From f43d7e1a0f79eb0a19ca04710503d7b03e8f05a6 Mon Sep 17 00:00:00 2001 From: Ryan Rock Date: Tue, 13 Oct 2026 18:55:19 -0500 Subject: [PATCH] [CI][AMD] Gate Aiter MHA FA on dtype (#60679) Signed-off-by: Ryan Rock Co-authored-by: Micah Williamson --- vllm/platforms/rocm.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/vllm/platforms/rocm.py b/vllm/platforms/rocm.py index c34ad0c363..ca5b9db603 100644 --- a/vllm/platforms/rocm.py +++ b/vllm/platforms/rocm.py @@ -887,7 +887,11 @@ class RocmPlatform(Platform): from vllm._aiter_ops import rocm_aiter_ops - if rocm_aiter_ops.is_mha_enabled() and on_cdna(): + if ( + rocm_aiter_ops.is_mha_enabled() + and on_cdna() + and (dtype == torch.float16 or dtype == torch.bfloat16) + ): logger.info_once("Using AITER Flash Attention backend for ViT model.") return AttentionBackendEnum.ROCM_AITER_FA