Add nilearn/nibabel/scipy; add get_fmri and beta_std to HTTPDynaDiffLoader

Move DynaDiff to Modal serverless endpoint; switch to CPU-only HF Space

Enable patch explorer GPU inference on HF Spaces

Set PYTHONPATH for dynadiff/model so import resolves before any thread runs

Switch to modular explorer package (scripts/explorer/)

Speed up builds: move CACHEBUST, drop pytorch-lightning conflict, add .dockerignore

Fix OOM: pin torch==2.3.0, remove VD build-time download

Pin xformers==0.0.27 to match torch 2.3.0, fix OOM build

Fix Dockerfile parse error: avoid 'from' at line start in RUN python -c

Fix DynaDiff VD weights, xformers, and brain thumbnail path resolution

Add fake nvcc stub: model/models.py imports deepspeed at top level which calls nvcc -V on import

Revert VD build-time download (OOMKilled); VD downloads at runtime in background thread

Pre-download VersatileDiffusion at build time to eliminate cold-start latency

Fix deepspeed build: unset CUDA_HOME during pip install, move ENV to post-install

Set CUDA_HOME and disable deepspeed runtime CUDA op compilation

Pin transformers==4.41.2 to fix FLAX_WEIGHTS_NAME import error

Fix dynadiff structure: diffusers repo in diffusers/ subdir