Staged bisect

app.py

@@ -27,9 +27,18 @@ from pathlib import Path
 APP_DIR = Path(__file__).parent.resolve()
 sys.path.insert(0, str(APP_DIR))
 
-if os.environ.get("BISECT") == "minimal":  # diagnostics: bare spaces+torch+gradio
-    exec(open(APP_DIR / "probe_app.py").read())
-    raise SystemExit(0)
+_BISECT = os.environ.get("BISECT", "")
+
+
+def _bisect(stage: str) -> None:
+    """Diagnostics: boot only the GPU probe after `stage` (set via the BISECT variable)."""
+    if _BISECT == stage:
+        print(f"[bisect] stopping after stage {stage}", flush=True)
+        exec(open(APP_DIR / "probe_app.py").read(), {"__name__": "probe"})
+        raise SystemExit(0)
+
+
+_bisect("minimal")
 
 os.environ.setdefault("TORCH_CUDA_ARCH_LIST", "12.0")
 os.environ.setdefault("MAX_JOBS", str(min(32, (os.cpu_count() or 8))))
@@ -67,6 +76,7 @@ def _nvidia_fds(stage: str) -> None:
 
 
 _nvidia_fds("after torch import")
+_bisect("s1_torch")
 
 
 # ---------------------------------------------------------------------------
@@ -96,6 +106,7 @@ 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", "")
+_bisect("s2_env")
 
 import torch.utils.cpp_extension as _cppext  # noqa: E402
 
@@ -109,6 +120,7 @@ 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")
+_bisect("s3_gsplat")
 
 # ---------------------------------------------------------------------------
 # 2. Compatibility shims for the vendored reference code
@@ -210,15 +222,18 @@ def build_model(checkpoint: str):
 for _mod in ("compressai", "pytorch_lightning", "wandb", "lpips", "pycolmap"):
     __import__(_mod)
     _nvidia_fds(f"after import {_mod}")
+_bisect("s4_imports")
 print("[setup] downloading checkpoint...", flush=True)
 CHECKPOINT_PATH = hf_hub_download(MODEL_REPO, CHECKPOINT_FILE)
 print("[setup] building CodecSplat...", flush=True)
 ENCODER, DECODER, DATA_SHIM = build_model(CHECKPOINT_PATH)
 _nvidia_fds("after build (cpu)")
+_bisect("s5_build")
 ENCODER.to("cuda")
 DECODER.to("cuda")
 print("[setup] model ready", flush=True)
 _nvidia_fds("after model build")
+_bisect("s6_cuda")
 
 from src.dataset.shims.crop_shim import rescale_and_crop  # noqa: E402
 from src.evaluation.metrics import compute_psnr, compute_ssim  # noqa: E402