next((m for m in self.modes if m.id == mode_id), None)
@@ -263,6 +265,8 @@ MUSIC_SPECS: dict[str, MusicSpec] = {
# Empty lyrics still sing; only this tag stays wordless.
instrumental_lyrics = "[Instrumental]",
rtf = 6.0,
+ # 1.0x cut long lyrics mid-song (#13337); 240 s x 1.5 = the runtime's 9000-token default.
+ overrun = 1.5,
),
}
diff --git a/studio/backend/core/inference/audio_cpp_music.py b/studio/backend/core/inference/audio_cpp_music.py
index 2e19537e6d7..1c851627df9 100644
--- a/studio/backend/core/inference/audio_cpp_music.py
+++ b/studio/backend/core/inference/audio_cpp_music.py
@@ -52,7 +52,7 @@ def seconds_text(value: float) -> str:
def timeout_seconds(family: Optional[str], seconds: float, variations: int, cpu: bool) -> float:
music = MUSIC_SPECS.get(str(family or ""))
- rtf = music.rtf if music is not None else 4.0
+ rtf = music.rtf * music.overrun if music is not None else 4.0
work = rtf * max(0.0, float(seconds)) * max(1, int(variations)) * (_CPU_FACTOR if cpu else 1.0)
return max(_TIMEOUT_FLOOR, min(_TIMEOUT_CEILING, _TIMEOUT_FLOOR + work))
@@ -192,9 +192,10 @@ def song_request(
request["text"] = lyrics
request_options["style"] = description
request_options["lyrics"] = lyrics
- # Length is the semantic token budget (25 fps); the default 200-frame floor outlasts short asks.
+ # Length caps the semantic tokens (25 fps); the default 200-frame floor outlasts short asks.
frames = int(round(seconds * 25))
- request_options["semantic_max_tokens"] = frames
+ overrun = MUSIC_SPECS["yue2"].overrun
+ request_options["semantic_max_tokens"] = int(round(seconds * overrun * 25))
request_options["semantic_min_tokens"] = min(200, frames)
return _finish(model, request, request_options, seed)
if instrumental and music is not None and music.instrumental_lyrics is not None:
diff --git a/studio/backend/tests/test_audio_cpp_models.py b/studio/backend/tests/test_audio_cpp_models.py
index 2c03e79ecff..71355f62ed7 100644
--- a/studio/backend/tests/test_audio_cpp_models.py
+++ b/studio/backend/tests/test_audio_cpp_models.py
@@ -1438,7 +1438,7 @@ def test_yue2_takes_the_description_as_its_required_style():
assert request["options"] == {
"style": "rock, male vocal",
"lyrics": "[chorus] oh",
- "semantic_max_tokens": 2048,
+ "semantic_max_tokens": 3072,
"semantic_min_tokens": 200,
}
with pytest.raises(RuntimeError, match = "style"):
@@ -2108,11 +2108,13 @@ def test_yue2_length_follows_the_requested_duration():
b, srv = _backend_with(_speech_model("yue2"), MUSIC_REPLY)
b.generate_audio_response("[chorus] oh", instructions = "rock", max_new_tokens = 125)
options = srv.calls[0][1]["request"]["options"]
- # 5 s at 25 frames per second, with the 200-frame default lowered to fit.
- assert options["semantic_max_tokens"] == 125 and options["semantic_min_tokens"] == 125
+ # 5 s at 25 frames per second, with the 200-frame default lowered to fit; the cap is 1.5x.
+ assert options["semantic_max_tokens"] == 188 and options["semantic_min_tokens"] == 125
b.generate_audio_response("[chorus] oh", instructions = "rock", max_new_tokens = 1500)
options = srv.calls[1][1]["request"]["options"]
- assert options["semantic_max_tokens"] == 1500 and options["semantic_min_tokens"] == 200
+ assert options["semantic_max_tokens"] == 2250 and options["semantic_min_tokens"] == 200
+ b.generate_audio_response("[chorus] oh", instructions = "rock", max_new_tokens = 6000)
+ assert srv.calls[2][1]["request"]["options"]["semantic_max_tokens"] == 9000
def test_music_requests_missing_a_required_field_are_a_400_message():
diff --git a/studio/backend/tests/test_audio_cpp_music_requests.py b/studio/backend/tests/test_audio_cpp_music_requests.py
index 7dfd5e3ab21..ea9fd8a1efb 100644
--- a/studio/backend/tests/test_audio_cpp_music_requests.py
+++ b/studio/backend/tests/test_audio_cpp_music_requests.py
@@ -375,7 +375,7 @@ def test_yue2_instrumental_sends_the_instrumental_tag_not_its_style():
assert request["text"] == "[Instrumental]"
assert request["options"]["lyrics"] == "[Instrumental]"
assert request["options"]["style"] == "ambient piano"
- assert request["options"]["semantic_max_tokens"] == 500
+ assert request["options"]["semantic_max_tokens"] == 750
def test_minimax_gguf_keeps_its_body_through_the_studio():
@@ -497,6 +497,7 @@ def test_the_wait_grows_with_the_audio_asked_for_and_is_capped():
assert t(10) < t(60) < t(60, 3) < t(60, 3, cpu = True)
assert t(0) == 300.0 and t(240, 4, cpu = True) == 4 * 3600.0
assert cm.timeout_seconds("stable_audio", 10, 1, False) == 310.0
+ assert t(60) == 300.0 + 6.0 * 1.5 * 60
b = _backend(_model("yue2"))
_run(b, text = "rock", lyrics = "x", duration_s = 60)
assert b._server.calls[0][2]["timeout"] == t(60)