From efd001756bff47791113857ef262d4683f7f7b55 Mon Sep 17 00:00:00 2001 From: amd-xavierwang Date: Mon, 5 Oct 2026 21:37:18 -0400 Subject: [PATCH] [ROCm] Disable AITER attention query quantization on RDNA3 (#57750) Signed-off-by: xawang Co-authored-by: TJian --- vllm/v1/attention/backends/rocm_aiter_unified_attn.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/vllm/v1/attention/backends/rocm_aiter_unified_attn.py b/vllm/v1/attention/backends/rocm_aiter_unified_attn.py index d57b420cd7..0d4fd55194 100644 --- a/vllm/v1/attention/backends/rocm_aiter_unified_attn.py +++ b/vllm/v1/attention/backends/rocm_aiter_unified_attn.py @@ -184,8 +184,11 @@ class RocmAiterUnifiedAttentionImpl(RocmAttentionImpl): ) from aiter.ops.triton.unified_attention import unified_attention + from vllm.platforms.rocm import on_gfx11 + self.unified_attention = unified_attention - self.supports_quant_query_input = True + # RDNA3 has no low precision MMA, preserve query precision + self.supports_quant_query_input = not on_gfx11() def _split_kv_cache( self, kv_cache: torch.Tensor