[inductor] fix TritonTemplateCaller.__str__ (#97578)

We remove TritonTemplateCaller.to_callable previously. But this method is still used in `TritonTemplateCaller.__str__` . The to_callable method in the base class will be used and raise an exception.

This PR fix TritonTemplateCaller.__str__ to return the string representation without calling to_callable.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/97578
Approved by: https://github.com/nmacchioni, https://github.com/ngimel

test/inductor/test_select_algorithm.py

@@ -9,6 +9,8 @@ import torch._inductor.select_algorithm as select_algorithm
 import torch.nn.functional as F
 from torch._dynamo.test_case import run_tests, TestCase
 from torch._dynamo.utils import counters
+from torch._inductor.autotune_process import BenchmarkRequest
+
 from torch.testing._internal.common_utils import IS_LINUX
 from torch.testing._internal.inductor_utils import HAS_CUDA
 
@@ -273,6 +275,28 @@ class TestSelectAlgorithm(TestCase):
         # Autotuning checks correctness of each version
         self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
 
+    def test_TritonTemplateCaller_str(self):
+        """
+        Make sure str(TritonTemplateCaller) does not raise exceptions.
+        """
+        module_path = "abc.py"
+        bmreq = BenchmarkRequest(
+            module_path=module_path,
+            module_cache_key=None,
+            kernel_name=None,
+            grid=None,
+            extra_args=None,
+            num_stages=None,
+            num_warps=None,
+            input_tensors=None,
+            output_tensor=None,
+        )
+        caller = select_algorithm.TritonTemplateCaller(
+            None, None, None, None, "extra", bmreq
+        )
+        caller_str = str(caller)
+        self.assertEqual(caller_str, f"TritonTemplateCaller({module_path}, extra)")
+
 
 if __name__ == "__main__":
     from torch._inductor.utils import is_big_gpu

torch/_inductor/select_algorithm.py

@@ -552,9 +552,7 @@ class TritonTemplateCaller(ChoiceCaller):
         return self.bmreq.benchmark(*args, output_tensor=out)
 
     def __str__(self):
-        return (
-            f"TritonTemplateCaller({self.to_callable().__file__}, {self.debug_extra})"
-        )
+        return f"TritonTemplateCaller({self.bmreq.module_path}, {self.debug_extra})"
 
     def call_name(self):
         return f"template_kernels.{self.name}"