fix div_tensor_mode accuracy failure on mthreads backends

[lyujheng]

PR Category

OP Test

Type of Change

Bug Fix

Description

Changed accuracy test file of div_tensor_mode, relaxed atol from 1e-4 to 1.0 for test_div_tensor_mode_float when rounding_mode is trunc or floor, reference runs on CPU (--ref cpu), and backend is mthreads.

case:
When the quotient a/b is extremely close to an integer boundary (e.g., 7.9999999 vs 8.0000001), a 1-ULP difference in float32 division between the mthreads GPU and CPU implementations gets amplified to ±1.0 after trunc/floor truncation. This is inherent floating-point behavior with discontinuous rounding functions, not an operator bug.

Issue

N/A

Progress

• Change is properly reviewed (1 reviewer required, 2 recommended).
• Change is responded to an issue.
• [√] Change is fully covered by a UT.

Performance

Accuracy