[Diffusion] [NPU] Wan2.2-T2V-A14B-Diffusers modelslim quantization support (#17996)

Co-authored-by: ronnie_zheng <zl19940307@163.com>

python/sglang/srt/hardware_backend/npu/quantization/linear_method_npu.py

@@ -105,7 +105,7 @@ class NPUW8A8Int8DynamicLinearMethod(_NPULinearMethodBase):
             quant_out,
             layer.weight,
             layer.weight_scale,
-            pertoken_scale=dynamic_scale,
+            pertoken_scale=dynamic_scale.flatten(),
             bias=bias,
             output_dtype=original_dtype,
         )
@@ -137,7 +137,7 @@ class NPU_W4A4DynamicLinearMethod(_NPULinearMethodBase):
             quant_out,
             layer.weight,
             layer.weight_scale,
-            pertoken_scale=dynamic_scale,
+            pertoken_scale=dynamic_scale.flatten(),
             bias=bias,
             output_dtype=original_dtype,
         )

python/sglang/srt/hardware_backend/npu/utils.py

@@ -123,9 +123,10 @@ def npu_format_cast(
     if envs.SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT.get():
         return tensor
 
-    import torch_npu
-
-    return torch_npu.npu_format_cast(tensor, acl_format.value)
+    if tensor.device == torch.device("cpu"):
+        return torch.ops.npu.npu_format_cast(tensor.npu(), acl_format.value).cpu()
+    else:
+        return torch.ops.npu.npu_format_cast(tensor, acl_format.value)
 
 
 def get_indexer_weight_stream():