Revert "fix Invalid call to aoti_torch_tensor_copy_ #123039 (#124037)"
This reverts commit f9379ebbbf1369aad8179cac4a2eb7d72f25739e.
Reverted https://github.com/pytorch/pytorch/pull/124037 on behalf of https://github.com/jeanschmidt due to introducing regressions in benchmark, see D56623194 for more details ([comment](https://github.com/pytorch/pytorch/pull/124037#issuecomment-2079574308))
diff --git a/test/inductor/test_cuda_cpp_wrapper.py b/test/inductor/test_cuda_cpp_wrapper.py
index fa717ab..b662e24 100644
--- a/test/inductor/test_cuda_cpp_wrapper.py
+++ b/test/inductor/test_cuda_cpp_wrapper.py
@@ -109,7 +109,9 @@
test_failures_cuda_wrapper[
f"{test_name}_dynamic_shapes"
] = test_torchinductor.TestFailure(("cuda_wrapper",), is_skip=False)
- skip_list = []
+ skip_list = [
+ "test_multi_device_cuda",
+ ]
for test_name in skip_list:
test_failures_cuda_wrapper[test_name] = test_torchinductor.TestFailure(
("cuda_wrapper",), is_skip=True
diff --git a/torch/_inductor/codegen/cpp_wrapper_cpu.py b/torch/_inductor/codegen/cpp_wrapper_cpu.py
index f50e258..95e4ef3 100644
--- a/torch/_inductor/codegen/cpp_wrapper_cpu.py
+++ b/torch/_inductor/codegen/cpp_wrapper_cpu.py
@@ -896,11 +896,9 @@
@cache_on_self
def get_output_refs(self):
return [
- (
- f"torch::tensor({x.codegen_reference(self.wrapper_call)})"
- if isinstance(x, ir.ShapeAsConstantBuffer) and not config.abi_compatible
- else x.codegen_reference(self.wrapper_call)
- )
+ f"torch::tensor({x.codegen_reference(self.wrapper_call)})"
+ if isinstance(x, ir.ShapeAsConstantBuffer) and not config.abi_compatible
+ else x.codegen_reference(self.wrapper_call)
for x in V.graph.graph_outputs
]
@@ -1100,11 +1098,9 @@
outputs_str = "output_tensors"
else:
outputs = [
- (
- f"output_tensors[{i}]"
- if self.output_is_tensor[i]
- else f"output_tensors[{i}].item()"
- )
+ f"output_tensors[{i}]"
+ if self.output_is_tensor[i]
+ else f"output_tensors[{i}].item()"
for i in range(len(V.graph.graph_outputs))
]
outputs_str = f"[{', '.join(outputs)}]"
@@ -1430,7 +1426,6 @@
and ir.is_contiguous_strides_for_shape(
buffer.get_stride(), buffer.get_size()
)
- and not buffer.is_extern()
)
def make_buffer_free(self, buffer):
