In-process COLMAP (thread-capped); CUDA-driver diagnostics

app.py

@@ -22,7 +22,6 @@ sys.path.insert(0, str(APP_DIR))
 # pycolmap's bundled OpenBLAS segfaults ("Bad memory unallocation") when many matcher
 # threads each spin up a multi-threaded BLAS; keep BLAS single-threaded.
 os.environ["OPENBLAS_NUM_THREADS"] = "1"
-os.environ.setdefault("OMP_NUM_THREADS", "4")
 
 import spaces  # noqa: E402  (must precede any CUDA-touching import)
 import torch  # noqa: E402
@@ -30,6 +29,26 @@ import torch  # noqa: E402
 torch.set_float32_matmul_precision("high")
 
 
+def _nvidia_fds(stage: str) -> None:
+    """Diagnostics: open /dev/nvidia* handles in the (GPU-less) main process mean the
+    CUDA driver was initialised there, which breaks ZeroGPU's forked workers."""
+    try:
+        fds = []
+        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.append(t)
+        print(f"[diag] {stage}: nvidia fds={sorted(set(fds))}", flush=True)
+    except Exception as exc:
+        print(f"[diag] {stage}: {exc}", flush=True)
+
+
+_nvidia_fds("after torch import")
+
+
 # ---------------------------------------------------------------------------
 # 2. Model
 # ---------------------------------------------------------------------------
@@ -41,7 +60,9 @@ from PIL import Image  # noqa: E402
 
 import codecsplat_pipeline as cp  # noqa: E402
 
-from colmap_sfm import IMAGE_EXT, run_sfm_subprocess  # noqa: E402
+_nvidia_fds("after pipeline import")
+
+from colmap_sfm import IMAGE_EXT, run_sfm  # noqa: E402
 
 MODEL_REPO = "pengpeng-yu/CodecSplat"
 CHECKPOINT_FILE = "codecsplat-base-dl3dv-256x448-view8-075062d0.pth"
@@ -55,6 +76,7 @@ print("[setup] building CodecSplat...", flush=True)
 ENCODER = cp.build_encoder(ckpt_path).to("cuda")
 DECODER = cp.build_decoder().to("cuda")
 print("[setup] model ready", flush=True)
+_nvidia_fds("after model build")
 
 
 def _u8(t: torch.Tensor) -> np.ndarray:
@@ -164,12 +186,13 @@ def reconstruct(images, lmb: float = 1024.0, matcher: str = "exhaustive",
     workdir = Path(tempfile.mkdtemp(prefix="codecsplat_"))
     progress(0.05, desc=f"COLMAP on {len(paths)} images")
     try:
-        sfm = run_sfm_subprocess(paths, workdir, max_image_size=1024, matcher=matcher,
+        sfm = run_sfm(paths, workdir, max_image_size=1024, matcher=matcher,
                       log=lambda m: print(m, flush=True))
     except (ValueError, RuntimeError) as exc:
         raise gr.Error(str(exc)) from exc
     t_sfm = time.time() - t_start
     img_paths, c2w, K = _scene_arrays(sfm["frames"])
+    _nvidia_fds("before GPU call")
 
     progress(0.55, desc="CodecSplat: encode -> bitstream -> decode -> render")
     r = codecsplat_gpu(img_paths, c2w, K, float(lmb))
@@ -273,5 +296,7 @@ with gr.Blocks(title="CodecSplat") as demo:
             label="Frames from the DL3DV demo scene (via the DepthSplat authors)",
         )
 
+_nvidia_fds("after UI build")
+
 if __name__ == "__main__":
     demo.queue(max_size=12).launch(theme=gr.themes.Citrus(), css=CSS, mcp_server=True, show_error=True)