Pin torch 2.11, drop libcudart preload, CUDA diagnostics

app.py

@@ -16,7 +16,6 @@ encoder/decoder overrides.
 
 from __future__ import annotations
 
-import ctypes
 import glob
 import os
 import sys
@@ -40,6 +39,23 @@ import torch  # noqa: E402
 torch.set_float32_matmul_precision("high")
 
 
+def _nvidia_fds(stage: str) -> None:
+    """Diagnostics: /dev/nvidia* handles in the GPU-less main process mean the CUDA
+    driver was touched there, which breaks ZeroGPU's forked workers."""
+    fds = set()
+    for fd in os.listdir("/proc/self/fd"):
+        try:
+            t = os.readlink(f"/proc/self/fd/{fd}")
+        except OSError:
+            continue
+        if "nvidia" in t:
+            fds.add(t)
+    print(f"[diag] {stage}: nvidia fds={sorted(fds)}", flush=True)
+
+
+_nvidia_fds("after torch import")
+
+
 # ---------------------------------------------------------------------------
 # 1. gsplat CUDA backend (JIT-compiled once at startup against the ZeroGPU
 #    image's CUDA toolkit -- gsplat has no prebuilt sm_120 wheel).
@@ -67,12 +83,6 @@ if CUDA_HOME:
     os.environ["PATH"] = os.path.join(CUDA_HOME, "bin") + os.pathsep + os.environ.get("PATH", "")
     _libdir = os.path.join(CUDA_HOME, "lib64")
     os.environ["LD_LIBRARY_PATH"] = _libdir + os.pathsep + os.environ.get("LD_LIBRARY_PATH", "")
-    for _so in sorted(glob.glob(os.path.join(_libdir, "libcudart.so*")), reverse=True):
-        try:
-            ctypes.CDLL(_so, mode=ctypes.RTLD_GLOBAL)
-            break
-        except OSError:
-            continue
 
 import torch.utils.cpp_extension as _cppext  # noqa: E402
 
@@ -85,6 +95,7 @@ from gsplat.cuda._backend import _C as _GSPLAT_C  # noqa: E402
 if _GSPLAT_C is None:
     raise RuntimeError(f"gsplat CUDA backend unavailable (CUDA_HOME={CUDA_HOME}).")
 print(f"[setup] gsplat backend ready in {time.time() - _t0:.1f}s", flush=True)
+_nvidia_fds("after gsplat")
 
 # ---------------------------------------------------------------------------
 # 2. Compatibility shims for the vendored reference code
@@ -190,6 +201,7 @@ ENCODER, DECODER, DATA_SHIM = build_model(CHECKPOINT_PATH)
 ENCODER.to("cuda")
 DECODER.to("cuda")
 print("[setup] model ready", flush=True)
+_nvidia_fds("after model build")
 
 from src.dataset.shims.crop_shim import rescale_and_crop  # noqa: E402
 from src.evaluation.metrics import compute_psnr, compute_ssim  # noqa: E402
@@ -421,6 +433,7 @@ def reconstruct(
         raise gr.Error(str(exc)) from exc
     t_sfm = time.time() - t_sfm
 
+    _nvidia_fds("after COLMAP")
     progress(0.6, desc="CodecSplat: encode -> bitstream -> decode -> render")
     viewer_ply, video, gallery, bin_path, ply_path, s = codecsplat_run(sfm["scene_dir"], float(lmb))
 

requirements.txt

@@ -1,4 +1,5 @@
-torchvision
+torch==2.11.0
+torchvision==0.26.0
 gsplat==1.5.3
 ninja
 rich