Severity: MEDIUM-HIGH — silent systematic bias, no error.
The aie_kernels elementwise kernels (generic add.cc, mul.cc, rope.cc; aie2/ + aie2p/ silu.cc) perform their final rounding with whatever mode the hardware was last set to. On the fused dispatch path a prior kernel can leave floor active, biasing large-magnitude results by ~0.7% (e.g. isolated add: max dev 0.25 at |~33|, systematic sign-dependent).
Root cause
No ::aie::set_rounding(aie::rounding_mode::conv_even) at kernel entry — rounding mode is global NPU state, not per-kernel.
Fix (committed in 76eee2a)
Explicit conv_even at entry of add.cc, mul.cc, rope.cc, silu.cc. Isolated add: bias 0.25 → 0.0000; operator tests 27/27 + 17/17 pass; llama prefill bit-identical across runs.
Regression coverage
iron/operators/ elementwise/silu/rope tests (rounding-sensitive cases)- llama-3.2-1B prefill determinism (bit-identical logits)
Related: mlir-aie #3481 (rounding-mode convention).
Severity: MEDIUM-HIGH — silent systematic bias, no error.
The
aie_kernelselementwise kernels (genericadd.cc,mul.cc,rope.cc;aie2/+aie2p/silu.cc) perform their final rounding with whatever mode the hardware was last set to. On the fused dispatch path a prior kernel can leave floor active, biasing large-magnitude results by ~0.7% (e.g. isolated add: max dev 0.25 at |~33|, systematic sign-dependent).Root cause
No
::aie::set_rounding(aie::rounding_mode::conv_even)at kernel entry — rounding mode is global NPU state, not per-kernel.Fix (committed in
76eee2a)Explicit
conv_evenat entry ofadd.cc,mul.cc,rope.cc,silu.cc. Isolated add: bias 0.25 → 0.0000; operator tests 27/27 + 17/17 pass; llama prefill bit-identical across runs.Regression coverage
iron/operators/elementwise/silu/rope tests (rounding-sensitive cases)Related: mlir-aie #3481 (rounding-mode convention).