[mark14wu]

Summary

A downcast with an explicit rounding mode, such as x.to(tl.float8e5, fp_downcast_rounding="rtne"), fails under the sanitizer and the race detector:

ValueError: Unsupported type: <class 'triton._C.libtriton.ir.ROUNDING_MODE'>

The builder's create_fp_to_fp receives an ir.ROUNDING_MODE enum. FpToFpSymbolicExpr added it as a child expression, and SymbolicExpr.from_value cannot infer a dtype for the enum. The child now holds the enum's integer value, so it is an ordinary constant and still distinguishes the two rounding modes in signature().

Tests

• New in tests/end_to_end/test_sanitizer.py: test_sanitizer_supports_fp_downcast_rounding, an fp32 to fp8e5 cast with fp_downcast_rounding="rtne" 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	0.072s	0.072s	+0.1%	20 / 20
gemm_oob	0.081s	0.081s	+0.4%	20 / 20
indirect_load	0.014s	0.014s	-0.4%	20 / 20
nested_loop	0.149s	0.149s	+0.1%	20 / 20
block_pointer_loop_advance	0.085s	0.084s	-0.9%	20 / 20
liger_jsd	0.100s	0.100s	-0.1%	20 / 20
flaggems_layernorm	0.260s	0.260s	+0.0%	20 / 20
swiglu	0.122s	0.121s	-0.1%	20 / 20
cross_entropy	0.679s	0.679s	-0.1%	20 / 20
fused_linear_jsd	0.152s	0.152s	-0.3%	20 / 20
Total	1.714s	1.713s	-0.1%	N/A

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