[ROCm] Disable AITER attention query quantization on RDNA3 (#57750)

Signed-off-by: xawang <xawang@amd.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
amd-xavierwang
2026-10-06 01:37:18 +00:00
committed by GitHub
co-authored by TJian
parent 2537629eb2
commit efd001756b
@@ -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