fix: authenticate hf_hub_download with HF_TOKEN env var (needed for private model repo)

fix: load middle slice instead of slice 0 — edge slices show only anatomy tip, center slice shows full knee

fix: padding_right=0 was zeroing entire kspace mask — set to W when 0 (means no right cutoff)

fix: pre-scale kspace to O(1) before VarNet — NormUnet.norm has no epsilon in std division, causing NaN on CPU when absmax~4e-4

fix: remove num_low_frequencies from model call — match training (line 71 train script)

debug: add tensor shape/nan prints + float32 cast + ssim ndarray fix

fix: pass num_low_frequencies (not max_value) to VarNet forward — fixes NaN output

fix: _w helper uses single unsqueeze for fastmri_ssim 3D input

fix: single unsqueeze for batch dim in _run_model (fixes ValueError 5-dim unpack)

fix: pass attrs dict to VarNetDataTransform (fixes padding_left TypeError)

Fix: remove middleware after launch, fix VarNetDataTransform slice_num arg

Fix: add REST routes after launch() on real app instance

Add @spaces.GPU decorator to satisfy spaces library startup check

Switch to Gradio 5.35.0, disable SSR

Pin gradio 4.44.1, restore launch call

Fix: let HF Spaces handle launch, disable SSR via env var

Fix: disable SSR mode for HF Spaces free tier

Add E2E-VarNet reconstruction API