From 3beb57a238b82fe90e8b99e009c876343b9d9703 Mon Sep 17 00:00:00 2001 From: Yan Ma Date: Wed, 15 Apr 2026 21:52:58 +0800 Subject: [PATCH] [XPU] properly handle q_descale on XPU as quant query input not supported (#39676) Signed-off-by: Yan Ma Co-authored-by: Kunshang Ji --- vllm/v1/attention/backends/flash_attn.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/vllm/v1/attention/backends/flash_attn.py b/vllm/v1/attention/backends/flash_attn.py index 0e926a99b8..6af0fa7c49 100755 --- a/vllm/v1/attention/backends/flash_attn.py +++ b/vllm/v1/attention/backends/flash_attn.py @@ -1031,7 +1031,9 @@ class FlashAttentionImpl(AttentionImpl): window_size=sliding_window_size, softcap=self.logits_soft_cap, fa_version=self.vllm_flash_attn_version, - q_descale=layer._q_scale.expand(descale_shape), + q_descale=layer._q_scale.expand(descale_shape) + if self.supports_quant_query_input + else None, k_descale=layer._k_scale.expand(descale_shape), v_descale=layer._v_scale.expand(descale_shape), num_splits=1 if self.batch_invariant_enabled else 0,