From c41b2d26e551a1628fc3c1c866b8cf0686e090e6 Mon Sep 17 00:00:00 2001 From: Aleksi Vesanto Date: Mon, 5 Oct 2026 10:21:04 +0300 Subject: [PATCH] [diffusion][ROCm][Perf]: Set gfx942 AITER FMHA rounding mode to rtz instead of rtna (#28650) Co-authored-by: jacky.cheng --- .../runtime/layers/attention/backends/aiter.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/python/sglang/multimodal_gen/runtime/layers/attention/backends/aiter.py b/python/sglang/multimodal_gen/runtime/layers/attention/backends/aiter.py index 095bdfd069..2d15e1ab2c 100755 --- a/python/sglang/multimodal_gen/runtime/layers/attention/backends/aiter.py +++ b/python/sglang/multimodal_gen/runtime/layers/attention/backends/aiter.py @@ -33,6 +33,12 @@ _FMHA_FP8_HEAD_DIM = 128 _FP8_MAX = torch.finfo(_fp8_dtype).max +# AITER FAv3 on gfx942 supports three different rounding modes: +# 0 = RTNE (round to nearest even) +# 1 = RTNA (round to nearest away) +# 2 = RTZ (round to zero) +AITER_BF16_CVT_MODE: int = 2 + if _use_fp8_attn: logger.info("DiT FP8 attention enabled via SGLANG_DIFFUSION_AITER_FP8_ATTN=1") @@ -344,6 +350,7 @@ class AITerImpl(AttentionImpl): causal=self.causal, return_attn_probs=False, return_lse=False, + how_v3_bf16_cvt=AITER_BF16_CVT_MODE, ) return output