- Add attention_dispatch module with sdpa, flash_attn (Tri Dao FA2/FA3),
and sage (auto-detects v3 Blackwell/v2 Ampere+) backends
- Add attention and memory_mode options to loader node INPUT_TYPES
- Replace dual-path model init with meta device + assign=True loading
- Add handle_post_inference_memory with cache_gpu/cpu_offload/unload modes
- Update all inference nodes to use centralized memory handler
- Set flash_attn + cpu_offload as defaults in all workflow JSONs
PozzettiAndrea
71bb3f86bb