Bump PyTorch pins to 2.12

[JacobSzwejbka]

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19643

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• Run pull request jobs on OSDC runners in shadow mode

❌ 3 New Failures, 118 Pending, 1 Unclassified Failure

As of commit cba52c9 with merge base 2a0a2f8 ():

NEW FAILURES - The following jobs have failed:

• MLX / test-mlx-qwen35-moe / test-mlx-qwen35-moe (gh)
  RuntimeError: Command bash /Users/runner/work/_temp/exec_script failed with exit code 1
• pull / test-llama-runner-qnn-linux (fp32, qnn_16a16w, qnn) / linux-job (gh)
  RuntimeError: Command docker exec -t 1c5eda488c794715dfb83edeb9a2806260ca2bfe4fc1104ba667d82133e82de0 /exec failed with exit code 1
• trunk / test-arm-backend-ethos-u (test_smaller_stories_llama) / linux-job (gh)
  RuntimeError: Command docker exec -t fcf7c1b1c3485dfac6cb8a042984bc5ab8065fa502ecd168ce42a74ebf128717 /exec failed with exit code 1

UNCLASSIFIED FAILURE - DrCI could not classify the following job because the workflow did not run on the merge base. The failure may be pre-existing on trunk or introduced by this PR:

• Check Labels / Check labels (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)
  RuntimeError: GraphQL query

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[JacobSzwejbka]

ARM TOSA failure context for the FoldAndAnnotateQParamsPass change:

The upstream PyTorch change that exposed this is pytorch/pytorch@d8253fa (Add dtype mismatch validation for mm/bmm in meta registrations, pytorch/pytorch#173459). That commit added a dtype check in common_meta_baddbmm_bmm:

torch._check(
    batch1.dtype == batch2.dtype,
    lambda: f"expected scalar type {batch1.dtype} but found {batch2.dtype}",
)

release/2.11 does not have this check; release/2.12 does, which is why the pin bump started failing ARM INT TOSA tests. Note: this was later reverted on PyTorch main by pytorch/pytorch@3208959, but the release/2.12 branch currently still contains the stricter meta check.

The ExecuTorch failure is in an intermediate graph state, not the final lowered TOSA graph. FoldAndAnnotateQParamsPass folds dq -> aten op -> q by removing the explicit q/dq nodes and storing their quantization params in node.meta["input_qparams"] / node.meta["output_qparams"]. Immediately after that fold, before later ARM passes insert rescales / rewrite bmm to TOSA MATMUL, the graph can temporarily contain quantized aten.bmm metadata. Retrying PyTorch's stricter fake aten.bmm kernel on that transient form now raises errors like:

FoldAndAnnotateQParamsPass: expected scalar type torch.float32 but found torch.int32

The ARM patch only special-cases quantized aten.bmm during FoldAndAnnotateQParamsPass metadata refresh. It preserves the existing output shape metadata and sets the fake output dtype from the folded output qparams, avoiding PyTorch's stricter fake bmm validation on the transient graph. The existing later ARM passes still own the actual rescale and TOSA MATMUL lowering path.