Defer native CUDA imports in ZeroGPU worker probe

app.py

@@ -28,7 +28,7 @@ OUTPUT_ROOT = Path(os.environ.get("CANIS_OUTPUT_ROOT", "/tmp/canis_outputs"))
 POOL_ROOT = Path(os.environ.get("CANIS_POOL_ROOT", "/tmp/canis_empty_pool"))
 PUBLIC_TRELLIS_REPO = "Louischong/Trellis-OA"
 PIPELINE_CONFIG_ROOT = Path("/tmp/canis_pipeline_config")
-DIAGNOSTIC_MODE = "zero_pack_native_probe"
+DIAGNOSTIC_MODE = "zero_pack_deferred_native_probe"
 
 @spaces.GPU(duration=360)
 def _canonicalize_on_gpu(
@@ -42,7 +42,7 @@ def _canonicalize_on_gpu(
     started = time.perf_counter()
     _trace(f"phase=gpu_callback event=START mode={DIAGNOSTIC_MODE}")
     _trace_gpu_environment()
-    if DIAGNOSTIC_MODE == "zero_pack_native_probe":
+    if DIAGNOSTIC_MODE in {"zero_pack_native_probe", "zero_pack_deferred_native_probe"}:
         return _run_zero_pack_native_probe(started)
     try:
         from canis_job_entry import run_job
@@ -545,8 +545,9 @@ def main() -> None:
     _verify_private_runtime()
     _install_python_compatibility()
     _install_trusted_checkpoint_compatibility()
-    _install_pytorch3d_compatibility()
-    _install_blackwell_attention_compatibility()
+    if DIAGNOSTIC_MODE != "zero_pack_deferred_native_probe":
+        _install_pytorch3d_compatibility()
+        _install_blackwell_attention_compatibility()
     _install_fpfh_only_compatibility()
     private_checkpoint, fcgf_checkpoint = _validate_models()
     checkpoint = _prepare_cached_pipeline_config(private_checkpoint)