Labyricorn/voicebox
1
0
Fork 0
mirror of https://github.com/jamiepine/voicebox.git synced 2026-09-29 07:05:14 -07:00
Code Issues Packages Projects Releases Wiki Activity

fix: load model into local var before patching to avoid half-initialised state

Browse Source 
Apply local-var-then-assign pattern to chatterbox_backend.py (multilingual)
to match the turbo backend. Also use _current_model_size fallback in
unload, delete, and status endpoints for consistent Qwen model size checks.
This commit is contained in:
James Pine
2026-03-13 05:40:18 -07:00
parent bfd7b815a5
commit 2f535a772f
3 changed files with 33 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
backend/backends/chatterbox_turbo_backend.py
+9 -4
View File
@@ -154,6 +154,8 @@ class ChatterboxTurboTTSBackend:
# Monkey-patch torch.load for CPU loading. The model's .pt files
# were saved on CUDA; from_local() doesn't pass map_location
# so loading on CPU fails without this.
# Load into a local var, apply patches, then publish to
# self.model so a failed patch doesn't leave us half-initialised.
if device == "cpu":
_orig_torch_load = torch.load
@@ -164,13 +166,13 @@ class ChatterboxTurboTTSBackend:
with ChatterboxTurboTTSBackend._load_lock:
torch.load = _patched_load
try:
self.model = ChatterboxTurboTTS.from_local(
model = ChatterboxTurboTTS.from_local(
local_path, device,
)
finally:
torch.load = _orig_torch_load
else:
self.model = ChatterboxTurboTTS.from_local(
model = ChatterboxTurboTTS.from_local(
local_path, device,
)
@@ -191,7 +193,7 @@ class ChatterboxTurboTTSBackend:
import types
# Patch S3Tokenizer (used by s3gen.tokenizer)
_tokzr = self.model.s3gen.tokenizer
_tokzr = model.s3gen.tokenizer
_orig_log_mel = _tokzr.log_mel_spectrogram.__func__
def _f32_log_mel(self_tokzr, audio, padding=0):
@@ -203,7 +205,7 @@ class ChatterboxTurboTTSBackend:
_tokzr.log_mel_spectrogram = types.MethodType(_f32_log_mel, _tokzr)
# Patch VoiceEncoder
_ve = self.model.ve
_ve = model.ve
_orig_ve_forward = _ve.forward.__func__
def _f32_ve_forward(self_ve, mels):
@@ -211,6 +213,9 @@ class ChatterboxTurboTTSBackend:
_ve.forward = types.MethodType(_f32_ve_forward, _ve)
# Only publish after all patches succeed
self.model = model
logger.info("Chatterbox Turbo TTS loaded successfully")
except ImportError as e: