[mark14wu]

Summary

Every kernel that calls tl.rand, tl.randn, or tl.randint fails under the sanitizer and the race detector:

NotImplementedError: Patching math ops not yet supported

Their Philox rounds call tl.umulhi. TRITON_NAMESPACES registered Umulhi twice: as the frontend op tl.math.umulhi and as the interpreter-builder op create_umulhi. A client that overrides Umulhi reached the frontend patch first, and run_op_overrider raised for every tl.math op. tl.math.umulhi calls the builder's create_umulhi, whose patch already dispatches to the clients (SymbolicClient._op_umulhi_overrider takes builder handles), so this PR removes the tl.math entry and the branch that raised.

Tests

• New in tests/end_to_end/test_sanitizer.py: test_sanitizer_supports_tl_rand, a kernel that stores tl.rand(seed, offs) under the sanitizer. It fails on main with the error above.
• pytest tests/: 366 passed, 6 skipped on main; 367 passed, 6 skipped with this PR (Python 3.12, Triton 3.8.0).
• ruff check and ruff format --check (v0.1.9) pass on the changed files.

🤖 Generated with Claude Code

[github-actions]

Performance Benchmark

Benchmark	main (min)	PR (min)	Change	Samples
gemm	0.051s	0.052s	+0.8%	20 / 20
gemm_oob	0.057s	0.058s	+0.5%	20 / 20
indirect_load	0.010s	0.010s	+0.1%	20 / 20
nested_loop	0.105s	0.106s	+0.8%	20 / 20
block_pointer_loop_advance	0.056s	0.057s	+0.2%	20 / 20
liger_jsd	0.072s	0.072s	+0.2%	20 / 20
flaggems_layernorm	0.183s	0.184s	+0.5%	20 / 20
swiglu	0.086s	0.086s	-0.0%	20 / 20
cross_entropy	0.481s	0.481s	-0.0%	20 / 20
fused_linear_jsd	0.110s	0.110s	+0.2%	20 / 20
Total	1.213s	1.216s	+0.2%	N/A

Iterations: 1 warmup + 20 measured
Samples are shown as main / PR; long pytest benchmarks may use fewer samples.