fix fp8 kv_cache bug

Author: k50112113

vllm/v1/attention/backends/rocm_aiter_fa.py

@@ -432,6 +432,8 @@ def __init__(
         if self.sinks is not None:
             raise NotImplementedError("Sinks are not supported for ROCM AITER")
 
+        self.fp8_dtype = current_platform.fp8_dtype()
+
     def forward(
         self,
         layer: torch.nn.Module,