pytorch - 💡(How to fix) Fix [Inductor] `pick_deterministic_choice` checks `ExternKernelChoice` instead of `ExternKernelCaller` — deterministic mode doesn't prefer extern kernels [1 participants]

🐛 Describe the bug

🐛 Describe the bug

In torch/_inductor/select_algorithm.py, pick_deterministic_choice() (called when torch._inductor.config.deterministic = True) tries to prefer extern (ATen/cuBLAS) kernels over Triton templates. However, it checks isinstance(choice, ExternKernelChoice) instead of isinstance(choice, ExternKernelCaller).

ExternKernelChoice is a standalone factory class (inherits from object), while the actual elements in the choices list are ExternKernelCaller instances (inherits from ChoiceCaller). The isinstance check is always False, so the function always falls through to return choices[0].

The same issue exists in _classify_kernel_operation() at line 3650.

Code

# select_algorithm.py, line 3771-3779
def pick_deterministic_choice(self, choices: list[ChoiceCaller]) -> ChoiceCaller:
    assert len(choices) >= 2
    externs = [
        choice for choice in choices if isinstance(choice, ExternKernelChoice)
        #                                          ^^^^^^^^^^^^^^^^^^^^^^^^^^
        #                                          should be ExternKernelCaller
    ]
    if len(externs) > 0:
        return externs[0]
    else:
        return choices[0]  # ← always reaches here

# select_algorithm.py, line 3650
elif isinstance(choice, ExternKernelChoice):  # same bug
    choice_name = choice.name

Class hierarchy:

object
└── ExternKernelChoice       ← factory; creates ExternKernelCaller instances (line 3024)

ChoiceCaller                 ← base for elements in `choices` list
├── ExternKernelCaller       ← wraps an ExternKernelChoice (line 3242)
├── TritonTemplateCaller
└── ...

Every other isinstance check in the same file correctly uses ExternKernelCaller:

# lines 3932, 3940, 3987, 4008, 4026, 4051, 4869, 5132, 5247 — all use ExternKernelCaller
if isinstance(choice, ExternKernelCaller):  ✓

Suggested Fix

# line 3774
choice for choice in choices if isinstance(choice, ExternKernelCaller)

# line 3650
elif isinstance(choice, ExternKernelCaller):

Impact

When torch._inductor.config.deterministic = True, the algorithm selector is supposed to prefer extern (cuBLAS) kernels over Triton templates for reproducibility. Instead, it picks choices[0] regardless of type.

In _classify_kernel_operation(), extern kernel names (mm, bmm, addmm, etc.) are never matched, so the classifier falls through to shape heuristics or "other", affecting autotuning logging and filtering.

Versions

Versions

PyTorch: 2.13.0.dev20260513+cu126

cc @mruberry @kurtamohler @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @aakhundov @coconutruben @jataylo