raise ask cap 512->768 tokens (still ~30s, under ZeroGPU 60s cap)

revert GPU duration to default (free ZeroGPU caps <120s); keep 512-token cap

raise generation cap (200→512 ask, 160→384 arena) + GPU duration 120s

raise generation cap (200→512 ask, 160→384 arena) + GPU duration 120s

resilient umap transform (point=None fallback) + pin umap-learn 0.5.12

resilient umap transform (point=None fallback) + pin umap-learn 0.5.12

fix device mismatch: move gating head to GPU alongside student

ZeroGPU: load student on CPU at startup with cuda.is_available guarded False

ZeroGPU: load student on CPU at startup with cuda.is_available guarded False

ZeroGPU: load student on CPU at startup (no device_map=auto), move to GPU in handler

back to torch backend + ZeroGPU; drop llama-cpp-python

fix n_threads oversubscription on cpu-basic (host reports 16+, container has 2)

llama.cpp speedups: n_threads, n_ctx 1024, gate off hot path, max_tokens 96

llama.cpp backend: requirements.txt

llama.cpp backend: _probe.py

fix deps: transformers>=5 (gradio 6.18 needs hub>=1.0)

drop stale __pycache__

Custom gr.Server frontend + 6-teacher model (Nemotron), token streaming, arena

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload _html.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload _glb.py with huggingface_hub

fix: gradio model 3d

fix: iframe Three.js, model response output, UI labels

feat: Three.js 3D soul space (remove Plotly UMAP)

fix: gating.pt path (root, not final/)

fix: gradio 5.33.2 (python 3.13 compat)

Initial Space — OFA soul transfer

Initial Space — OFA soul transfer

initial commit