Cap duration booking at 890s (ZeroGPU max 900s); AoTI is active so all UI requests fit b92404e verified fdgfdbvdb commited on Aug 27
Guard _drop_xet_cas against deleting cwd (HF_XET_CACHE unset -> Path(.)) 636d9b1 verified fdgfdbvdb commited on Aug 27
Split fetch/apply; drop Xet CAS store after loading PinkCherry f4ebe64 verified fdgfdbvdb commited on Aug 27
Fetch PinkCherry first, free its disk footprint, then load base model (stays under 150G) 2f4cdcd verified fdgfdbvdb commited on Aug 27
Plain-HTTP temp-file download for PinkCherry (avoids Xet CAS storage double-count) 5a8b225 verified fdgfdbvdb commited on Aug 27
Revert Xet disable; loader now fetches PinkCherry via plain HTTP temp file (no CAS double-copy) 74b67de verified fdgfdbvdb commited on Aug 27
Disable Xet downloads (CAS blob + weights bust 150G storage) e8ac48d verified fdgfdbvdb commited on Aug 27
Swap swiglu fc1 halves [gate;value]->[value;gate] (fixes noise output) 1539a5f verified fdgfdbvdb commited on Aug 27
Only delete PinkCherry cache file; keep base model shards (mmap) to fix noise output 660a95a verified fdgfdbvdb commited on Aug 27
1.25x duration safety factor (eager blocks run slower than AoTI estimate) 3fa6160 verified fdgfdbvdb commited on Aug 27
Drop rope.inv_freq from remap (diffusers computes it internally) 352d3ea verified fdgfdbvdb commited on Aug 27
Load PinkCherry via key remapper (fixes state_dict mismatch) 993f6f7 verified fdgfdbvdb commited on Aug 27
Add ComfyUI->diffusers key remapper for PinkCherry weights 4731bb5 verified fdgfdbvdb commited on Aug 27
Fix auth format for spaces ZeroGPU wrapper (list of tuples) ed0cc0f verified fdgfdbvdb commited on Aug 27
Free cached weights after load to stay under the 150G storage limit 5330fe0 verified fdgfdbvdb commited on Aug 27