[Bugfix] Fix deepseek V0 percision issue and add acc ci for it

Signed-off-by: MengqingCao <[email protected]>

Author: MengqingCao

.github/workflows/vllm_ascend_test_long_term.yaml

@@ -41,9 +41,19 @@ jobs:
     strategy:
       max-parallel: 2
       matrix:
+        os: [linux-arm64-npu-1, linux-arm64-npu-4]
         vllm_version: [main, v0.9.0]
+    concurrency:
+      group: >
+        ${{
+        matrix.os == 'linux-arm64-npu-4'
+          && github.event.pull_request.number
+          && format('pr-{0}-limit-npu-4', github.event.pull_request.number)
+        || format('job-{0}-{1}-{2}', matrix.os, matrix.vllm_version, github.event.pull_request.number)
+        }}
+      cancel-in-progress: false
     name: vLLM Ascend long term test
-    runs-on: linux-arm64-npu-1
+    runs-on: ${{ matrix.os }}
     container:
       # TODO(yikun): Remove m.daocloud.io prefix when infra proxy ready
       image: m.daocloud.io/quay.io/ascend/cann:8.1.rc1-910b-ubuntu22.04-py3.10
@@ -92,8 +102,13 @@ jobs:
 
       - name: Run vllm-project/vllm-ascend long term test
         run: |
-          # spec decode test
-          VLLM_USE_MODELSCOPE=True pytest -sv tests/long_term/spec_decode/e2e/test_v1_mtp_correctness.py
-          VLLM_USE_MODELSCOPE=true pytest -sv tests/long_term/spec_decode/e2e/test_v1_spec_decode.py
-          VLLM_USE_MODELSCOPE=True pytest -sv tests/long_term/spec_decode/e2e/test_mtp_correctness.py  # it needs a clean process
-          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
+          if [[ "${{ matrix.os }}" == "linux-arm64-npu-1" ]]; then
+            # spec decode test
+            VLLM_USE_MODELSCOPE=True pytest -sv tests/long_term/spec_decode/e2e/test_v1_mtp_correctness.py
+            VLLM_USE_MODELSCOPE=True pytest -sv tests/long_term/spec_decode/e2e/test_v1_spec_decode.py
+            VLLM_USE_MODELSCOPE=True pytest -sv tests/long_term/spec_decode/e2e/test_mtp_correctness.py  # it needs a clean process
+            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
+          fi

tests/conftest.py

@@ -354,4 +354,4 @@ def prompt_template(request):
 
 @pytest.fixture(scope="session")
 def ilama_lora_files():
-    return snapshot_download(repo_id="jeeejeee/ilama-text2sql-spider")
+    return snapshot_download(repo_id="jeeejeee/ilama-text2sql-spider")

tests/long_term/test_deepseek_v2_lite_tp2_accuracy.py

@@ -0,0 +1,71 @@
+#
+# Copyright (c) 2025 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.
+# Adapted from vllm-project/blob/main/tests/entrypoints/llm/test_accuracy.py
+#
+
+import gc
+import multiprocessing
+from multiprocessing import Queue
+
+import lm_eval
+import pytest
+import torch
+
+# pre-trained model path on Hugging Face.
+MODELS = ["deepseek-ai/DeepSeek-V2-Lite"]
+# Math reasoning benchmark (Grade School Math 8K).
+TASK = "gsm8k"
+# Answer validation requiring format consistency.
+FILTER = "exact_match,strict-match"
+# 3% relative tolerance for numerical accuracy.
+RTOL = 0.03
+# Baseline accuracy after VLLM optimization.
+EXPECTED_VALUE = 0.316
+
+
+def run_test(model_name, queue, more_args=None):
+    model_args = f"pretrained={model_name},max_model_len=4096,trust_remote_code=True,tensor_parallel_size=4"
+    if more_args is not None:
+        model_args = f"{model_args},{more_args}"
+    results = lm_eval.simple_evaluate(
+        model="vllm",
+        model_args=model_args,
+        tasks=TASK,
+        batch_size="auto",
+    )
+    result = results["results"][TASK][FILTER]
+    print(100 * "*", "\nThe accuracy test result:", result)
+    queue.put(result)
+    del results
+    torch.npu.empty_cache()
+    gc.collect()
+
+
+@pytest.mark.parametrize("model", MODELS)
+def test_lm_eval_accuracy(model, monkeypatch: pytest.MonkeyPatch):
+    with monkeypatch.context():
+        result_queue: Queue[float] = multiprocessing.Queue()
+        p = multiprocessing.Process(target=run_test,
+                                    args=(
+                                        model,
+                                        result_queue,
+                                    ))
+        p.start()
+        p.join()
+        result = result_queue.get()
+        assert (EXPECTED_VALUE - RTOL < result < EXPECTED_VALUE + RTOL), \
+            f"Expected: {EXPECTED_VALUE}±{RTOL} | Measured: {result}"

vllm_ascend/ops/fused_moe.py

@@ -329,6 +329,7 @@ def fused_experts(
     num_experts = w1.shape[0]
     dtype = hidden_states.dtype
     device = hidden_states.device
+    topk_weights = topk_weights.to(dtype)
     # assert dtype in [torch.float32, torch.float16, torch.bfloat16
     #                  ], "Only float32, float16, and bfloat16 are supported"
 

vllm_ascend/quantization/w8a8_dynamic.py

@@ -340,6 +340,7 @@ def fused_experts(hidden_states: torch.Tensor,
     num_experts = w1.shape[0]
     dtype = hidden_states.dtype
     device = hidden_states.device
+    topk_weights = topk_weights.to(dtype)
 
     if expert_map is not None:
         # Generate token indices and flatten