Replicate

gptqmodel/utils/env.py

@@ -13,13 +13,17 @@
 _TRUTHY = {"1", "true", "yes", "on", "y"}
 
 
-def env_flag(name: str, default: str | None = "0") -> bool:
+def env_flag(name: str, default: str | bool | None = "0") -> bool:
     """Return ``True`` when an env var is set to a truthy value."""
 
-    value = os.getenv(name, default)
+    value = os.getenv(name)
     if value is None:
-        return False
-    return value.strip().lower() in _TRUTHY
+        if default is None:
+            return False
+        if isinstance(default, bool):
+            return default
+        value = default
+    return str(value).strip().lower() in _TRUTHY
 
 
 __all__ = ["env_flag"]

gptqmodel/utils/looper_helpers.py

@@ -24,7 +24,7 @@
 from ..utils.torch import ALL_DEVICES, CPU, torch_sync
 
 
-USE_TORCH_REPLICATE = env_flag("GPTQMODEL_USE_TORCH_REPLICATE")
+USE_TORCH_REPLICATE = env_flag("GPTQMODEL_USE_TORCH_REPLICATE", True)
 
 
 _THREAD_SAFE_PARALLEL = ThreadSafe(torch_parallel)

tests/pytest.ini

@@ -5,3 +5,6 @@ norecursedirs = tasks evalplus_results
 markers =
     ci: CPU-only CI regression coverage for DeviceThreadPool affinity behaviour
     cuda: Requires CUDA device
+    inference: Inference workloads that replicate models across devices
+filterwarnings =
+    ignore:Warning only once for all operators.*:UserWarning