Flip triton kernel default layout constraint to "needs_fixed_stride_order"

zou3519 · zou3519 · commit 276f87219887 · 2024-09-10T07:59:56.000-07:00
This is to match the default layout constraint for custom operators. By
default, Inductor should match the stride order of inputs to a triton
kernel.

Test Plan:
- existing tests

[ghstack-poisoned]
diff --git a/torch/_inductor/config.py b/torch/_inductor/config.py
@@ -74,7 +74,7 @@ def autotune_remote_cache_default() -> Optional[bool]:
 
 # The default layout constraint for user-defined triton kernels.
 # See "The default layout constraint for custom operators" for options.
-triton_kernel_default_layout_constraint = "flexible_layout"
+triton_kernel_default_layout_constraint = "needs_fixed_stride_order"
 
 # use cpp wrapper instead of python wrapper
 cpp_wrapper = os.environ.get("TORCHINDUCTOR_CPP_WRAPPER", "0") == "1"
