[refactor] refactor weight trans nz and transpose (#4878)

### What this PR does / why we need it?

Now `VLLM_ASCEND_ENABLE_NZ` will have three options:
0: disable nz;
1: only quant case enable nz;
2: enable nz as long as possible;

And `VLLM_ASCEND_ENABLE_NZ`=1 by default.

All cases are shown in the table below:

|  | W4A4 | W4A8 | W8A8 | fp16/bf16 | fp32 |
|---|---|---|---|---|---|
| trans nz | can't support nz | trans nz by default | trans nz by
default | trans nz when VLLM_ASCEND_ENABLE_NZ is 2 | can't support nz |
| transpose | only support not transpose case | only support transpose
case | only support transpose case | linear: only support not transpose
case<br>gmm: only support transpose case | same to fp16/bf16 |

Some exceptional cases:
1. MLAPO op need to do some additional processing on the weights,
including trans nz. If use MLAPO op, some weight will be transformed to
nz forcely;
2. MLA/SFA's weight `W_UV` will be used by op
`torch.ops._C_ascend.batch_matmul_transpose`, and this op can't support
nz currently;

### Does this PR introduce _any_ user-facing change?
Now fp16/bf16 weight will not trans nz by default.

### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: zzzzwwjj <1183291235@qq.com>

vllm_ascend/quantization/w8a8.py

@@ -21,9 +21,8 @@ import torch
 import torch_npu
 from vllm.forward_context import get_forward_context
 
-from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ,
-                               COMPRESSED_TENSORS_METHOD, AscendDeviceType,
-                               get_ascend_device_type, is_enable_nz)
+from vllm_ascend.utils import (COMPRESSED_TENSORS_METHOD, AscendDeviceType,
+                               get_ascend_device_type, maybe_trans_nz)
 
 
 def quant_per_tensor(in_tensor: torch.Tensor,
@@ -42,9 +41,7 @@ class AscendW8A8LinearMethod:
     """
 
     def __init__(self) -> None:
-        # aclnn quant matmul requires to transpose matrix B, set to true by default.
-        self.transpose_weight = get_ascend_device_type(
-        ) != AscendDeviceType._310P
+        pass
 
     @staticmethod
     def get_weight(
@@ -189,11 +186,9 @@ class AscendW8A8LinearMethod:
         layer.aclnn_input_offset = torch.nn.Parameter(
             layer.input_offset.data.repeat(expanding_factor),
             requires_grad=False).to(layer.aclnn_input_scale.dtype)
-        if self.transpose_weight:
+        if get_ascend_device_type() != AscendDeviceType._310P:
             layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()
-        if is_enable_nz():
-            layer.weight.data = torch_npu.npu_format_cast(
-                layer.weight.data, ACL_FORMAT_FRACTAL_NZ)
+        layer.weight.data = maybe_trans_nz(layer.weight.data)
         layer.weight_scale.data = torch.flatten(layer.weight_scale.data)
         layer.weight_offset.data = torch.flatten(layer.weight_offset.data)
         ascend_quant_method = getattr(layer, "ascend_quant_method", "")