Skip to content

Fix speaker_encoder silent decay under AdamW weight decay in SFT #351

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
MR-AI-HuiMin wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix speaker_encoder silent decay under AdamW weight decay in SFT #351

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 19 additions & 3 deletions finetuning/sft_12hz.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,7 +57,17 @@ def train():
dataset = TTSDataset(train_data, qwen3tts.processor, config)
train_dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=True, collate_fn=dataset.collate_fn)

optimizer = AdamW(qwen3tts.model.parameters(), lr=args.lr, weight_decay=0.01)
# speaker_encoder is only used to extract a detached speaker vector for conditioning /
# checkpoint identity. It must not be optimized: AdamW weight_decay still shrinks
# its weights even when gradients are ~0 (because of .detach()), which silently
# corrupts speaker embeddings over long SFT runs. See PR description for risk notes.
if hasattr(qwen3tts.model, "speaker_encoder") and qwen3tts.model.speaker_encoder is not None:
qwen3tts.model.speaker_encoder.requires_grad_(False)
trainable_params = [
p for n, p in qwen3tts.model.named_parameters()
if p.requires_grad and not n.startswith("speaker_encoder")
]
optimizer = AdamW(trainable_params, lr=args.lr, weight_decay=0.01)

model, optimizer, train_dataloader = accelerator.prepare(
qwen3tts.model, optimizer, train_dataloader
Expand All @@ -79,9 +89,15 @@ def train():
codec_0_labels = batch['codec_0_labels']
codec_mask = batch['codec_mask']

speaker_embedding = model.speaker_encoder(ref_mels.to(model.device).to(model.dtype)).detach()
# Extract once, lock the first observation, then reuse it for training so
# conditioning matches the vector written into codec_embedding at save time
# (CustomVoice inference also uses that fixed row, not a live encoder).
extracted = model.speaker_encoder(ref_mels.to(model.device).to(model.dtype)).detach()
if target_speaker_embedding is None:
target_speaker_embedding = speaker_embedding
target_speaker_embedding = extracted[:1].detach().clone()
speaker_embedding = target_speaker_embedding.to(
device=extracted.device, dtype=extracted.dtype
).expand(extracted.size(0), -1)

input_text_ids = input_ids[:, :, 0]
input_codec_ids = input_ids[:, :, 1]
Expand Down