win10/QMTAP-0.8B-exp

Qwen3.5 DreamVAE AR

This architecture is rebuilt as an autoregressive Qwen3.5-compatible model with native DreamVAE memory.

For a full engineering and research-oriented description, see ARCHITECTURE_GUIDE.md.

• Base behavior follows Qwen3.5 causal next-token loss via self.loss_function.
• No mask-denoising, block parallel denoising, or self-speculative parallel denoising path is implemented.
• Dream memory writes during labeled AR loss are not visible to later layers in the same forward pass by default (memory_training_writes_visible=false) to avoid same-sequence target leakage.
• Persistent memory is stored as safetensors under memory_model/states with key/latent/strength/age/source/modality/timestamp/checksum metadata.
• Inference automatically loads the latest persistent state. A direct unlabeled forward commits its new runtime memory automatically, while generate() commits once after the complete generation instead of once per token.
• Training and labeled evaluation never write persistent memory. Persistent slots are pruned automatically from configurable strength, age, and maximum-slot conditions; explicit removal APIs remain available only as administrative overrides.
• Persistent state files and index.json use a cross-process lock and atomic replacement; startup repairs index entries for completed state files left by an interrupted writer.