add XPU path in apply_rotary_pos_emb_flashattn for Keye-VL models (vllm-project#7)

* add xpu path

Signed-off-by: Lin, Fanli <[email protected]>

* use partial to create a function wrapper

Co-authored-by: Copilot <[email protected]>

---------

Signed-off-by: Lin, Fanli <[email protected]>
Co-authored-by: Copilot <[email protected]>

Author: faaany

vllm/model_executor/models/keye.py

@@ -345,6 +345,10 @@ def apply_rotary_pos_emb_flashatt(
         from vllm.vllm_flash_attn.layers.rotary import apply_rotary_emb
     elif current_platform.is_rocm():
         from flash_attn.ops.triton.rotary import apply_rotary as apply_rotary_emb
+    else:
+        # For XPU and other platforms, use PyTorch fallback
+        from vllm.model_executor.layers.rotary_embedding.common import apply_rotary_emb_torch
+        apply_rotary_emb = partial(apply_rotary_emb_torch, is_neox_style=True)
 
     q_embed = apply_rotary_emb(q.float(), cos.float(), sin.float()).type_as(q)
     k_embed = apply_rotary_emb(k.float(), cos.float(), sin.float()).type_as(k)