test(pt_expt): reduce AOTInductor compile time with fast configs

Set inductor configs in conftest to skip expensive C++ optimizations
during .pt2 compilation: max_fusion_size=8, epilogue_fusion=False,
pattern_matcher=False, package_cpp_only=True, compile_opt_level=O0.
Tests only validate correctness so runtime performance is irrelevant.
Cuts per-model compile time from ~50s to ~30s.

Author: Han Wang

source/tests/pt_expt/conftest.py

@@ -12,11 +12,21 @@
 """
 
 import pytest
+import torch._inductor.config as _inductor_config
 import torch.utils._device as _device
 from torch.overrides import (
     _get_current_function_mode_stack,
 )
 
+# Reduce AOTInductor (.pt2) compile time for unit tests.
+# Tests only validate correctness, not runtime performance, so we can
+# skip expensive C++ optimizations.  This cuts compile time by ~50%.
+_inductor_config.max_fusion_size = 8
+_inductor_config.epilogue_fusion = False
+_inductor_config.pattern_matcher = False
+_inductor_config.aot_inductor.package_cpp_only = True
+_inductor_config.aot_inductor.compile_wrapper_opt_level = "O0"
+
 
 def _pop_device_contexts() -> list:
     """Pop all stale DeviceContext modes from the torch function mode stack."""