wait for thread tasks only when every module has completed. (#1852)

Qubitium · Qubitium · commit b03d7fe71010 · 2025-09-23T23:22:39.000Z
Signed-off-by: Qubitium &lt;Qubitium@modelcloud.ai&gt;
diff --git a/gptqmodel/looper/module_looper.py b/gptqmodel/looper/module_looper.py
@@ -561,14 +561,12 @@ def finalize_module(module):
 
                     #del module
 
-
                 if auto_gc:
                     torch_empty_cache()
 
-                # LifeCycle: All sub-modules have finalized meaning quantization work is complete
-                # wait for all thread tasks
-                ASYNC_WORKER.join()
-
+        # LifeCycle: All sub-modules have finalized meaning quantization work is complete
+        # wait for all thread tasks
+        ASYNC_WORKER.join()
 
         # paranoid safety check
         torch_sync()
diff --git a/gptqmodel/utils/__init__.py b/gptqmodel/utils/__init__.py
@@ -11,7 +11,7 @@
 
 log = setup_logger()
 
-ASYNC_WORKER = AsyncManager(threads=1)
+ASYNC_WORKER = AsyncManager(threads=4)
 
 # TODO: datasets is not compatible with free threading
 if has_gil_disabled():
