[CI][xlite] Match xlite backend with the current main and add e2e MoE model test coverage (vllm-project#12272)

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

This PR brings the `xlite` backend in line with the current main and
`vllm==0.24.0`, and adds e2e coverage for the xlite graph mode to catch
future changes.

Refer to PR vllm-project#12019 for the context.

- vLLM version: v0.24.0
- vLLM main:
vllm-project/vllm@85c09e9

Signed-off-by: Sijie Fu <fusijie@huawei.com>

Author: SijieFu

.github/workflows/scripts/test_config.yaml

@@ -577,6 +577,7 @@
   tests:
     - tests/ut/xlite
     - tests/e2e/pull_request/one_card/test_xlite.py
+    - tests/e2e/pull_request/two_card/test_xlite.py
 
 # Batch Invariant
 - name: batch_invariant
@@ -756,6 +757,7 @@ estimated_times:
   tests/e2e/pull_request/two_card/test_shared_expert_dp.py: 400
   tests/e2e/pull_request/two_card/test_sp_pass.py: 260
   tests/e2e/pull_request/two_card/test_hccl_weight_transfer.py: 110
+  tests/e2e/pull_request/two_card/test_xlite.py: 180
   tests/e2e/pull_request/four_card/_310p/test_dense_model_310p.py: 410
   tests/e2e/pull_request/four_card/_310p/test_moe_model_310p.py: 510
   tests/e2e/pull_request/four_card/_310p/test_vl_model_310p.py: 360

tests/e2e/pull_request/one_card/test_xlite.py

@@ -1,5 +1,5 @@
 #
-# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
 # Copyright 2023 The vLLM team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -27,39 +27,53 @@
 
 import pytest
 
-from tests.e2e.pull_request.utils import PROMPTS_SHORT, LLMTestCase, compare_logprobs
+from tests.e2e.conftest import wait_until_npu_memory_free
+from tests.e2e.pull_request.utils import PROMPTS_SHORT, compare_logprobs
 
 os.environ["VLLM_ASCEND_ENABLE_NZ"] = "2"
 
-CASE_DECODE_ONLY = LLMTestCase(
-    model="Qwen/Qwen3-0.6B",
-    prompts=PROMPTS_SHORT,
-)
-
-CASE_FULL = LLMTestCase(
-    model="Qwen/Qwen3-0.6B",
-    prompts=PROMPTS_SHORT,
-)
+MODELS: list[str] = ["Qwen/Qwen3-0.6B"]
 
 
-@pytest.mark.skip(reason="TODO: Re-enable xlite_decode_only e2e test when stable.")
-@pytest.mark.parametrize("cur_case", [CASE_DECODE_ONLY])
-def test_models_with_xlite_decode_only(cur_case: LLMTestCase):
+@pytest.mark.e2e_model(*MODELS)
+@pytest.mark.e2e_coverage(
+    arch="dense",
+    feature="xlite",
+    parallel="TP",
+    deploy="pd_mix",
+    hardware="A2",
+    quantization="BF16",
+    graph_mode="xlite_decode_only",
+)
+@pytest.mark.parametrize("model", MODELS)
+@wait_until_npu_memory_free()
+def test_models_with_xlite_decode_only(model: str):
     runner_kwargs = {
-        "model_name": cur_case.model,
+        "model_name": model,
         "max_model_len": 1024,
         "block_size": 128,
-        "additional_config": {"xlite_graph_config": {"enabled": True}},
+        "additional_config": {"xlite_graph_config": {"enabled": True, "full_mode": False}},
     }
-    compare_logprobs(runner_kwargs=runner_kwargs, prompts=cur_case.prompts)
+    compare_logprobs(runner_kwargs=runner_kwargs, prompts=PROMPTS_SHORT)
 
 
-@pytest.mark.parametrize("cur_case", [CASE_FULL])
-def test_models_with_xlite_full_mode(cur_case: LLMTestCase):
+@pytest.mark.e2e_model(*MODELS)
+@pytest.mark.e2e_coverage(
+    arch="dense",
+    feature="xlite",
+    parallel="TP",
+    deploy="pd_mix",
+    hardware="A2",
+    quantization="BF16",
+    graph_mode="xlite_full",
+)
+@pytest.mark.parametrize("model", MODELS)
+@wait_until_npu_memory_free()
+def test_models_with_xlite_full_mode(model: str):
     runner_kwargs = {
-        "model_name": cur_case.model,
+        "model_name": model,
         "max_model_len": 1024,
         "block_size": 128,
         "additional_config": {"xlite_graph_config": {"enabled": True, "full_mode": True}},
     }
-    compare_logprobs(runner_kwargs=runner_kwargs, prompts=cur_case.prompts)
+    compare_logprobs(runner_kwargs=runner_kwargs, prompts=PROMPTS_SHORT)

tests/e2e/pull_request/two_card/test_xlite.py

@@ -0,0 +1,86 @@
+#
+# Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Copyright 2023 The vLLM team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+#
+"""Verify MoE serving with xlite graph mode on two NPU cards.
+
+Run `pytest tests/e2e/pull_request/two_card/test_xlite.py`.
+"""
+
+import os
+
+import pytest
+
+from tests.e2e.conftest import VllmRunner, wait_until_npu_memory_free
+from tests.e2e.pull_request.utils import PROMPTS_SHORT
+
+os.environ["VLLM_ASCEND_ENABLE_NZ"] = "2"
+
+MODELS: list[str] = ["Qwen/Qwen3-30B-A3B"]
+
+
+@pytest.mark.e2e_model(*MODELS)
+@pytest.mark.e2e_coverage(
+    arch="moe",
+    feature="xlite",
+    parallel="TP,EP",
+    deploy="pd_mix",
+    hardware="A3",
+    quantization="BF16",
+    graph_mode="xlite_decode_only",
+)
+@pytest.mark.parametrize("model", MODELS)
+@wait_until_npu_memory_free()
+def test_models_with_xlite_decode_only(model: str):
+    with VllmRunner(
+        model,
+        tensor_parallel_size=2,
+        enable_expert_parallel=True,
+        distributed_executor_backend="mp",
+        block_size=128,
+        max_model_len=2048,
+        additional_config={"xlite_graph_config": {"enabled": True, "full_mode": False}},
+    ) as vllm_model:
+        outputs = vllm_model.generate_greedy(PROMPTS_SHORT, 3)
+
+    assert all(output[1] for output in outputs)
+
+
+@pytest.mark.e2e_model(*MODELS)
+@pytest.mark.e2e_coverage(
+    arch="moe",
+    feature="xlite",
+    parallel="TP,EP",
+    deploy="pd_mix",
+    hardware="A3",
+    quantization="BF16",
+    graph_mode="xlite_full",
+)
+@pytest.mark.parametrize("model", MODELS)
+@wait_until_npu_memory_free()
+def test_models_with_xlite_full_mode(model: str):
+    with VllmRunner(
+        model,
+        tensor_parallel_size=2,
+        enable_expert_parallel=True,
+        distributed_executor_backend="mp",
+        block_size=128,
+        max_model_len=2048,
+        additional_config={"xlite_graph_config": {"enabled": True, "full_mode": True}},
+    ) as vllm_model:
+        outputs = vllm_model.generate_greedy(PROMPTS_SHORT, 3)
+
+    assert all(output[1] for output in outputs)

vllm_ascend/xlite/xlite.py

@@ -436,15 +436,15 @@ def _build_model(self) -> None:
         layers, _ = self._get_layers_and_model_prefix()
 
         xlite_model.gate = get_layer_weights(layers, "mlp.gate.weight")
-        prefix = "mlp.experts."
+        prefix = "mlp.experts.routed_experts."
         kwargs: WeightGetterConfig = {"secondary_flattening": f"{prefix}local_num_experts", "post_processor": None}
         xlite_model.re_up_gate = get_layer_weights(layers, f"{prefix}w13_weight", **kwargs)
         xlite_model.re_down = get_layer_weights(layers, f"{prefix}w2_weight", **kwargs)
         xlite_config.experts_weight_nz = self.is_tensor_nz(xlite_model.re_up_gate[0])
 
         if self.quantization:
             kwargs["post_processor"] = self._transform_deq_scale
-            xlite_model.re_up_gate_scale = get_layer_weights(layers, f"{prefix}w13_weight_scale_fp32", **kwargs)
+            xlite_model.re_up_gate_scale = get_layer_weights(layers, f"{prefix}w13_weight_scale", **kwargs)
             xlite_model.re_down_scale = get_layer_weights(layers, f"{prefix}w2_weight_scale", **kwargs)
 
 
@@ -486,7 +486,7 @@ def _build_model(self) -> None:
         self.init_matmul_weights(layers, "se_up_gate", "mlp.shared_experts.gate_up_proj")
         self.init_matmul_weights(layers, "se_down", "mlp.shared_experts.down_proj")
 
-        prefix = "mlp.experts."
+        prefix = "mlp.experts.routed_experts."
         kwargs: WeightGetterConfig = {"secondary_flattening": f"{prefix}local_num_experts", "post_processor": None}
         xlite_model.re_up_gate = get_layer_weights(layers, f"{prefix}w13_weight", **kwargs)
         xlite_model.re_down = get_layer_weights(layers, f"{prefix}w2_weight", **kwargs)
@@ -495,7 +495,7 @@ def _build_model(self) -> None:
 
         if self.quantization:
             kwargs["post_processor"] = self._transform_deq_scale
-            xlite_model.re_up_gate_scale = get_layer_weights(layers, f"{prefix}w13_weight_scale_fp32", **kwargs)
+            xlite_model.re_up_gate_scale = get_layer_weights(layers, f"{prefix}w13_weight_scale", **kwargs)
             xlite_model.re_down_scale = get_layer_weights(layers, f"{prefix}w2_weight_scale", **kwargs)
 
 
@@ -530,7 +530,7 @@ def _build_model(self) -> None:
             layers, "block_sparse_moe.e_score_correction_bias", post_processor=lambda b: b.to(torch.float32)
         )
 
-        prefix = "block_sparse_moe.experts."
+        prefix = "block_sparse_moe.experts.routed_experts."
         kwargs: WeightGetterConfig = {"secondary_flattening": f"{prefix}local_num_experts", "post_processor": None}
         xlite_model.re_up_gate = get_layer_weights(layers, f"{prefix}w13_weight", **kwargs)
         xlite_model.re_down = get_layer_weights(layers, f"{prefix}w2_weight", **kwargs)
@@ -539,7 +539,7 @@ def _build_model(self) -> None:
 
         if self.quantization:
             kwargs["post_processor"] = self._transform_deq_scale
-            xlite_model.re_up_gate_scale = get_layer_weights(layers, f"{prefix}w13_weight_scale_fp32", **kwargs)
+            xlite_model.re_up_gate_scale = get_layer_weights(layers, f"{prefix}w13_weight_scale", **kwargs)
             xlite_model.re_down_scale = get_layer_weights(layers, f"{prefix}w2_weight_scale", **kwargs)