Add MLX hardtanh op handler

[atharvjairath]

Fixes #18921

Adds MLX delegate support for aten.hardtanh.default by lowering it to the existing ClipNode path with the operator's min_val and max_val bounds. This enables bounded activation models, including ReLU6-style hardtanh usage, to stay delegated to MLX instead of failing as an unsupported op.

This also adds focused MLX op tests for:

• default hardtanh bounds [-1.0, 1.0]
• ReLU6 bounds [0.0, 6.0]
• symmetric custom bounds [-2.0, 2.0]
• asymmetric custom bounds [-0.25, 0.75]

Test plan:

lintrunner backends/mlx/ops.py backends/mlx/test/test_ops.py

ok No lint issues.

CPLUS_INCLUDE_PATH=/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1 python -m executorch.backends.mlx.test.run_all_tests --rebuild hardtanh

Rebuilding op_test_runner in /Users/atharvjairath/Desktop/timepass/executorch/cmake-out...
Build succeeded.

Running test: hardtanh_min-1.0_max1.0_2x3x4
✓ MLX delegation verified
C++ binary output: OK
✓ PASSED: All outputs match

Running test: hardtanh_min0.0_max6.0_4x8
✓ MLX delegation verified
C++ binary output: OK
✓ PASSED: All outputs match

Running test: hardtanh_min-2.0_max2.0_10
✓ MLX delegation verified
C++ binary output: OK
✓ PASSED: All outputs match

Running test: hardtanh_min-0.25_max0.75_2x8x16
✓ MLX delegation verified
C++ binary output: OK
✓ PASSED: All outputs match

TEST SUMMARY
Passed: 4
Failed: 0

This follows up on #18986 by adding custom min/max bound coverage and including the requested local test output.

cc @metascroy

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19776

• 📄 Preview Python docs built from this PR

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

❌ 1 New Failure, 6 Pending

As of commit 542e9dc with merge base eeb0646 ():

NEW FAILURE - The following job has failed:

• pull / test-llama-runner-qnn-linux (fp32, qnn_8a8w, qnn) / linux-job (gh)
  RuntimeError: Command docker exec -t ab80ad944718fe95ea1bb8c20664ab85f20a4f09b0d63a1f0e581b78059c0587 /exec failed with exit code 92

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

[linux-foundation-easycla]

The committers listed above are authorized under a signed CLA.

• ✅ login: atharvjairath / name: atharvjairath (19fa148)

[atharvjairath]

@pytorchbot label "release notes: apple"

[atharvjairath]

cc: @nil-is-all

[atharvjairath]

Friendly bump, PR is ready for review. Added MLX support for aten.hardtanh.default via the existing ClipNode path, with tests for default/custom bounds and ReLU6 usage.

Would appreciate a review when possible. cc @metascroy

[mergennachin]

@claude Review this PR

[claude]

Claude finished @mergennachin's task in 1m 37s —— View job

---

Review of PR #19776 — Add MLX hardtanh op handler

• Read changed files and gather context
• Review backends/mlx/ops.py changes
• Review backends/mlx/test/test_ops.py changes
• Post review feedback

---

Summary

Clean, well-structured PR that adds aten.hardtanh.default support to the MLX backend by lowering to the existing ClipNode. The implementation correctly mirrors the adjacent _clamp_handler pattern and the tests provide good coverage across bound configurations.

backends/mlx/ops.py

Looks good. The handler correctly:

• Uses require_args / require_kwargs matching the established pattern
• Defaults to [-1.0, 1.0] matching the ATen schema (Scalar min_val=-1, Scalar max_val=1)
• Retrieves dtype from input tensor metadata for emit_lifted_constant
• Emits a ClipNode with both bounds always present (valid since hardtanh always has finite bounds)

One minor observation: since the ATen schema for hardtanh.default always includes the min/max scalars as positional args (they have defaults but torch.export materializes them explicitly), the len(args) > 1 / len(args) > 2 fallback guards on lines 2937-2938 are defensive-only and will likely never be hit in practice. This is fine for robustness — just noting it's technically dead code under normal export paths.

backends/mlx/test/test_ops.py

Looks good. The HardtanhTest class:

• Follows the same OpTestCase / @register_test pattern as neighboring tests (e.g., ClampTest)
• Covers default bounds, ReLU6, symmetric, and asymmetric configurations
• Uses * 4 scaling on random inputs to ensure values actually exceed bounds (good practice)
• Tight tolerances (1e-5) are appropriate for a float32 clamp op

Verdict

Approve — this is a straightforward, correct addition that follows existing patterns. No issues found.

---

[metascroy]

Looks great! Let's see what CI says.