[Multimodal][XPU]Enable vision attn backend for xpu platform (#27525)

Signed-off-by: Yan Ma <yan.ma@intel.com>
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Co-authored-by: Yejing Lai <yejing.lai@intel.com>
Co-authored-by: Guancheng Fu <110874468+gc-fu@users.noreply.github.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>

vllm/_ipex_ops.py

@@ -270,21 +270,23 @@ class ipex_ops:
 
     @staticmethod
     def flash_attn_varlen_func(
-        out: torch.Tensor,
         q: torch.Tensor,
         k: torch.Tensor,
         v: torch.Tensor,
         cu_seqlens_q: torch.Tensor,
-        seqused_k: torch.Tensor,  # we don't support this in ipex kernel
         max_seqlen_q: int,
         max_seqlen_k: int,
-        softmax_scale: float,
-        causal: bool,
-        block_table: torch.Tensor,
-        alibi_slopes: torch.Tensor | None,
+        softmax_scale: float | None = None,
+        causal: bool = False,
+        out: torch.Tensor | None = None,
+        block_table: torch.Tensor | None = None,
+        alibi_slopes: torch.Tensor | None = None,
         window_size: list[int] | None = None,
         softcap: float | None = 0.0,
+        seqused_k: torch.Tensor | None = None,
         cu_seqlens_k: torch.Tensor | None = None,
+        # passed in qwen vl
+        dropout_p: float = 0.0,
         # The following parameters are not used in ipex kernel currently,
         # we keep API compatible to CUDA's.
         scheduler_metadata=None,
@@ -295,12 +297,43 @@ class ipex_ops:
         num_splits=0,
         s_aux: torch.Tensor | None = None,
     ):
+        if out is None:
+            out = torch.empty(q.shape, dtype=q.dtype, device=q.device)
         real_window_size: tuple[int, int]
         if window_size is None:
             real_window_size = (-1, -1)
         else:
             assert len(window_size) == 2
             real_window_size = (window_size[0], window_size[1])
+
+        if block_table is None:
+            assert cu_seqlens_k is not None, (
+                "cu_seqlens_k can't be None when calling varlen_attention."
+            )
+            if softmax_scale is None:
+                softmax_scale = q.shape[-1] ** (-0.5)
+            ipex_ops.varlen_attention(
+                q.contiguous(),
+                k.contiguous(),
+                v.contiguous(),
+                out,
+                cu_seqlens_q,
+                cu_seqlens_k,
+                None,
+                max_seqlen_q,
+                max_seqlen_k,
+                0.0,
+                softmax_scale,
+                False,
+                causal,
+                False,
+                None,
+                real_window_size[0],
+                real_window_size[1],
+                -1,
+            )
+            return out
+        else:
             return ipex.llm.modules.PagedAttention.flash_attn_varlen_func(
                 out,
                 q.contiguous(),
@@ -314,6 +347,7 @@ class ipex_ops:
                 causal,
                 block_table,
                 alibi_slopes,
+                sink=s_aux,
                 softcap=softcap,
                 window_size_left=real_window_size[0],
                 window_size_right=real_window_size[1],

vllm/attention/layer.py

@@ -123,6 +123,11 @@ def maybe_get_vit_flash_attn_backend(
         ):
             attn_backend = _Backend.FLASH_ATTN
             use_upstream_fa = True
+    elif current_platform.is_xpu():
+        assert attn_backend == _Backend.FLASH_ATTN, (
+            "XPU platform only supports FLASH_ATTN as vision attention backend."
+        )
+        use_upstream_fa = False
     else:
         return _Backend.TORCH_SDPA, None
 
@@ -133,7 +138,7 @@ def maybe_get_vit_flash_attn_backend(
             if use_upstream_fa:
                 from flash_attn import flash_attn_varlen_func
             else:
-                from vllm.vllm_flash_attn import flash_attn_varlen_func
+                from vllm.attention.utils.fa_utils import flash_attn_varlen_func
     else:
         flash_attn_varlen_func = None
 
@@ -521,10 +526,6 @@ class MultiHeadAttention(nn.Module):
         # If vllm native fa is selected, we use it directly.
         use_upstream_fa = False
 
-        if current_platform.is_xpu():
-            # currently, only torch_sdpa is supported on xpu
-            self.attn_backend = _Backend.TORCH_SDPA
-        else:
         self.attn_backend = (
             backend
             if backend

vllm/attention/ops/vit_attn_wrappers.py

@@ -70,7 +70,7 @@ def flash_attn_maxseqlen_wrapper(
         if use_upstream_fa:
             from flash_attn import flash_attn_varlen_func
         else:
-            from vllm.vllm_flash_attn import flash_attn_varlen_func
+            from vllm.attention.utils.fa_utils import flash_attn_varlen_func
     q, k, v = (einops.rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v])
     output = flash_attn_varlen_func(
         q,

vllm/model_executor/models/qwen2_5_vl.py

@@ -364,6 +364,8 @@ class Qwen2_5_VisionAttention(nn.Module):
 
         if current_platform.is_rocm() and self.attn_backend == _Backend.FLASH_ATTN:
             self.use_upstream_fa = True
+        if current_platform.is_xpu():
+            self.use_upstream_fa = False
         self.is_flash_attn_backend = self.attn_backend in {
             _Backend.FLASH_ATTN,
             _Backend.ROCM_AITER_FA,
@@ -856,10 +858,7 @@ class Qwen2_5_VisionTransformer(nn.Module):
     ) -> tuple[torch.Tensor, torch.Tensor]:
         max_seqlen = torch.zeros([], device=cu_seqlens.device)
         seqlens = torch.zeros(1, device=cu_seqlens.device)
-        if (
-            self.attn_backend == _Backend.FLASH_ATTN
-            or self.attn_backend == _Backend.ROCM_AITER_FA
-        ):
+        if self.attn_backend in {_Backend.FLASH_ATTN, _Backend.ROCM_AITER_FA}:
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
         elif self.attn_backend == _Backend.XFORMERS:
             seqlens = cu_seqlens[1:] - cu_seqlens[:-1]

vllm/model_executor/models/qwen2_vl.py

@@ -789,10 +789,7 @@ class Qwen2VisionTransformer(nn.Module):
         self, cu_seqlens: torch.Tensor
     ) -> tuple[int | None, list[int] | None]:
         max_seqlen, seqlens = None, None
-        if (
-            self.attn_backend == _Backend.FLASH_ATTN
-            or self.attn_backend == _Backend.ROCM_AITER_FA
-        ):
+        if self.attn_backend in {_Backend.FLASH_ATTN, _Backend.ROCM_AITER_FA}:
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
         elif self.attn_backend == _Backend.XFORMERS:
             seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()

vllm/platforms/xpu.py

@@ -115,6 +115,12 @@ class XPUPlatform(Platform):
         device_props = torch.xpu.get_device_properties(device_id)
         return device_props.total_memory
 
+    @classmethod
+    def get_vit_attn_backend(cls, head_size: int, dtype: torch.dtype) -> _Backend:
+        from vllm.attention.backends.registry import _Backend
+
+        return _Backend.FLASH_ATTN
+
     @classmethod
     def inference_mode(cls):
         return torch.no_grad()