[CI] Add accuracy ci for DP and EP and TP

Signed-off-by: hfadzxy <starmoon_zhang@163.com>

Author: zhangxinyuehfad

.github/workflows/vllm_ascend_test_long_term.yaml

@@ -102,5 +102,5 @@ jobs:
             pytest -sv tests/long_term/spec_decode --ignore=tests/long_term/spec_decode/e2e/test_mtp_correctness.py --ignore=tests/long_term/spec_decode/e2e/test_v1_spec_decode.py --ignore=tests/long_term/spec_decode/e2e/test_v1_mtp_correctness.py
             pytest -sv tests/long_term/test_accuracy.py
           else
-            VLLM_USE_MODELSCOPE=True pytest -sv tests/long_term/test_deepseek_v2_lite_tp2_accuracy.py
+            pytest -sv tests/long_term/test_accuracy.py
           fi

tests/long_term/test_accuracy.py

@@ -16,74 +16,133 @@
 # This file is a part of the vllm-ascend project.
 # Adapted from vllm-project/blob/main/tests/entrypoints/llm/test_accuracy.py
 #
-
 import gc
 import multiprocessing
+import os
 import sys
+import time
+from vllm import SamplingParams
 from multiprocessing import Queue
 
+import signal
+import requests
+import subprocess
 import lm_eval
 import pytest
 import torch
 
+SERVER_HOST = "127.0.0.1"
+SERVER_PORT = 8000
+HEALTH_URL = f"http://{SERVER_HOST}:{SERVER_PORT}/health"
+COMPLETIONS_URL = f"http://{SERVER_HOST}:{SERVER_PORT}/v1/completions"
+
 # pre-trained model path on Hugging Face.
-MODEL_NAME = ["Qwen/Qwen2.5-0.5B-Instruct", "Qwen/Qwen2.5-VL-3B-Instruct"]
+# Qwen/Qwen2.5-0.5B-Instruct: accuracy test for unimodal model.
+# Qwen/Qwen2.5-VL-3B-Instruct: accuracy test for multimodal model.
+# Qwen/Qwen3-30B-A3B: accuracy test for EP.
+# deepseek-ai/DeepSeek-V2-Lite: accuracy test for TP.
+MODEL_NAME = [
+    "Qwen/Qwen2.5-0.5B-Instruct", "Qwen/Qwen2.5-VL-3B-Instruct",
+    "Qwen/Qwen3-30B-A3B", "deepseek-ai/DeepSeek-V2-Lite"
+]
+# Qwen/Qwen2.5-7B-Instruct: accuracy test for DP
+MODEL_NAME_DP = ["Qwen/Qwen2.5-0.5B-Instruct"]
+
 # Benchmark configuration mapping models to evaluation tasks:
 # - Text model: GSM8K (grade school math reasoning)
 # - Vision-language model: MMMU Art & Design validation (multimodal understanding)
 TASK = {
     "Qwen/Qwen2.5-0.5B-Instruct": "gsm8k",
-    "Qwen/Qwen2.5-VL-3B-Instruct": "mmmu_val_art_and_design"
+    "Qwen/Qwen2.5-VL-3B-Instruct": "mmmu_val_art_and_design",
+    "Qwen/Qwen3-30B-A3B": "gsm8k",
+    "deepseek-ai/DeepSeek-V2-Lite": "gsm8k"
 }
 # Answer validation requiring format consistency.
 FILTER = {
     "Qwen/Qwen2.5-0.5B-Instruct": "exact_match,strict-match",
-    "Qwen/Qwen2.5-VL-3B-Instruct": "acc,none"
+    "Qwen/Qwen2.5-VL-3B-Instruct": "acc,none",
+    "Qwen/Qwen3-30B-A3B": "exact_match,strict-match",
+    "deepseek-ai/DeepSeek-V2-Lite": "exact_match,strict-match"
 }
 # 3% relative tolerance for numerical accuracy.
 RTOL = 0.03
 # Baseline accuracy after VLLM optimization.
 EXPECTED_VALUE = {
     "Qwen/Qwen2.5-0.5B-Instruct": 0.316,
-    "Qwen/Qwen2.5-VL-3B-Instruct": 0.541
+    "Qwen/Qwen2.5-VL-3B-Instruct": 0.541,
+    "Qwen/Qwen3-30B-A3B": 0.888,
+    "deepseek-ai/DeepSeek-V2-Lite": 0.376
 }
 # Maximum context length configuration for each model.
 MAX_MODEL_LEN = {
     "Qwen/Qwen2.5-0.5B-Instruct": 4096,
-    "Qwen/Qwen2.5-VL-3B-Instruct": 8192
+    "Qwen/Qwen2.5-VL-3B-Instruct": 8192,
+    "Qwen/Qwen3-30B-A3B": 4096,
+    "deepseek-ai/DeepSeek-V2-Lite": 4096
 }
 # Model types distinguishing text-only and vision-language models.
 MODEL_TYPE = {
     "Qwen/Qwen2.5-0.5B-Instruct": "vllm",
-    "Qwen/Qwen2.5-VL-3B-Instruct": "vllm-vlm"
+    "Qwen/Qwen2.5-VL-3B-Instruct": "vllm-vlm",
+    "Qwen/Qwen3-30B-A3B": "vllm",
+    "deepseek-ai/DeepSeek-V2-Lite": "vllm"
 }
 # wrap prompts in a chat-style template.
-APPLY_CHAT_TEMPLATE = {"vllm": False, "vllm-vlm": True}
+APPLY_CHAT_TEMPLATE = {
+    "Qwen/Qwen2.5-0.5B-Instruct": False,
+    "Qwen/Qwen2.5-VL-3B-Instruct": True,
+    "Qwen/Qwen3-30B-A3B": False,
+    "deepseek-ai/DeepSeek-V2-Lite": False
+}
 # Few-shot examples handling as multi-turn dialogues.
-FEWSHOT_AS_MULTITURN = {"vllm": False, "vllm-vlm": True}
+FEWSHOT_AS_MULTITURN = {
+    "Qwen/Qwen2.5-0.5B-Instruct": False,
+    "Qwen/Qwen2.5-VL-3B-Instruct": True,
+    "Qwen/Qwen3-30B-A3B": False,
+    "deepseek-ai/DeepSeek-V2-Lite": False
+}
+MORE_ARGS = {
+    "Qwen/Qwen2.5-0.5B-Instruct":
+    None,
+    "Qwen/Qwen2.5-VL-3B-Instruct":
+    None,
+    "Qwen/Qwen3-30B-A3B":
+    "tensor_parallel_size=4,enable_expert_parallel=True,enforce_eager=True",
+    "deepseek-ai/DeepSeek-V2-Lite":
+    "tensor_parallel_size=4,trust_remote_code=True,enforce_eager=True"
+}
+
+multiprocessing.set_start_method("spawn", force=True)
 
 
-def run_test(queue, model, max_model_len, model_type):
+def get_available_npu_count():
+    return torch.npu.device_count()
+
+
+def run_test(queue, model, max_model_len, model_type, more_args):
     try:
         if model_type == "vllm-vlm":
             model_args = (f"pretrained={model},max_model_len={max_model_len},"
                           "dtype=auto,max_images=2")
         else:
             model_args = (f"pretrained={model},max_model_len={max_model_len},"
                           "dtype=auto")
+        if more_args is not None:
+            model_args = f"{model_args},{more_args}"
         results = lm_eval.simple_evaluate(
             model=model_type,
             model_args=model_args,
             tasks=TASK[model],
             batch_size="auto",
-            apply_chat_template=APPLY_CHAT_TEMPLATE[model_type],
-            fewshot_as_multiturn=FEWSHOT_AS_MULTITURN[model_type],
+            apply_chat_template=APPLY_CHAT_TEMPLATE[model],
+            fewshot_as_multiturn=FEWSHOT_AS_MULTITURN[model],
         )
         result = results["results"][TASK[model]][FILTER[model]]
         print("result:", result)
         queue.put(result)
     except Exception as e:
-        queue.put(e)
+        error_msg = f"{type(e).__name__}: {str(e)}"
+        queue.put(error_msg)
         sys.exit(1)
     finally:
         gc.collect()
@@ -93,19 +152,89 @@ def run_test(queue, model, max_model_len, model_type):
 @pytest.mark.parametrize("model", MODEL_NAME)
 @pytest.mark.parametrize("VLLM_USE_V1", ["0", "1"])
 def test_lm_eval_accuracy(monkeypatch: pytest.MonkeyPatch, model, VLLM_USE_V1):
+    npu_count = get_available_npu_count()
     if model == "Qwen/Qwen2.5-VL-3B-Instruct" and VLLM_USE_V1 == "1":
         pytest.skip(
             "Qwen2.5-VL-3B-Instruct is not supported when VLLM_USE_V1=1")
-    with monkeypatch.context() as m:
-        m.setenv("VLLM_USE_V1", VLLM_USE_V1)
+    if (model == "Qwen/Qwen2.5-VL-3B-Instruct"
+            or model == "Qwen/Qwen2.5-0.5B-Instruct") and npu_count != 1:
+        pytest.skip(
+            "test accuarcy for Qwen2.5-0.5B-Instruct and Qwen2.5-VL-3B-Instruct when tp != 1"
+        )
+    if (model == "Qwen/Qwen3-30B-A3B"
+            or model == "deepseek-ai/DeepSeek-V2-Lite") and (
+                os.getenv("VLLM_USE_V1") != "1" or npu_count != 4):
+        pytest.skip(
+            "test ep accuracy for Qwen/Qwen3-30B-A3B when VLLM_USE_V1=1 and tp=4"
+        )
+    with monkeypatch.context():
         result_queue: Queue[float] = multiprocessing.Queue()
         p = multiprocessing.Process(target=run_test,
                                     args=(result_queue, model,
                                           MAX_MODEL_LEN[model],
-                                          MODEL_TYPE[model]))
+                                          MODEL_TYPE[model], MORE_ARGS[model]))
         p.start()
         p.join()
         result = result_queue.get()
         print(result)
         assert (EXPECTED_VALUE[model] - RTOL < result < EXPECTED_VALUE[model] + RTOL), \
             f"Expected: {EXPECTED_VALUE[model]}±{RTOL} | Measured: {result}"
+
+
+@pytest.mark.parametrize("max_tokens", [10])
+@pytest.mark.parametrize("model", MODEL_NAME_DP)
+def test_lm_eval_accuracy_dp(model, max_tokens):
+    npu_count = get_available_npu_count()
+    if (model != "Qwen/Qwen2.5-0.5B-Instruct"
+            or os.getenv("VLLM_USE_V1") != "1" or npu_count != 4):
+        pytest.skip(
+            "test accuracy for DP when model is Qwen2.5-0.5B-Instruct and engine is V1"
+        )
+
+    log_file = open("accuracy.log", "a")
+    cmd = [
+        "vllm", "serve", model, "--tensor_parallel_size", "2",
+        "--data_parallel_size", "2"
+    ]
+    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:
+            pytest.fail(
+                f"vLLM serve did not become healthy after 300s: {HEALTH_URL}")
+
+        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 north china, it has many famous attractions"
+        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()