feat(gpt): add GPT-5.4 support

fix(texture): reuse pipeline for IP adapter

fix(zerogpu): restore Blackwell runtime patches

update

update

update

update

switch to cu128 wheels (sm_120 cubin) for Blackwell support

diag: per-frame timing in render_gs_frames/render_mesh_frames

diag: add per-stage timing in image_to_3d to find sm_120 slowdown source

fix(zerogpu): bump heavy GPU calls to duration=180 (PTX JIT eats default 60s on sm_120)

fix(sm_120): disable xformers flash3 dispatcher (no FA3 kernel for Blackwell)

fix(zerogpu): neutralize warp.init in parent to prevent fork-after-cuInit hang

fix: re-import spaces BEFORE trellis/warp imports (spaces 0.50 strict check)

debug: inject ZeroGPU stderr probes (schedule/handler/worker_init)

deps: switch open3d -> open3d-cpu (no CUDA .so, avoids ZeroGPU fork pollution)

fix: force open3d to CPU-only to keep ZeroGPU worker CUDA context clean

debug: remove queue=False and add pre-dispatch print

update

debug: add diagnostic prints in preprocess_image_fn to locate hang

fix: disable gradio SSR to avoid hang on HF Spaces

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

update

Merge branch 'new-branch'