Labyricorn/voicebox
1
0
Fork 0
mirror of https://github.com/jamiepine/voicebox.git synced 2026-09-19 14:50:38 -07:00
Code Issues Packages Projects Releases Wiki Activity

Fix runaway MLX Qwen audio chunks (#964)

Browse Source 
* fix runaway MLX Qwen audio chunks

* test: tighten runaway retry coverage

---------

Co-authored-by: huanghua01 <[email protected]>
This commit is contained in:
AirHua
2026-07-26 23:31:23 -07:00
committed by GitHub
co-authored by huanghua01
parent 624f6a2140
commit ed54347e81
7 changed files with 268 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files
backend/backends/__init__.py
+15
View File
@@ -56,6 +56,7 @@ class ModelConfig:
model_size: str = "default"
size_mb: int = 0
needs_trim: bool = False
retries_runaway: bool = False
supports_instruct: bool = False
languages: list[str] = field(default_factory=lambda: ["en"])
@@ -232,6 +233,10 @@ def _get_qwen_model_configs() -> list[ModelConfig]:
repo_1_7b = "Qwen/Qwen3-TTS-12Hz-1.7B-Base"
repo_0_6b = "Qwen/Qwen3-TTS-12Hz-0.6B-Base"
# mlx-audio can continue after an EOS miss with silence followed by
# codec noise. Retry only the affected text as smaller chunks.
retries_runaway = backend_type == "mlx"
return [
ModelConfig(
model_name="qwen-tts-1.7B",
@@ -240,6 +245,7 @@ def _get_qwen_model_configs() -> list[ModelConfig]:
hf_repo_id=repo_1_7b,
model_size="1.7B",
size_mb=3500,
retries_runaway=retries_runaway,
supports_instruct=False, # Base model drops instruct silently
languages=["zh", "en", "ja", "ko", "de", "fr", "ru", "pt", "es", "it"],
),
@@ -250,6 +256,7 @@ def _get_qwen_model_configs() -> list[ModelConfig]:
hf_repo_id=repo_0_6b,
model_size="0.6B",
size_mb=1200,
retries_runaway=retries_runaway,
supports_instruct=False,
languages=["zh", "en", "ja", "ko", "de", "fr", "ru", "pt", "es", "it"],
),
@@ -504,6 +511,14 @@ def engine_needs_trim(engine: str) -> bool:
return False
def engine_retries_runaway(engine: str) -> bool:
"""Whether unstable output should be retried in smaller chunks."""
for cfg in get_tts_model_configs():
if cfg.engine == engine:
return cfg.retries_runaway
return False
def engine_has_model_sizes(engine: str) -> bool:
"""Whether this engine supports multiple model sizes (only Qwen currently)."""
configs = [c for c in get_tts_model_configs() if c.engine == engine]