Traceback (most recent call last): File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/common_utils.py", line 3370, in wrapper method(*args, **kwargs) ~~~~~~^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/convert_frame.py", line 2340, in call result = self._torchdynamo_orig_backend( frame, cache_entry, self.hooks, frame_state, skip=1 ) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/convert_frame.py", line 2076, in call result = self._inner_convert( frame, cache_entry, hooks, frame_state, skip=skip + 1 ) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/convert_frame.py", line 727, in call result = _compile( frame.f_code, ...<16 lines>... convert_frame_box=self._box, ) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/convert_frame.py", line 1851, in _compile guarded_code, tracer_output = compile_inner(code, one_graph, hooks) ~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_utils_internal.py", line 96, in wrapper_function return function(*args, **kwargs) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/convert_frame.py", line 1495, in compile_inner return _compile_inner(code, one_graph, hooks) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/convert_frame.py", line 1540, in _compile_inner dynamo_output = compile_frame( code, ...<11 lines>... package=package, ) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/convert_frame.py", line 1403, in compile_frame bytecode, tracer_output = transform_code_object(code, transform) ~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/bytecode_transformation.py", line 1626, in transform_code_object tracer_output = transformations(instructions, code_options) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/convert_frame.py", line 1375, in transform tracer_output = trace_frame( code, ...<14 lines>... package=package, ) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/convert_frame.py", line 368, in _fn assert guards.check(), ( ~~~~~~~~~~~~^^ AssertionError: Global torch_function state changed while dynamo tracing, please report a bug

To execute this test, run the following from the base repo dir: PYTORCH_TEST_WITH_DYNAMO=1 python test/test_overrides.py TestTorchFunctionMode.test_disable_enable_torch_function_ctx

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0

Fix Plan

1. Update test_disable_enable_torch_function_ctx to use torch.dynamo.freeze instead of torch.dynamo.convert

The issue is caused by the test modifying the global torch_function state while dynamo tracing is active. We can fix this by using torch.dynamo.freeze to freeze the tracing state before modifying the global state.

import torch
from torch.dynamo import freeze

def test_disable_enable_torch_function_ctx():
    # Freeze the tracing state
    with freeze():
        # Modify the global torch_function state
        torch._C._set_torch_function_state('new_state')

2. Update the test to use PYTORCH_TEST_WITH_DYNAMO=0 to disable dynamo tracing

We can also update the test to use PYTORCH_TEST_WITH_DYNAMO=0 to disable dynamo tracing, which will prevent the test from modifying the global torch_function state.

import os

def test_disable_enable_torch_function_ctx():
    # Disable dynamo tracing
    os.environ['PYTORCH_TEST_WITH_DYNAMO'] = '0'
    # Modify the global torch_function state
    torch._C._set_torch_function_state('new_state')

3. Update the test to use torch.dynamo.reset to reset the tracing state

Alternatively, we can update the test to use torch.dynamo.reset to reset the tracing state before modifying the global torch_function state.

import torch
from torch.dynamo import reset

def test_disable_enable_torch_function_ctx():
    # Reset the tracing state
    reset()
    # Modify the global torch_function state
    torch._C._set_torch_function_state('new_state')

Verification

To verify that the fix worked, you can run the test with the updated code