A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 03:30 UTC
[XPU] properly handle q_descale on XPU as quant query input not supported (#39676)
Signed-off-by: Yan Ma <yan.ma@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Yan Ma committed
Apr 15, 2026 at 21:52 UTC
3beb57a238b82fe90e8b99e009c876343b9d9703
1 file changed
+3
-1
vllm/v1/attention/backends/flash_attn.py
+3
-1
@@ -1031,7 +1031,9 @@ class FlashAttentionImpl(AttentionImpl):
1031
window_size=sliding_window_size,
1032
softcap=self.logits_soft_cap,
1033
fa_version=self.vllm_flash_attn_version,
1034
- q_descale=layer._q_scale.expand(descale_shape),
1034
+ q_descale=layer._q_scale.expand(descale_shape)
1035
+ if self.supports_quant_query_input
1036
+ else None,
1037
k_descale=layer._k_scale.expand(descale_shape),
1038
v_descale=layer._v_scale.expand(descale_shape),
1039
num_splits=1 if self.batch_invariant_enabled else 0,