add FULL graph.

Signed-off-by: wangxiaoxin-sherie <[email protected]>

Author: wangxiaoxin-sherie

.github/workflows/_e2e_test.yaml

@@ -175,6 +175,7 @@ jobs:
           VLLM_USE_MODELSCOPE: True
         if: ${{ inputs.type == 'full' }}
         run: |
+          pytest -sv tests/e2e/multicard/test_full_graph_mode.py
           pytest -sv tests/e2e/multicard/test_data_parallel.py
           pytest -sv tests/e2e/multicard/test_expert_parallel.py
           # pytest -sv tests/e2e/multicard/test_external_launcher.py

tests/e2e/multicard/test_full_graph_mode.py

@@ -29,7 +29,7 @@
 from tests.e2e.model_utils import check_outputs_equal
 
 
-def test_models_distributed_Qwen3_MOE_TP2_WITH_FULLGRAPH():
+def test_models_distributed_Qwen3_MOE_TP2_WITH_FULL_DECODE_ONLY():
     if 'HCCL_OP_EXPANSION_MODE' in os.environ:
         del os.environ['HCCL_OP_EXPANSION_MODE']
     prompts = [
@@ -70,3 +70,45 @@ def test_models_distributed_Qwen3_MOE_TP2_WITH_FULLGRAPH():
         name_0="vllm_eager_outputs",
         name_1="vllm_fullgraph_outputs",
     )
+
+def test_models_distributed_Qwen3_MOE_TP2_WITH_FULL():
+    if 'HCCL_OP_EXPANSION_MODE' in os.environ:
+        del os.environ['HCCL_OP_EXPANSION_MODE']
+    prompts = [
+        "Hello, my name is", "The president of the United States is",
+        "The capital of France is", "The future of AI is"
+    ]
+    model = "Qwen/Qwen3-30B-A3B"
+    sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
+    with VllmRunner(model,
+                    max_model_len=1024,
+                    tensor_parallel_size=2,
+                    enforce_eager=False,
+                    compilation_config={"cudagraph_mode":
+                                        "FULL"}) as runner:
+        vllm_fullgraph_outputs = runner.model.generate(prompts,
+                                                       sampling_params)
+
+    with VllmRunner(
+            model,
+            max_model_len=1024,
+            enforce_eager=True,
+    ) as runner:
+        vllm_eager_outputs = runner.model.generate(prompts, sampling_params)
+
+    vllm_fullgraph_outputs_list = []
+    for output in vllm_fullgraph_outputs:
+        vllm_fullgraph_outputs_list.append(
+            (output.outputs[0].index, output.outputs[0].text))
+
+    vllm_eager_outputs_list = []
+    for output in vllm_eager_outputs:
+        vllm_eager_outputs_list.append(
+            (output.outputs[0].index, output.outputs[0].text))
+
+    check_outputs_equal(
+        outputs_0_lst=vllm_eager_outputs_list,
+        outputs_1_lst=vllm_fullgraph_outputs_list,
+        name_0="vllm_eager_outputs",
+        name_1="vllm_fullgraph_outputs",
+    )