fix(ci): disable torch.compile in orientation training to prevent segfault

torch.compile=on combined with DataLoader spawn workers during LER
validation causes a segfault (20 leaked semaphores, core dumped).
Set PREDECODER_TORCH_COMPILE=0 for the Train all orientations step.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: ivanbasov

code/evaluation/logical_error_rate.py

@@ -1046,6 +1046,10 @@ def run_inference_and_decode_pre_decoder_memory(model, device, dist, cfg) -> dic
                         pass
     except Exception:
         pass
+    # torch.compile + DataLoader spawn workers causes a segfault; fall back to
+    # single-process loading when compile was applied.
+    if _applied_compile and int(test_loader_kwargs.get("num_workers", 0)) > 0:
+        test_loader_kwargs["num_workers"] = 0
     # Handle prefetch_factor when num_workers=0
     if test_loader_kwargs.get('num_workers', 0) == 0:
         test_loader_kwargs.pop('prefetch_factor', None)
@@ -1653,6 +1657,10 @@ def compute_syndrome_density_reduction(model, device, dist, cfg) -> dict:
                         pass
     except Exception:
         pass
+    # torch.compile + DataLoader spawn workers causes a segfault; fall back to
+    # single-process loading when compile will be applied.
+    if _get_env_bool("PREDECODER_TORCH_COMPILE", True) and int(test_loader_kwargs.get("num_workers", 0)) > 0:
+        test_loader_kwargs["num_workers"] = 0
     if int(test_loader_kwargs.get("num_workers", 0)) == 0:
         test_loader_kwargs.pop("prefetch_factor", None)
         if test_loader_kwargs.get("persistent_workers", False):

code/tests/test_v3_optimizations.py

@@ -211,5 +211,62 @@ def test_exact_threshold_qualifies(self):
         )
 
 
+# ---------------------------------------------------------------------------
+# torch.compile + DataLoader spawn-worker guard
+# ---------------------------------------------------------------------------
+class TestCompileSpawnGuard(unittest.TestCase):
+    """Guard that forces num_workers=0 when torch.compile is active.
+
+    torch.compile + DataLoader multiprocessing_context=spawn segfaults (leaked
+    semaphores, core dump). Both LER inference and SDR functions apply this
+    guard before constructing the DataLoader.
+    """
+
+    def _apply_ler_guard(self, applied_compile: bool, num_workers: int) -> dict:
+        """Mirror the guard in run_inference_and_decode_pre_decoder_memory."""
+        kwargs = {"num_workers": num_workers, "prefetch_factor": 2}
+        if applied_compile and int(kwargs.get("num_workers", 0)) > 0:
+            kwargs["num_workers"] = 0
+        if kwargs.get("num_workers", 0) == 0:
+            kwargs.pop("prefetch_factor", None)
+        return kwargs
+
+    def _apply_sdr_guard(self, compile_env: str, num_workers: int) -> dict:
+        """Mirror the guard in compute_syndrome_density_reduction."""
+        kwargs = {"num_workers": num_workers}
+        with patch.dict(os.environ, {"PREDECODER_TORCH_COMPILE": compile_env}):
+            if _get_env_bool("PREDECODER_TORCH_COMPILE", True) and int(kwargs.get("num_workers", 0)) > 0:
+                kwargs["num_workers"] = 0
+        return kwargs
+
+    # --- LER inference guard ---
+
+    def test_ler_guard_forces_zero_workers_when_compiled(self):
+        kwargs = self._apply_ler_guard(applied_compile=True, num_workers=16)
+        self.assertEqual(kwargs["num_workers"], 0)
+
+    def test_ler_guard_keeps_workers_when_not_compiled(self):
+        kwargs = self._apply_ler_guard(applied_compile=False, num_workers=16)
+        self.assertEqual(kwargs["num_workers"], 16)
+
+    def test_ler_guard_noop_when_workers_already_zero(self):
+        kwargs = self._apply_ler_guard(applied_compile=True, num_workers=0)
+        self.assertEqual(kwargs["num_workers"], 0)
+
+    def test_ler_guard_drops_prefetch_factor_when_workers_zeroed(self):
+        kwargs = self._apply_ler_guard(applied_compile=True, num_workers=16)
+        self.assertNotIn("prefetch_factor", kwargs)
+
+    # --- SDR guard (env-var based) ---
+
+    def test_sdr_guard_forces_zero_workers_when_compile_env_on(self):
+        kwargs = self._apply_sdr_guard(compile_env="1", num_workers=16)
+        self.assertEqual(kwargs["num_workers"], 0)
+
+    def test_sdr_guard_keeps_workers_when_compile_env_off(self):
+        kwargs = self._apply_sdr_guard(compile_env="0", num_workers=16)
+        self.assertEqual(kwargs["num_workers"], 16)
+
+
 if __name__ == "__main__":
     unittest.main()