[Dist][EP] Remove ETP/EP maintained in vllm-ascend (#1681)

### What this PR does / why we need it?
Remove ETP/EP maintained in branch main. We drop this as there is no
relevant scenarios to use ETP now, and we may subsequently advocate
implementing expert tensor parallelism in vLLM to support scenarios
where the expert is needed to be sliced

This is a part of #1422 backport.

Fixes #1396
#1154

### Does this PR introduce _any_ user-facing change?
We'll not maintain etp/ep in vllm-ascend anymore, and use the tp/ep in
vllm instead.

### How was this patch tested?
CI passed with new added and existing test.


- vLLM version: v0.9.2
- vLLM main:
vllm-project/vllm@fe8a2c5

Signed-off-by: MengqingCao <cmq0113@163.com>

Author: MengqingCao

docs/source/tutorials/multi_npu_moge.md

@@ -48,6 +48,7 @@ Run the following script to start the vLLM server on Multi-NPU:
 ```bash
 vllm serve /path/to/pangu-pro-moe-model \
 --tensor-parallel-size 4 \
+--enable-expert-parallel \
 --trust-remote-code \
 --enforce-eager
 ```
@@ -145,6 +146,7 @@ if __name__ == "__main__":
 
     llm = LLM(model="/path/to/pangu-pro-moe-model",
             tensor_parallel_size=4,
+            enable_expert_parallel=True,
             distributed_executor_backend="mp",
             max_model_len=1024,
             trust_remote_code=True,

docs/source/user_guide/configuration/additional_config.md

@@ -28,7 +28,6 @@ The following table lists the additional configuration options available in vLLM
 |-------------------------------| ---- |------|-----------------------------------------------------------------------------------------------|
 | `torchair_graph_config`       | dict | `{}` | The config options for torchair graph mode                                                    |
 | `ascend_scheduler_config`     | dict | `{}` | The config options for ascend scheduler                                                       |
-| `expert_tensor_parallel_size` | str  | `0`  | Expert tensor parallel size the model to use.                                                 |
 | `refresh`                     | bool | `false` | Whether to refresh global ascend config content. This value is usually used by rlhf or ut/e2e test case.     |
 | `expert_map_path`             | str  | `None` | When using expert load balancing for the MOE model, an expert map path needs to be passed in. |
 | `chunked_prefill_for_mla`     | bool | `False` | Whether to enable the fused operator-like chunked_prefill. |
@@ -75,7 +74,6 @@ An example of additional configuration is as follows:
         "enabled": True,
         "enable_chunked_prefill": True,
     },
-    "expert_tensor_parallel_size": 1,
     "refresh": False,
 }
 ```

examples/run_dp_attention_etp16.sh

@@ -36,7 +36,7 @@
 
 # pre-trained model path on Hugging Face.
 # Qwen/Qwen2.5-0.5B-Instruct: accuracy test for DP.
-# Qwen/Qwen3-30B-A3B: accuracy test for EP and ETP.
+# Qwen/Qwen3-30B-A3B: accuracy test for EP.
 # deepseek-ai/DeepSeek-V2-Lite: accuracy test for TP.
 MODEL_NAME = ["Qwen/Qwen3-30B-A3B", "deepseek-ai/DeepSeek-V2-Lite"]
 
@@ -200,62 +200,3 @@ def test_lm_eval_accuracy_dp(model, max_tokens):
         except subprocess.TimeoutExpired:
             server_proc.kill()
             server_proc.wait()
-
-
-@pytest.mark.parametrize("max_tokens", [10])
-@pytest.mark.parametrize("model", ["Qwen/Qwen3-30B-A3B"])
-def test_lm_eval_accuracy_etp(model, max_tokens):
-    log_file = open("accuracy_etp.log", "a+")
-    cmd = [
-        "vllm", "serve", model, "--max_model_len", "4096",
-        "--tensor_parallel_size", "4", "--enforce_eager",
-        "--enable_expert_parallel", "--additional_config",
-        '{"expert_tensor_parallel_size": "4"}'
-    ]
-    server_proc = subprocess.Popen(cmd,
-                                   stdout=log_file,
-                                   stderr=subprocess.DEVNULL)
-
-    try:
-        for _ in range(300):
-            try:
-                r = requests.get(HEALTH_URL, timeout=1)
-                if r.status_code == 200:
-                    break
-            except requests.exceptions.RequestException:
-                pass
-            time.sleep(1)
-        else:
-            log_file.flush()
-            log_file.seek(0)
-            log_content = log_file.read()
-            pytest.fail(
-                f"vLLM serve did not become healthy after 300s: {HEALTH_URL}\n"
-                f"==== vLLM Serve Log Start ===\n{log_content}\n==== vLLM Serve Log End ==="
-            )
-
-        prompt = "bejing is a"
-        payload = {
-            "prompt": prompt,
-            "max_tokens": max_tokens,
-            "sampling_params": {
-                "temperature": 0.0,
-                "top_p": 1.0,
-                "seed": 123
-            }
-        }
-        resp = requests.post(COMPLETIONS_URL, json=payload, timeout=30)
-        resp.raise_for_status()
-        data = resp.json()
-
-        generated = data["choices"][0]["text"].strip()
-        expected = "city in china. it is the capital city of"
-        assert generated == expected, f"Expected `{expected}`, got `{generated}`"
-
-    finally:
-        server_proc.send_signal(signal.SIGINT)
-        try:
-            server_proc.wait(timeout=10)
-        except subprocess.TimeoutExpired:
-            server_proc.kill()
-            server_proc.wait()

examples/run_dp_attention_etp16_benmark.sh

@@ -0,0 +1,30 @@
+import pytest
+
+from tests.e2e.conftest import VllmRunner
+from tests.e2e.model_utils import check_outputs_equal
+
+
+@pytest.mark.parametrize("model_name", ["deepseek-ai/DeepSeek-V2-Lite-Chat"])
+def test_e2e_ep_correctness(model_name):
+    example_prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
+    max_tokens = 5
+
+    with VllmRunner(model_name, tensor_parallel_size=2) as vllm_model:
+        tp_output = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    with VllmRunner(model_name,
+                    tensor_parallel_size=2,
+                    enable_expert_parallel=True) as vllm_model:
+        ep_output = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=ep_output,
+        outputs_1_lst=tp_output,
+        name_0="ep_output",
+        name_1="tp_output",
+    )

tests/e2e/long_term/accuracy/accuracy_multicard.py

@@ -50,7 +50,6 @@ def test_generate_with_allgather():
                             "enabled": True,
                             "chunked_prefill_enabled": False,
                         },
-                        "expert_tensor_parallel_size": 1
                     }) as vllm_model:
         vllm_model.generate(example_prompts, sampling_params)
 
@@ -74,6 +73,5 @@ def test_generate_with_alltoall():
                             "enabled": True,
                             "chunked_prefill_enabled": False,
                         },
-                        "expert_tensor_parallel_size": 1
                     }) as vllm_model:
         vllm_model.generate(example_prompts, sampling_params)

tests/e2e/multicard/test_expert_parallel.py

@@ -123,6 +123,7 @@ def _pangu_torchair_test_fixture(
             distributed_executor_backend="mp",
             enforce_eager=False,
             additional_config=additional_config,
+            enable_expert_parallel=True,
     ) as vllm_model:
         # use greedy sampler to make sure the generated results are fix
         vllm_output = vllm_model.generate_greedy(example_prompts, 5)