simplify

sam2/modeling/sam/transformer.py

@@ -27,8 +27,10 @@ def _can_use_flash_attn(q: torch.Tensor) -> bool:
     # FlashAttention works on CUDA with fp16/bf16 and (usually) Ampere+ GPUs
     if not q.is_cuda:
         return False
-    major, _ = torch.cuda.get_device_capability(q.device)
-    return q.dtype in (torch.float16, torch.bfloat16) and major >= 8  # A100/RTX30+ typically
+    # major, _ = torch.cuda.get_device_capability(q.device)
+    # return q.dtype in (torch.float16, torch.bfloat16) and major >= 8  # A100/RTX30+ typically
+
+    return True
 
 
 class TwoWayTransformer(nn.Module):