[BugFix][310p] Qwen3.5 shared-experts linear failuer when out_dim=1 (vllm-project#11714)

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

This PR fixes a failure in Qwen3.5 shared-experts linear layer on 310P
devices when `out_dim=1`.
On 310P, `torch_npu` rejects `FRACTAL_NZ` matmul when the weight-side
matrix has $N=1$ or $K=1$.
To resolve this, we introduce `_should_keep_nd_for_310p_weight` to
detect if the weight has a dimension of 1 on 310P, and if so, keep the
weight in ND format instead of converting it to `FRACTAL_NZ`.

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

No.

### How was this patch tested?

Tested on 310P Ascend NPU.


- vLLM version: v0.23.0
- vLLM main:
vllm-project/vllm@1f486d9

Signed-off-by: Tflowers-0129 <2906339855@qq.com>
Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>

Author: Tflowers-0129

vllm_ascend/ops/linear.py

@@ -45,6 +45,7 @@
     AscendDeviceType,
     enable_sp,
     get_ascend_device_type,
+    is_310p,
     maybe_trans_nz,
 )
 
@@ -75,16 +76,26 @@ def unquantized_gemm_fake(
 )
 
 
+def _should_keep_nd_for_310p_weight(weight: torch.Tensor) -> bool:
+    return is_310p() and weight.ndim >= 2 and (weight.shape[-1] == 1 or weight.shape[-2] == 1)
+
+
 class AscendUnquantizedLinearMethod(UnquantizedLinearMethod):
     """Linear method without quantization"""
 
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
         super().process_weights_after_loading(layer)
+        keep_nd_weight = _should_keep_nd_for_310p_weight(layer.weight.data)
         # must use fp32 to avoid accuracy degradation in dsv4.
         if getattr(layer, "precast_fp32_weight", False):
-            layer.weight_fp32 = maybe_trans_nz(layer.weight.data.to(torch.float32))
+            weight_fp32 = layer.weight.data.to(torch.float32)
+            layer.weight_fp32 = weight_fp32 if keep_nd_weight else maybe_trans_nz(weight_fp32)
         if "conv1d" not in layer.prefix:
-            layer.weight.data = maybe_trans_nz(layer.weight.data)
+            # 310P torch_npu rejects FRACTAL_NZ matmul when the weight-side
+            # matrix has n=1 or k=1. Keep scalar gates such as Qwen MoE's
+            # shared_expert_gate in ND format, leaving non-310P policy intact.
+            if not keep_nd_weight:
+                layer.weight.data = maybe_trans_nz(layer.weight.data)
 
     def apply(
         self,