[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)
+                origin_weight = weight.data.transpose(1, 2)
-                weight.data = npu_format_cast(weight.data)
+                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=}")
 

test/registered/ascend/basic_function/offloading/test_npu_offload_modes.py

@@ -0,0 +1,104 @@
+import unittest
+from urllib.parse import urlparse
+
+import requests
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ascend.test_ascend_utils import DEEPSEEK_CODER_V2_LITE_WEIGHTS_PATH
+from sglang.test.ci.ci_register import register_npu_ci
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+register_npu_ci(est_time=800, suite="nightly-2-npu-a3", nightly=True)
+
+TEST_MODEL_MATRIX = {
+    DEEPSEEK_CODER_V2_LITE_WEIGHTS_PATH,
+}
+
+
+class TestAscendOffloadModes(CustomTestCase):
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = TEST_MODEL_MATRIX
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.url = urlparse(DEFAULT_URL_FOR_TEST)
+        cls.common_args = [
+            "--trust-remote-code",
+            "--disable-cuda-graph",
+            "--mem-fraction-static",
+            0.9,
+            "--attention-backend",
+            "ascend",
+            "--offload-group-size",
+            4,
+            "--offload-num-in-group",
+            1,
+            "--offload-prefetch-step",
+            1,
+            "--dp-size",
+            2,
+        ]
+
+    def run_a_test(self, offload_mode, additional_args=None):
+        """Run test for a specific offload mode."""
+        for model in self.models:
+            with self.subTest(model=model, offload_mode=offload_mode):
+                print(f"##=== Testing {offload_mode} offload: {model} ===##")
+
+                args = [
+                    *self.common_args,
+                    "--offload-mode",
+                    offload_mode,
+                ]
+
+                if additional_args:
+                    args.extend(additional_args)
+
+                process = popen_launch_server(
+                    model,
+                    self.base_url,
+                    timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+                    other_args=args,
+                )
+
+                try:
+                    # Check if server is running (basic functionality test)
+                    response = requests.post(
+                        f"{DEFAULT_URL_FOR_TEST}/generate",
+                        json={
+                            "text": "Where is the capital of France?",
+                            "sampling_params": {
+                                "temperature": 0,
+                                "max_new_tokens": 32,
+                            },
+                        },
+                    )
+                    self.assertEqual(
+                        response.status_code,
+                        200,
+                        f"The request status code is not 200, server failed to respond for {offload_mode}",
+                    )
+                    self.assertIn(
+                        "Paris",
+                        response.text,
+                        f"The inference result does not include Paris, server failed to respond for {offload_mode}",
+                    )
+                finally:
+                    kill_process_tree(process.pid)
+
+    def test_offload_mode_cpu(self):
+        """Test offload mode: cpu"""
+        self.run_a_test("cpu")
+
+    def test_offload_mode_sharded_gpu(self):
+        """Test offload mode: sharded_gpu"""
+        self.run_a_test("sharded_gpu")
+
+
+if __name__ == "__main__":
+    unittest.main()