[Feat] Unquantized Linear to nz and control all nz-cast (#3356)

### What this PR does / why we need it?
Currently, when executing to the Linear layer of models in vLLM-Ascend,
the weights format is ND in unquantized case and skipped ascend case.
This PR supplements the execution logic for Linear layer. We use a new
global variable: VLLM_ASCEND_ENABLE_NZ. When VLLM_ASCEND_ENABLE_NZ=1 and
CANN version is 8.3, the weights of the Linear layer will be converted
to FRACTAL_NZ, in both unquantized case and skipped ascend case. We also
use VLLM_ASCEND_ENABLE_NZ to control the existing NZ conversion, such as
w8a8-quantized case.

### Does this PR introduce _any_ user-facing change?
Add a new global variable VLLM_ASCEND_ENABLE_NZ. If you want to use NZ
format, you should set VLLM_ASCEND_ENABLE_NZ=1.

### How was this patch tested?

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

Signed-off-by: anon189Ty <Stari_Falcon@outlook.com>

vllm_ascend/attention/mla_v1.py

@@ -27,6 +27,8 @@ from vllm_ascend.multistream.base import MSAttentionMetadataSplitConfig
 from vllm_ascend.multistream.context import get_multistream_comm_context
 from vllm_ascend.multistream.ms_split import model_input_split_v1_mla_attn
 from vllm_ascend.ops.weight_prefetch import maybe_npu_prefetch
+from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, ACL_FORMAT_FRACTAL_NZ,
+                               is_enable_nz)
 from vllm_ascend.worker.npu_input_batch import InputBatch
 
 if TYPE_CHECKING:
@@ -595,6 +597,10 @@ class AscendMLAImpl(MLAAttentionImpl):
                 del eye
                 # standardize to (output, input)
                 return dequant_weights.T
+            # Weight will be reshaped next. To be on the safe side, the format
+            # of the weight should be reverted to FRACTAL_AND.
+            layer.weight.data = torch_npu.npu_format_cast(
+                layer.weight.data, ACL_FORMAT_FRACTAL_ND)
             return layer.weight
 
         # we currently do not have quantized bmm's which are needed for
@@ -623,6 +629,12 @@ class AscendMLAImpl(MLAAttentionImpl):
         # Convert from (L, N, P) to (N, P, L)
         self.W_UK_T = W_UK.permute(1, 2, 0).contiguous()
 
+        # Function `get_and_maybe_dequant_weights` will cast the weights to
+        # FRACTAL_AND. So we need to cast to FRACTAL_NZ again.
+        if is_enable_nz():
+            self.kv_b_proj.weight.data = torch_npu.npu_format_cast(
+                self.kv_b_proj.weight.data, ACL_FORMAT_FRACTAL_NZ)
+
         # Waiting for BMM NZ support
         # self.W_UV.data = torch_npu.npu_format_cast(self.W_UV.data, 29)
         # self.W_UK_T.data = torch_npu.npu_format_cast(self.W_UK_T.data, 29)