…H-149591) (#149653)

gh-149083: use sentinel to fix _functools.reduce() signature (GH-149591)
(cherry picked from commit c6fd7de)

Co-authored-by: Sergey B Kirpichev <skirpichev@gmail.com>

….15 (#185682)

## Summary

Python 3.15 will expose an introspectable signature for `functools.reduce` using a PEP 661 sentinel default: `(function, iterable, /, initial=functools._initial_missing)`. The CPython change is [#149591](python/cpython#149591), which merged into CPython main on 2026-05-10, after 3.15.0b1 was tagged. So on 3.15.0b1 the default is still `<unrepresentable>` and `inspect.signature()` raises `ValueError`, which `substitute_in_graph` swallows and skips the check. From 3.15.0b2 the check will run.

When it runs, it compares positional parameter names, keyword-only names, and default values. The polyfill declares its own local sentinel:

```python
_initial_missing = object()

@substitute_in_graph(functools.reduce)
def reduce(function, iterable, initial=_initial_missing, /):
    ...
```

That local `object()` is a different instance from `functools._initial_missing`, so the defaults dict compares unequal and `substitute_in_graph` raises:

```
File ".../torch/_dynamo/polyfills/functools.py", line N, in <module>
  @substitute_in_graph(functools.reduce)
TypeError: Signature mismatch between <built-in function reduce> and <function reduce at 0x...>:
  (function, iterable, /, initial=_initial_missing)
  != (function, iterable, initial=<object object at 0x...>, /)
```

The polyfills loader imports every polyfill module, so this will break `import torch._dynamo` entirely on Python 3.15.0b2+. Same failure mode as #185403's `struct.pack` fix, just one step ahead of the next beta.

The fix imports `functools._initial_missing` instead of declaring a local one. The polyfill's internal `if initial is _initial_missing` identity check still works because both sides now reference the same object. No behavior change on any Python version. `functools._initial_missing` has existed in `functools` since well before 3.10 (it's used by the pure-Python `functools.reduce` fallback), so the import is safe on every supported version.

Part of #184352 (Python 3.15 support). Similar to #185403.

## Test plan

Verified on a locally-built Python 3.15-dev (CPython main at `heads/3.15:863c7e0`, which includes [#149591](python/cpython#149591)). The upstream signature is exposed natively:

```
$ python3.15 -c "import functools, inspect; print(inspect.signature(functools.reduce))"
(function, iterable, /, initial=_initial_missing)
$ python3.15 -c "import functools, inspect; sig = inspect.signature(functools.reduce); print(sig.parameters['initial'].default is functools._initial_missing)"
True
```

Extracted the actual `substitute_in_graph` from `torch/_dynamo/decorators.py` and ran it against both polyfill versions with no monkey-patching.

Unmodified polyfill (`_initial_missing = object()` declared locally):

```
TypeError: Signature mismatch between <built-in function reduce> and <function unmodified_reduce at 0x...>:
  (function, iterable, /, initial=_initial_missing)
  != (function, iterable, initial=<object object at 0x...>, /)
```

Fixed polyfill (`from functools import _initial_missing`): no `TypeError` at the signature check.

Regression on Python 3.12.13 (oldest currently supported):

```
$ python3 -c "import inspect, functools; inspect.signature(functools.reduce)"
ValueError: no signature found for builtin <built-in function reduce>
```

The signature check is silently skipped on Python <= 3.14, so the fix is a no-op there. `functools._initial_missing` is confirmed present on 3.12.13. Eager-vs-polyfill parity verified on every input shape:

```
reduce(lambda a, b: a+b, [1,2,3,4])       == 10     (matches functools.reduce)
reduce(lambda a, b: a+b, [1,2,3], 100)    == 106    (matches)
reduce(lambda a, b: a+b, [])              -> TypeError("reduce() of empty iterable with no initial value")  (matches)
reduce(lambda a, b: a+b, [], 99)          == 99     (matches)
```

Pre-commit checks: `python3 -m py_compile torch/_dynamo/polyfills/functools.py` and `git diff --staged --check` both clean.

Pull Request resolved: #185682
Approved by: https://github.com/ezyang

Co-authored-by: Edward Z. Yang via mergedog <ezyang@meta.com>

….15 (pytorch#185682)

## Summary

Python 3.15 will expose an introspectable signature for `functools.reduce` using a PEP 661 sentinel default: `(function, iterable, /, initial=functools._initial_missing)`. The CPython change is [pytorch#149591](python/cpython#149591), which merged into CPython main on 2026-05-10, after 3.15.0b1 was tagged. So on 3.15.0b1 the default is still `<unrepresentable>` and `inspect.signature()` raises `ValueError`, which `substitute_in_graph` swallows and skips the check. From 3.15.0b2 the check will run.

When it runs, it compares positional parameter names, keyword-only names, and default values. The polyfill declares its own local sentinel:

```python
_initial_missing = object()

@substitute_in_graph(functools.reduce)
def reduce(function, iterable, initial=_initial_missing, /):
    ...
```

That local `object()` is a different instance from `functools._initial_missing`, so the defaults dict compares unequal and `substitute_in_graph` raises:

```
File ".../torch/_dynamo/polyfills/functools.py", line N, in <module>
  @substitute_in_graph(functools.reduce)
TypeError: Signature mismatch between <built-in function reduce> and <function reduce at 0x...>:
  (function, iterable, /, initial=_initial_missing)
  != (function, iterable, initial=<object object at 0x...>, /)
```

The polyfills loader imports every polyfill module, so this will break `import torch._dynamo` entirely on Python 3.15.0b2+. Same failure mode as pytorch#185403's `struct.pack` fix, just one step ahead of the next beta.

The fix imports `functools._initial_missing` instead of declaring a local one. The polyfill's internal `if initial is _initial_missing` identity check still works because both sides now reference the same object. No behavior change on any Python version. `functools._initial_missing` has existed in `functools` since well before 3.10 (it's used by the pure-Python `functools.reduce` fallback), so the import is safe on every supported version.

Part of pytorch#184352 (Python 3.15 support). Similar to pytorch#185403.

## Test plan

Verified on a locally-built Python 3.15-dev (CPython main at `heads/3.15:863c7e0`, which includes [pytorch#149591](python/cpython#149591)). The upstream signature is exposed natively:

```
$ python3.15 -c "import functools, inspect; print(inspect.signature(functools.reduce))"
(function, iterable, /, initial=_initial_missing)
$ python3.15 -c "import functools, inspect; sig = inspect.signature(functools.reduce); print(sig.parameters['initial'].default is functools._initial_missing)"
True
```

Extracted the actual `substitute_in_graph` from `torch/_dynamo/decorators.py` and ran it against both polyfill versions with no monkey-patching.

Unmodified polyfill (`_initial_missing = object()` declared locally):

```
TypeError: Signature mismatch between <built-in function reduce> and <function unmodified_reduce at 0x...>:
  (function, iterable, /, initial=_initial_missing)
  != (function, iterable, initial=<object object at 0x...>, /)
```

Fixed polyfill (`from functools import _initial_missing`): no `TypeError` at the signature check.

Regression on Python 3.12.13 (oldest currently supported):

```
$ python3 -c "import inspect, functools; inspect.signature(functools.reduce)"
ValueError: no signature found for builtin <built-in function reduce>
```

The signature check is silently skipped on Python <= 3.14, so the fix is a no-op there. `functools._initial_missing` is confirmed present on 3.12.13. Eager-vs-polyfill parity verified on every input shape:

```
reduce(lambda a, b: a+b, [1,2,3,4])       == 10     (matches functools.reduce)
reduce(lambda a, b: a+b, [1,2,3], 100)    == 106    (matches)
reduce(lambda a, b: a+b, [])              -> TypeError("reduce() of empty iterable with no initial value")  (matches)
reduce(lambda a, b: a+b, [], 99)          == 99     (matches)
```

Pre-commit checks: `python3 -m py_compile torch/_dynamo/polyfills/functools.py` and `git diff --staged --check` both clean.

Pull Request resolved: pytorch#185682
Approved by: https://github.com/ezyang

Co-authored-by: Edward Z. Yang via mergedog <ezyang@meta.com>

….15 (pytorch#185682)

## Summary

Python 3.15 will expose an introspectable signature for `functools.reduce` using a PEP 661 sentinel default: `(function, iterable, /, initial=functools._initial_missing)`. The CPython change is [pytorch#149591](python/cpython#149591), which merged into CPython main on 2026-05-10, after 3.15.0b1 was tagged. So on 3.15.0b1 the default is still `<unrepresentable>` and `inspect.signature()` raises `ValueError`, which `substitute_in_graph` swallows and skips the check. From 3.15.0b2 the check will run.

When it runs, it compares positional parameter names, keyword-only names, and default values. The polyfill declares its own local sentinel:

```python
_initial_missing = object()

@substitute_in_graph(functools.reduce)
def reduce(function, iterable, initial=_initial_missing, /):
    ...
```

That local `object()` is a different instance from `functools._initial_missing`, so the defaults dict compares unequal and `substitute_in_graph` raises:

```
File ".../torch/_dynamo/polyfills/functools.py", line N, in <module>
  @substitute_in_graph(functools.reduce)
TypeError: Signature mismatch between <built-in function reduce> and <function reduce at 0x...>:
  (function, iterable, /, initial=_initial_missing)
  != (function, iterable, initial=<object object at 0x...>, /)
```

The polyfills loader imports every polyfill module, so this will break `import torch._dynamo` entirely on Python 3.15.0b2+. Same failure mode as pytorch#185403's `struct.pack` fix, just one step ahead of the next beta.

The fix imports `functools._initial_missing` instead of declaring a local one. The polyfill's internal `if initial is _initial_missing` identity check still works because both sides now reference the same object. No behavior change on any Python version. `functools._initial_missing` has existed in `functools` since well before 3.10 (it's used by the pure-Python `functools.reduce` fallback), so the import is safe on every supported version.

Part of pytorch#184352 (Python 3.15 support). Similar to pytorch#185403.

## Test plan

Verified on a locally-built Python 3.15-dev (CPython main at `heads/3.15:863c7e0`, which includes [pytorch#149591](python/cpython#149591)). The upstream signature is exposed natively:

```
$ python3.15 -c "import functools, inspect; print(inspect.signature(functools.reduce))"
(function, iterable, /, initial=_initial_missing)
$ python3.15 -c "import functools, inspect; sig = inspect.signature(functools.reduce); print(sig.parameters['initial'].default is functools._initial_missing)"
True
```

Extracted the actual `substitute_in_graph` from `torch/_dynamo/decorators.py` and ran it against both polyfill versions with no monkey-patching.

Unmodified polyfill (`_initial_missing = object()` declared locally):

```
TypeError: Signature mismatch between <built-in function reduce> and <function unmodified_reduce at 0x...>:
  (function, iterable, /, initial=_initial_missing)
  != (function, iterable, initial=<object object at 0x...>, /)
```

Fixed polyfill (`from functools import _initial_missing`): no `TypeError` at the signature check.

Regression on Python 3.12.13 (oldest currently supported):

```
$ python3 -c "import inspect, functools; inspect.signature(functools.reduce)"
ValueError: no signature found for builtin <built-in function reduce>
```

The signature check is silently skipped on Python <= 3.14, so the fix is a no-op there. `functools._initial_missing` is confirmed present on 3.12.13. Eager-vs-polyfill parity verified on every input shape:

```
reduce(lambda a, b: a+b, [1,2,3,4])       == 10     (matches functools.reduce)
reduce(lambda a, b: a+b, [1,2,3], 100)    == 106    (matches)
reduce(lambda a, b: a+b, [])              -> TypeError("reduce() of empty iterable with no initial value")  (matches)
reduce(lambda a, b: a+b, [], 99)          == 99     (matches)
```

Pre-commit checks: `python3 -m py_compile torch/_dynamo/polyfills/functools.py` and `git diff --staged --check` both clean.

Pull Request resolved: pytorch#185682
Approved by: https://github.com/ezyang

Co-authored-by: Edward Z. Yang via mergedog <ezyang@meta.com>

….15 (pytorch#185682)

## Summary

Python 3.15 will expose an introspectable signature for `functools.reduce` using a PEP 661 sentinel default: `(function, iterable, /, initial=functools._initial_missing)`. The CPython change is [pytorch#149591](python/cpython#149591), which merged into CPython main on 2026-05-10, after 3.15.0b1 was tagged. So on 3.15.0b1 the default is still `<unrepresentable>` and `inspect.signature()` raises `ValueError`, which `substitute_in_graph` swallows and skips the check. From 3.15.0b2 the check will run.

When it runs, it compares positional parameter names, keyword-only names, and default values. The polyfill declares its own local sentinel:

```python
_initial_missing = object()

@substitute_in_graph(functools.reduce)
def reduce(function, iterable, initial=_initial_missing, /):
    ...
```

That local `object()` is a different instance from `functools._initial_missing`, so the defaults dict compares unequal and `substitute_in_graph` raises:

```
File ".../torch/_dynamo/polyfills/functools.py", line N, in <module>
  @substitute_in_graph(functools.reduce)
TypeError: Signature mismatch between <built-in function reduce> and <function reduce at 0x...>:
  (function, iterable, /, initial=_initial_missing)
  != (function, iterable, initial=<object object at 0x...>, /)
```

The polyfills loader imports every polyfill module, so this will break `import torch._dynamo` entirely on Python 3.15.0b2+. Same failure mode as pytorch#185403's `struct.pack` fix, just one step ahead of the next beta.

The fix imports `functools._initial_missing` instead of declaring a local one. The polyfill's internal `if initial is _initial_missing` identity check still works because both sides now reference the same object. No behavior change on any Python version. `functools._initial_missing` has existed in `functools` since well before 3.10 (it's used by the pure-Python `functools.reduce` fallback), so the import is safe on every supported version.

Part of pytorch#184352 (Python 3.15 support). Similar to pytorch#185403.

## Test plan

Verified on a locally-built Python 3.15-dev (CPython main at `heads/3.15:863c7e0`, which includes [pytorch#149591](python/cpython#149591)). The upstream signature is exposed natively:

```
$ python3.15 -c "import functools, inspect; print(inspect.signature(functools.reduce))"
(function, iterable, /, initial=_initial_missing)
$ python3.15 -c "import functools, inspect; sig = inspect.signature(functools.reduce); print(sig.parameters['initial'].default is functools._initial_missing)"
True
```

Extracted the actual `substitute_in_graph` from `torch/_dynamo/decorators.py` and ran it against both polyfill versions with no monkey-patching.

Unmodified polyfill (`_initial_missing = object()` declared locally):

```
TypeError: Signature mismatch between <built-in function reduce> and <function unmodified_reduce at 0x...>:
  (function, iterable, /, initial=_initial_missing)
  != (function, iterable, initial=<object object at 0x...>, /)
```

Fixed polyfill (`from functools import _initial_missing`): no `TypeError` at the signature check.

Regression on Python 3.12.13 (oldest currently supported):

```
$ python3 -c "import inspect, functools; inspect.signature(functools.reduce)"
ValueError: no signature found for builtin <built-in function reduce>
```

The signature check is silently skipped on Python <= 3.14, so the fix is a no-op there. `functools._initial_missing` is confirmed present on 3.12.13. Eager-vs-polyfill parity verified on every input shape:

```
reduce(lambda a, b: a+b, [1,2,3,4])       == 10     (matches functools.reduce)
reduce(lambda a, b: a+b, [1,2,3], 100)    == 106    (matches)
reduce(lambda a, b: a+b, [])              -> TypeError("reduce() of empty iterable with no initial value")  (matches)
reduce(lambda a, b: a+b, [], 99)          == 99     (matches)
```

Pre-commit checks: `python3 -m py_compile torch/_dynamo/polyfills/functools.py` and `git diff --staged --check` both clean.

Pull Request resolved: pytorch#185682
Approved by: https://github.com/ezyang

Co-authored-by: Edward Z. Yang via mergedog <ezyang@meta.com>