v0.8.1

What's new

Added

• examples/config-mps.yaml: Recommended training config for Apple Silicon — batch_size=2, use_fp16=false, use_gradient_checkpointing=true, workers=0.
• scripts/profile_mps.py: Profiles CPU fallbacks during a VAE forward+backward pass on MPS.
• ctx_loss console logging: Context dim loss shown in training progress line as | Ctx: 0.0123.
• ctx_loss graph curve: Plotted as a separate curve in training_losses.png.
• ctx_loss_weight config param: Tunes the relative weight of context-dim v-prediction loss.

Fixed

• MPS cache clearing: torch.mps.empty_cache() added alongside CUDA clearing in VAETrainer and FlowTrainer.
• MPS adaptive pooling: Replaced ad-hoc try/except with mps_safe_pool2d from fluxflow.utils.mps.
• v-prediction loss target: Flow trainer was computing loss against clean x0 despite prediction_type="v_prediction". Now correctly computes v = alpha_t * noise - sigma_t * x0.
• Context dim train/inference mismatch: Training now noises all dims uniformly to match inference.
• Context dim loss scale: VAE and context dims normalised independently to prevent ~10x loss imbalance.
• Gradient clipping: Replaced broken adaptive formula with straightforward clip_grad_norm_.
• GAN instance noise never applied: Inverted guard removed; noise now always applied.
• GAN adaptive weight explosion at startup: _compute_adaptive_weight clamped to max_weight=5.0.
• Discriminator trained on deterministic latents: Changed to training=True with torch.no_grad() for reparameterisation.
• ctx_loss normalisation scale: Now normalises by v-target's own std for consistent scale across timesteps.

Changed

• fluxflow dependency bumped to >=0.8.1: Required for fluxflow.utils.mps.mps_safe_pool2d.