[ROCm][Bugfix] Fix `fa_version` argument error in...

[ROCm][Bugfix] Fix `fa_version` argument error in `flash_attn_maxseqlen_wrapper` for ROCm without aiter (#30909)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>

vllm/attention/ops/vit_attn_wrappers.py

@@ -28,7 +28,7 @@ def flash_attn_maxseqlen_wrapper(
     max_seqlen: torch.Tensor,
     batch_size: int,
     is_rocm_aiter: bool,
-    fa_version: int,
+    fa_version: int | None,
 ) -> torch.Tensor:
     kwargs = {}
     if is_rocm_aiter:
@@ -36,6 +36,7 @@ def flash_attn_maxseqlen_wrapper(
     else:
         from vllm.attention.utils.fa_utils import flash_attn_varlen_func
 
+        if not current_platform.is_rocm() and fa_version is not None:
             kwargs["fa_version"] = fa_version
     q, k, v = (einops.rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v])
     output = flash_attn_varlen_func(
@@ -62,7 +63,7 @@ def flash_attn_maxseqlen_wrapper_fake(
     max_seqlen: torch.Tensor,
     batch_size: int,
     is_rocm_aiter: bool,
-    fa_version: int,
+    fa_version: int | None,
 ) -> torch.Tensor:
     return torch.empty_like(q)
 
@@ -82,7 +83,7 @@ def vit_flash_attn_wrapper(
     max_seqlen: torch.Tensor,
     batch_size: int,
     is_rocm_aiter: bool,
-    fa_version: int,
+    fa_version: int | None,
 ) -> torch.Tensor:
     return torch.ops.vllm.flash_attn_maxseqlen_wrapper(
         q, k, v, cu_seqlens, max_seqlen, batch_size, is_rocm_aiter, fa_version