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

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

Author: MengqingCao

tests/conftest.py

@@ -354,4 +354,12 @@ 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")
+
+
+@pytest.fixture
+def enable_modelscope_env():
+    import os
+    from unittest.mock import patch
+    with patch.dict(os.environ, {"VLLM_USE_MODEL_SCOPE": "TRUE"}):
+        yield

tests/long_term/spec_decode/e2e/test_mtp_correctness.py

@@ -93,11 +93,17 @@
 ])
 @pytest.mark.parametrize("batch_size", [1, 32])
 @pytest.mark.parametrize("seed", [1])
-def test_mtp_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
-                                    per_test_common_llm_kwargs,
-                                    baseline_llm_kwargs, test_llm_kwargs,
-                                    batch_size: int, output_len: int,
-                                    seed: int):
+def test_mtp_e2e_greedy_correctness(
+    enable_modelscope_env,
+    vllm_runner,
+    common_llm_kwargs,
+    per_test_common_llm_kwargs,
+    baseline_llm_kwargs,
+    test_llm_kwargs,
+    batch_size: int,
+    output_len: int,
+    seed: int,
+):
 
     run_equality_correctness_test(vllm_runner, common_llm_kwargs,
                                   per_test_common_llm_kwargs,
@@ -138,12 +144,17 @@ def test_mtp_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
 ])
 @pytest.mark.parametrize("batch_size", [1, 32])
 @pytest.mark.parametrize("seed", [1])
-def test_mtp_e2e_quant_greedy_correctness(vllm_runner, common_llm_kwargs,
-                                          per_test_common_llm_kwargs,
-                                          baseline_llm_kwargs, test_llm_kwargs,
-                                          batch_size: int, output_len: int,
-                                          seed: int):
-
+def test_mtp_e2e_quant_greedy_correctness(
+    enable_modelscope_env,
+    vllm_runner,
+    common_llm_kwargs,
+    per_test_common_llm_kwargs,
+    baseline_llm_kwargs,
+    test_llm_kwargs,
+    batch_size: int,
+    output_len: int,
+    seed: int,
+):
     run_equality_correctness_test(vllm_runner, common_llm_kwargs,
                                   per_test_common_llm_kwargs,
                                   baseline_llm_kwargs, test_llm_kwargs,
@@ -192,12 +203,18 @@ def test_mtp_e2e_quant_greedy_correctness(vllm_runner, common_llm_kwargs,
 @pytest.mark.parametrize("batch_size", [8])
 @pytest.mark.parametrize("seed", [1])
 @pytest.mark.parametrize("logprobs", [1, 6])
-def test_mtp_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
-                                 per_test_common_llm_kwargs,
-                                 baseline_llm_kwargs, test_llm_kwargs,
-                                 batch_size: int, output_len: int, seed: int,
-                                 logprobs: int):
-
+def test_mtp_e2e_greedy_logprobs(
+    enable_modelscope_env,
+    vllm_runner,
+    common_llm_kwargs,
+    per_test_common_llm_kwargs,
+    baseline_llm_kwargs,
+    test_llm_kwargs,
+    batch_size: int,
+    output_len: int,
+    seed: int,
+    logprobs: int,
+):
     run_equality_correctness_test(
         vllm_runner,
         common_llm_kwargs,
@@ -246,9 +263,16 @@ def test_mtp_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
 @pytest.mark.parametrize("batch_size", [1, 32])
 @pytest.mark.parametrize("seed", [1])
 def test_mtp_e2e_greedy_correctness_torchair_graph(
-        vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
-        baseline_llm_kwargs, test_llm_kwargs, batch_size: int, output_len: int,
-        seed: int):
+    enable_modelscope_env,
+    vllm_runner,
+    common_llm_kwargs,
+    per_test_common_llm_kwargs,
+    baseline_llm_kwargs,
+    test_llm_kwargs,
+    batch_size: int,
+    output_len: int,
+    seed: int,
+):
     """Verify greedy equality with torchair graph enabled and different
     batch sizes using bfloat16 weights."""
     run_equality_correctness_test(vllm_runner, common_llm_kwargs,
@@ -290,9 +314,16 @@ def test_mtp_e2e_greedy_correctness_torchair_graph(
 @pytest.mark.parametrize("batch_size", [1, 32])
 @pytest.mark.parametrize("seed", [1])
 def test_mtp_e2e_quant_greedy_correctness_torchair_graph(
-        vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
-        baseline_llm_kwargs, test_llm_kwargs, batch_size: int, output_len: int,
-        seed: int):
+    enable_modelscope_env,
+    vllm_runner,
+    common_llm_kwargs,
+    per_test_common_llm_kwargs,
+    baseline_llm_kwargs,
+    test_llm_kwargs,
+    batch_size: int,
+    output_len: int,
+    seed: int,
+):
     """Verify greedy equality with torchair graph enabled and different
     batch sizes using quant weights."""
     run_equality_correctness_test(vllm_runner, common_llm_kwargs,
@@ -341,9 +372,16 @@ def test_mtp_e2e_quant_greedy_correctness_torchair_graph(
 @pytest.mark.parametrize("batch_size", [4])
 @pytest.mark.parametrize("seed", [1])
 def test_mtp_e2e_greedy_correctness_with_preemption(
-        vllm_runner, common_llm_kwargs, per_test_common_llm_kwargs,
-        baseline_llm_kwargs, test_llm_kwargs, batch_size: int, output_len: int,
-        seed: int):
+    enable_modelscope_env,
+    vllm_runner,
+    common_llm_kwargs,
+    per_test_common_llm_kwargs,
+    baseline_llm_kwargs,
+    test_llm_kwargs,
+    batch_size: int,
+    output_len: int,
+    seed: int,
+):
     """Verify greedy equality, even when some sequences are preempted mid-
     generation.
     """
@@ -391,10 +429,17 @@ def test_mtp_e2e_greedy_correctness_with_preemption(
         32,
     ])
 @pytest.mark.parametrize("seed", [1])
-def test_mtp_different_k(vllm_runner, common_llm_kwargs,
-                         per_test_common_llm_kwargs, baseline_llm_kwargs,
-                         test_llm_kwargs, batch_size: int, output_len: int,
-                         seed: int):
+def test_mtp_different_k(
+    enable_modelscope_env,
+    vllm_runner,
+    common_llm_kwargs,
+    per_test_common_llm_kwargs,
+    baseline_llm_kwargs,
+    test_llm_kwargs,
+    batch_size: int,
+    output_len: int,
+    seed: int,
+):
     """Verify that mtp speculative decoding produces exact equality
     to without spec decode with different values of num_speculative_tokens.
     """
@@ -437,10 +482,17 @@ def test_mtp_different_k(vllm_runner, common_llm_kwargs,
         32,
     ])
 @pytest.mark.parametrize("seed", [1])
-def test_mtp_disable_queue(vllm_runner, common_llm_kwargs,
-                           per_test_common_llm_kwargs, baseline_llm_kwargs,
-                           test_llm_kwargs, batch_size: int, output_len: int,
-                           seed: int):
+def test_mtp_disable_queue(
+    enable_modelscope_env,
+    vllm_runner,
+    common_llm_kwargs,
+    per_test_common_llm_kwargs,
+    baseline_llm_kwargs,
+    test_llm_kwargs,
+    batch_size: int,
+    output_len: int,
+    seed: int,
+):
     """Verify that mtp speculative decoding produces exact equality
     to without spec decode when speculation is disabled for large
     batch sizes.

tests/long_term/spec_decode/e2e/test_v1_spec_decode.py

@@ -60,6 +60,7 @@ def eagle3_model_name():
 
 
 def test_ngram_correctness(
+    enable_modelscope_env,
     monkeypatch: pytest.MonkeyPatch,
     test_prompts: list[list[dict[str, Any]]],
     sampling_config: SamplingParams,
@@ -71,8 +72,10 @@ def test_ngram_correctness(
     '''
     with monkeypatch.context() as m:
         m.setenv("VLLM_USE_V1", "1")
+        m.setenv("VLLM_USE_MODELSCOPE", "True")
 
-        ref_llm = LLM(model=model_name, max_model_len=1024)
+        ref_llm = LLM(model="LLM-Research/Meta-Llama-3.1-8B-Instruct",
+                      max_model_len=1024)
         ref_outputs = ref_llm.chat(test_prompts, sampling_config)
         del ref_llm
 
@@ -105,6 +108,7 @@ def test_ngram_correctness(
 
 @pytest.mark.parametrize("use_eagle3", [False, True], ids=["eagle", "eagle3"])
 def test_eagle_correctness(
+    enable_modelscope_env,
     monkeypatch: pytest.MonkeyPatch,
     test_prompts: list[list[dict[str, Any]]],
     sampling_config: SamplingParams,

tests/multicard/test_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}"

tests/singlecard/test_offline_inference.py

@@ -41,7 +41,12 @@
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("dtype", ["half", "float16"])
 @pytest.mark.parametrize("max_tokens", [5])
-def test_models(model: str, dtype: str, max_tokens: int) -> None:
+def test_models(
+    enable_modelscope_env,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
     # 5042 tokens for gemma2
     # gemma2 has alternating sliding window size of 4096
     # we need a prompt with more than 4096 tokens to test the sliding window
@@ -60,7 +65,12 @@ def test_models(model: str, dtype: str, max_tokens: int) -> None:
 @pytest.mark.parametrize("model", MULTIMODALITY_MODELS)
 @pytest.mark.skipif(os.getenv("VLLM_USE_V1") == "1",
                     reason="qwen2.5_vl is not supported on v1")
-def test_multimodal(model, prompt_template, vllm_runner):
+def test_multimodal(
+    enable_modelscope_env,
+    model,
+    prompt_template,
+    vllm_runner,
+):
     image = ImageAsset("cherry_blossom") \
         .pil_image.convert("RGB")
     img_questions = [

vllm_ascend/ops/fused_moe.py

@@ -337,6 +337,7 @@ 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

@@ -342,6 +342,7 @@ 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