[ROCm][Bugfix] Use platform specific FP8 dtype (#15717)

Signed-off-by: Gregory Shtrasberg <[email protected]>

Author: gshtras

vllm/attention/ops/prefix_prefill.py

@@ -753,7 +753,7 @@ def context_attention_fwd(q,
             assert (v_cache.dtype == torch.uint8)
 
             if kv_cache_dtype in ("fp8", "fp8_e4m3"):
-                target_dtype = torch.float8_e4m3fn
+                target_dtype = current_platform.fp8_dtype()
             elif kv_cache_dtype == "fp8_e5m2":
                 target_dtype = torch.float8_e5m2
             else: