[NPU] fix some npu error with OffloaderV2 (#19541)

Co-authored-by: Jianzhao Xu <xujianchao@huawei.com>
Co-authored-by: sglang-npu-bot <sglangnpu@163.com>

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

@@ -170,6 +170,10 @@ def npu_format_cast(
         )
         return tensor
 
+    # Skip format cast for meta tensors (used in offloader)
+    if tensor.device.type == "meta":
+        return tensor
+
     return torch.ops.npu.npu_format_cast(tensor, acl_format.value)
 
 

python/sglang/srt/layers/quantization/unquant.py

@@ -317,8 +317,10 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
         if _is_npu:
             for weight_name in ["w13_weight", "w2_weight"]:
                 weight = getattr(layer, weight_name)
-                weight.data = weight.data.transpose(1, 2)
-                weight.data = npu_format_cast(weight.data)
+                origin_weight = weight.data.transpose(1, 2)
+                new_weight = origin_weight.contiguous()
+                origin_weight.untyped_storage().resize_(0)
+                weight.data = npu_format_cast(new_weight)
 
         return
 

python/sglang/srt/utils/offloader.py

@@ -452,6 +452,10 @@ def _move_param_to_meta(module, param_name):
             data=new_data,
             requires_grad=False,
         )
+        if hasattr(old_param, "weihgt_loader"):
+            new_param.weight_loader = old_param.weight_loader
+        else:
+            new_param.weight_loader = lambda *args, **kwargs: None
     else:
         raise ValueError(f"Unknown {old_param_type=} {old_param=}")