disable padding for non-cuda cases.

Signed-off-by: charlifu <charlifu@amd.com>

Author: charlifu

tests/compile/test_fusion.py

@@ -51,8 +51,7 @@ def forward(self, x):
 @pytest.mark.parametrize("hidden_size", [64, 3392, 4096])
 @pytest.mark.parametrize("num_tokens", [7, 256, 533, 2048, 2049])
 @pytest.mark.parametrize("eps", [1e-5, 1e-6])
-@pytest.mark.skipif(envs.VLLM_TARGET_DEVICE != "cuda"
-                    and envs.VLLM_TARGET_DEVICE != "rocm",
+@pytest.mark.skipif(envs.VLLM_TARGET_DEVICE not in ["cuda", "rocm"],
                     reason="Only test on CUDA and Rocm")
 def test_fusion_rmsnorm_quant(dtype, hidden_size, num_tokens, eps):
     torch.set_default_device("cuda")

vllm/model_executor/layers/quantization/utils/w8a8_utils.py

@@ -125,7 +125,7 @@ def apply_fp8_linear(
         qinput, x_scale = ops.scaled_fp8_quant(
             input_2d,
             input_scale,
-            num_token_padding=None,
+            num_token_padding=17 if current_platform.is_cuda() else None,
             use_per_token_if_dynamic=use_per_token_if_dynamic)
 
         per_tensor_weights = (weight_scale.numel() == 1)