dynamic distpatch of fp8 kernels (#14245)

Signed-off-by: Jeff Daily <jeff.daily@amd.com>

tests/kernels/quant_utils.py

@@ -9,8 +9,7 @@ from vllm.platforms import current_platform
 # Using the default value (240.0) from pytorch will cause accuracy
 # issue on dynamic quantization models. Here use 224.0 for rocm.
 ROCM_FP8_MAX = 224.0
-FP8_DTYPE = torch.float8_e4m3fnuz if current_platform.is_rocm() \
-                else torch.float8_e4m3fn
+FP8_DTYPE = current_platform.fp8_dtype()
 
 
 def as_float32_tensor(x: Union[float, torch.tensor]) -> torch.tensor: