fix(lora): resume from checkpoint fails due to strict state_dict loading

[duchengyao]

Is this PR adding new feature or fix a BUG?

Fix BUG.

Is this pull request related to any issue? If yes, please link the issue.

#1295

Problem

TextToSemantic.on_save_checkpoint intentionally saves only LoRA parameters to reduce checkpoint size (~100MB vs ~9GB). However, this causes Lightning's restore_model() to fail during resume because load_state_dict is called with strict=True, and the frozen base model weights are missing from the saved state_dict.

Fix

Override load_state_dict in TextToSemantic to always use strict=False.

• Only LoRA weights are updated (base weights remain from from_pretrained)
• Optimizer states and LR schedulers are correctly restored
• Full fine-tuning is unaffected (non-LoRA checkpoints have all keys)

Before

RuntimeError: Error(s) in loading state_dict for TextToSemantic:
	Missing key(s): model.embeddings.weight, model.codebook_embeddings.weight, ...

After

LoRA training resumes from checkpoint with no errors.

Files changed

• fish_speech/models/text2semantic/lit_module.py — add load_state_dict override

Testing

Tested on a single RTX 4090 (48GB) with LoRA r=8 on s2-pro:

1. Train for N steps → interrupt
2. Re-run training script → resume from latest checkpoint successfully
3. Loss curve is continuous, optimizer/scheduler states restored