[BE] Extend test_remove_no_ops

test/inductor/test_torchinductor.py

@@ -5991,7 +5991,6 @@ def fn(x1, x2, x3, x4):
         )
 
     @skip_if_gpu_halide
-    @skipCUDAIf(not SM80OrLater, "uses bfloat16 which requires SM >= 80")
     # Constant folding was explicitly turned off due to issue #108388
     # Turn it back on for test
     @torch._inductor.config.patch(joint_graph_constant_folding=True)
@@ -6033,21 +6032,26 @@ def matmul_with_op(x, y, fn):
                 FileCheck().check_not("triton.jit").run(source_codes[0])
 
         # test dtype conversion
-        inps = [
-            torch.rand([256, 256], device=self.device, dtype=torch.bfloat16)
-            for _ in range(2)
-        ]
-        for fn in fns:
+        device_interface = get_interface_for_device(self.device)
+
+        for lowp_dtype in [torch.float16, torch.bfloat16]:
+            if not device_interface.is_dtype_supported(lowp_dtype):
+                continue
+            inps = [
+                torch.rand([256, 256], device=self.device, dtype=lowp_dtype)
+                for _ in range(2)
+            ]
+            for fn in fns:
+                out, source_codes = run_and_get_code(foo_opt, inps[0], inps[1], fn)
+                self.assertEqual(out, matmul_with_op(inps[0], inps[1], fn))
+
+            # test broadcasted shape bail
+            fn = lambda x: x + torch.zeros(  # noqa: E731
+                [256, 256, 256], dtype=lowp_dtype, device=self.device
+            )
             out, source_codes = run_and_get_code(foo_opt, inps[0], inps[1], fn)
             self.assertEqual(out, matmul_with_op(inps[0], inps[1], fn))
 
-        # test broadcasted shape bail
-        fn = lambda x: x + torch.zeros(  # noqa: E731
-            [256, 256, 256], dtype=torch.bfloat16, device=self.device
-        )
-        out, source_codes = run_and_get_code(foo_opt, inps[0], inps[1], fn)
-        self.assertEqual(out, matmul_with_op(inps[0], inps[1], fn))
-
     def test_remove_noop_copy(self):
         def fn(x, y):
             x = x.cos()

torch/_dynamo/device_interface.py

@@ -319,6 +319,10 @@ def record(self, stream=None):
     def is_available() -> bool:
         return True
 
+    @staticmethod
+    def is_bf16_supported(including_emulation: bool = False):
+        return True
+
     @staticmethod
     def get_compute_capability(device: _device_t = None) -> str:
         return ""