Make CI error on inductor fallback when decomp is available (#99473)

Fixes #99446

Remove the warning, as that annoyed end-users who don't know what to do about it.

Instead, try to hold the line by preventing any decomp from being added without making
the corresponding change to inductor's fallbacks.

Note: we probably still need to better document how to update inductor's decomps,
for now it's pretty much "go ask the inductor team for advice"

Pull Request resolved: #99473
Approved by: https://github.com/ezyang, https://github.com/ngimel, https://github.com/jansel

torch/_decomp/__init__.py

@@ -203,6 +203,7 @@ def core_aten_decompositions() -> Dict[OpOverload, Callable]:
             aten.elu_backward,
             aten._embedding_bag,
             aten.embedding_dense_backward,
+            aten._euclidean_dist.default,
             aten.expand_as,
             aten.eye,
             aten.fill,

torch/_inductor/lowering.py

@@ -22,6 +22,7 @@
     type_to_dtype,
 )
 from torch.fx.experimental.symbolic_shapes import magic_methods, method_to_operator
+from torch.testing._internal.common_utils import IS_CI
 from torch.utils._pytree import tree_flatten
 from .._dynamo.utils import import_submodule
 
@@ -1107,9 +1108,22 @@ def make_fallback(kernel, layout_constraint=None, warn=True):
     assert (
         kernel not in decompositions
     ), f"both a fallback and a decomp for same kernel: {kernel}"
-    if get_decompositions([kernel]) and warn:
-        developer_warning(
-            f"make_fallback({kernel}): a decomposition exists, we should switch to it"
+    if get_decompositions([kernel]) and warn and IS_CI:
+        # Note: 'warn' is holdover from when this was a warning, but for ops that previously
+        # set warn=False we do not want a CI error.
+        # Ignore the 'suppress errors' configs in CI, as this particular warning happens on startup anyway and is not
+        # likely to be triggered preferentially on one CI config over another.
+        if torch._dynamo.config.suppress_errors:
+            torch._dynamo.config.suppress_errors = False
+            log.warning(
+                "A make_fallback error occured in suppress_errors config,"
+                " and suppress_errors is being disabled to surface it."
+            )
+        raise AssertionError(
+            f"make_fallback({kernel}): a decomposition exists, we should switch to it."
+            " To fix this error, either add a decomposition to core_aten_decompositions (preferred)"
+            " or inductor_decompositions, and delete the corresponding `make_fallback` line."
+            " Get help from the inductor team if unsure, don't pick arbitrarily to unblock yourself.",
         )
 
     add_needs_realized_inputs(kernel)
@@ -1481,6 +1495,9 @@ def apply_constraint(arg, fx_arg):
 make_fallback(aten._linalg_eigh)
 make_fallback(aten.zeros.names)
 
+# fails accuracy on test_torch.py, and explicit fallback required to avoid warn=True on implicit
+make_fallback(aten.exponential.default, warn=False)
+
 
 @register_lowering(aten.clone)
 def clone(x, *, memory_format=0):