support mxfp8 quantization (qwen dense) (#5723)

### What this PR does / why we need it?
support mxfp8 quantization (qwen liner layer)

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef


Signed-off-by: wangyao <iwangyao@outlook.com>

vllm_ascend/quantization/w8a8mxfp8.py

@@ -0,0 +1,98 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# This file is a part of the vllm-ascend project.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+from typing import Any, Dict, Optional
+
+import torch
+import torch_npu
+from vllm.config import get_current_vllm_config
+
+
+class AscendW8A8MXFP8DynamicLinearMethod:
+    """Linear method for Ascend W8A8_DYNAMIC.
+    """
+    model_dtype = None
+
+    def __init__(self):
+        vllm_config = get_current_vllm_config()
+        self.group_size = vllm_config.quant_config.quant_description.get(
+            "group_size", 32)
+
+    @staticmethod
+    def get_weight(input_size: int, output_size: int,
+                   params_dtype: torch.dtype) -> Dict[str, Any]:
+        params_dict = {
+            "weight":
+            torch.empty(output_size, input_size, dtype=torch.float8_e4m3fn)
+        }
+        return params_dict
+
+    @staticmethod
+    def get_pertensor_param(params_dtype: torch.dtype) -> Dict[str, Any]:
+        return {}
+
+    @staticmethod
+    def get_perchannel_param(
+        output_size: int,
+        params_dtype: torch.dtype,
+    ) -> Dict[str, Any]:
+        return {}
+
+    def get_pergroup_param(self,
+                           input_size: int,
+                           output_size: int,
+                           params_dtype: torch.dtype,
+                           layer_type: Optional[str] = None) -> Dict[str, Any]:
+        params_dict = {}
+        params_dict["weight_scale"] = torch.empty(output_size,
+                                                  input_size //
+                                                  self.group_size,
+                                                  dtype=torch.uint8)
+        return params_dict
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+        tp_rank: Optional[int] = 0,
+    ) -> torch.Tensor:
+
+        quantized_x, dynamic_scale = torch_npu.npu_dynamic_mx_quant(
+            x, dst_type=torch.float8_e4m3fn)
+        pertoken_scale = dynamic_scale
+        output_dtype = x.dtype
+
+        output = torch_npu.npu_quant_matmul(
+            quantized_x,
+            layer.weight,
+            layer.weight_scale,
+            scale_dtype=torch_npu.float8_e8m0fnu,
+            pertoken_scale=pertoken_scale,
+            pertoken_scale_dtype=torch_npu.float8_e8m0fnu,
+            bias=bias,
+            output_dtype=output_dtype,
+            group_sizes=[1, 1, self.group_size])
+
+        return output
+
+    def process_weights_after_loading(self, layer):
+        n_dim, k_dim = layer.weight_scale.data.shape
+        layer.weight_scale.data = layer.weight_scale.data.reshape(
+            n_dim, k_dim // 2, 2)
+        layer.weight.data = layer.weight.data.transpose(0, 1)
+        layer.weight_scale.data = layer.weight_scale.data.transpose(0, 1)