Update

[ghstack-poisoned]
pytorch · benjaminglass1 · Mar 25, 2025 · Mar 26, 2025 · Apr 9, 2025 · Apr 10, 2025
commit 5bfee45fbc88ffcc0dbe89e31f2bcef600c57c11
diff --git a/test/inductor/test_aot_inductor.py b/test/inductor/test_aot_inductor.py
@@ -4672,11 +4672,7 @@ def forward(self, a, b, c):
         so_path, code = run_and_get_cpp_code(
             AOTIRunnerUtil.legacy_compile, model, example_inputs
         )
-        is_cpu_freezing = (
-            config.freezing is None or config.freezing
-        ) and self.device != GPU_TYPE
-        varname = f"u{int(mark_unbacked) + (2 if is_cpu_freezing else 0)}"
-        lowerbound_check = f"{varname} >= {1 if mark_unbacked else 2}"
+        lowerbound_check = f"u{int(mark_unbacked)} >= {1 if mark_unbacked else 2}"
         FileCheck().check_count(lowerbound_check, 1).run(code)
 
         compiled = AOTIRunnerUtil.legacy_load(self.device, so_path)

diff --git a/torch/_inductor/fx_passes/mkldnn_fusion.py b/torch/_inductor/fx_passes/mkldnn_fusion.py
@@ -1096,6 +1096,10 @@ def is_const_or_cat_by_const(weight):
         if input_meta_value is None or weight_meta_value is None:
             return False
         batch_size = input_meta_value.shape[0]
+        # One of our tests attempts to run linear over a zero-element tensor.  This is
+        # not supported by MKLDNN, so refuse to fuse.
+        if input_meta_value.numel() == 0 or weight_meta_value.numel() == 0:
+            return False
         if (
             input_meta_value.dtype == torch.float64
             or weight_meta_value.dtype == torch.float64