- Wrap training loop in try/finally so _unapply_lora always runs.
Without this, an exception mid-training would leave LoRALinear wrappers
in the cached DiTWrapper; a subsequent training run would then apply LoRA
on top of existing LoRA, silently doubling the effective rank.
- Fix misleading comment: diffusion.model is DiTWrapper (not DiffusionTransformer).
DiffusionTransformer is at diffusion.model.model; _apply_lora reaches it
recursively but the direct attribute is the wrapper.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds PrismAudioLoRATrainer and PrismAudioLoRALoader nodes enabling
low-rank adaptation of the DiT on paired (video features + audio) datasets.
- LoRALinear wraps nn.Linear with trainable lora_A/lora_B matrices
- Rectified flow training loop with fp16 GradScaler, AdamW, cfg dropout
- Checkpoint saving every N steps + _config.json metadata alongside weights
- _unapply_lora restores base model state after training completes
- Weight-merge loader: delta_W added in-place, no deep copy overhead
- Three target presets: attn_only, attn_ffn (default), full
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Ethanfel