Add models test and add serval new models yaml (#3394)

### What this PR does / why we need it?
This PR added Add accuracy CI for servals new models
- `ascend test / accuracy` is for PR triggered check popluar models
accuracy
- `ascedn test / models` is for accuracy report, full models test,
nightly model test
- Add Qwen2-Audio-7B-Instruct, Qwen2-VL-7B-Instruct, Qwen3-8B,
Qwen3-VL-30B-A3B-Instruct

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
CI passed

Closes: https://github.com/vllm-project/vllm-ascend/pull/2330
Closes: https://github.com/vllm-project/vllm-ascend/pull/3362


- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

Signed-off-by: hfadzxy <starmoon_zhang@163.com>
Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
Co-authored-by: hfadzxy <starmoon_zhang@163.com>

tests/e2e/models/configs/Qwen2-Audio-7B-Instruct.yaml

@@ -0,0 +1,10 @@
+model_name: "Qwen/Qwen2-Audio-7B-Instruct"
+hardware: "Atlas A2 Series"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.44
+  - name: "exact_match,flexible-extract"
+    value: 0.45
+num_fewshot: 5

tests/e2e/models/configs/Qwen2-VL-7B-Instruct.yaml

@@ -0,0 +1,10 @@
+model_name: "Qwen/Qwen2-VL-7B-Instruct"
+hardware: "Atlas A2 Series"
+model: "vllm-vlm"
+tasks:
+- name: "mmmu_val"
+  metrics:
+  - name: "acc,none"
+    value: 0.50
+max_model_len: 8192
+gpu_memory_utilization: 0.7

tests/e2e/models/configs/Qwen3-8B.yaml

@@ -0,0 +1,11 @@
+model_name: "Qwen/Qwen3-8B"
+hardware: "Atlas A2 Series"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.765
+  - name: "exact_match,flexible-extract"
+    value: 0.81
+num_fewshot: 5
+enable_thinking: False

tests/e2e/models/configs/Qwen3-VL-30B-A3B-Instruct.yaml

@@ -0,0 +1,12 @@
+model_name: "Qwen/Qwen3-VL-30B-A3B-Instruct"
+hardware: "Atlas A2 Series"
+model: "vllm-vlm"
+tasks:
+- name: "mmmu_val"
+  metrics:
+  - name: "acc,none"
+    value: 0.58
+max_model_len: 8192
+tensor_parallel_size: 2
+gpu_memory_utilization: 0.7
+enable_expert_parallel: True

tests/e2e/models/configs/accuracy.txt

@@ -1,4 +1,8 @@
 DeepSeek-V2-Lite.yaml
-Qwen3-8B-Base.yaml
 Qwen2.5-VL-7B-Instruct.yaml
-Qwen3-30B-A3B.yaml
+Qwen3-30B-A3B.yaml
+Qwen3-8B.yaml
+Qwen2-7B.yaml
+Qwen2-VL-7B-Instruct.yaml
+Qwen2-Audio-7B-Instruct.yaml
+Qwen3-VL-30B-A3B-Instruct.yaml

tests/e2e/models/test_lm_eval_correctness.py

@@ -7,7 +7,7 @@ import pytest
 import yaml
 from jinja2 import Environment, FileSystemLoader
 
-RTOL = 0.03
+RTOL = 0.05
 TEST_DIR = os.path.dirname(__file__)
 
 
@@ -48,7 +48,7 @@ def build_model_args(eval_config, tp_size):
     }
     for s in [
             "max_images", "gpu_memory_utilization", "enable_expert_parallel",
-            "tensor_parallel_size", "enforce_eager"
+            "tensor_parallel_size", "enforce_eager", "enable_thinking"
     ]:
         val = eval_config.get(s, None)
         if val is not None: