remove unsupported tests from kernels and add pytest html

requirements-test.txt

@@ -5,6 +5,8 @@ pytest-forked
 pytest-asyncio
 pytest-rerunfailures
 pytest-shard
+pytest-html
+pytest-timeout
 
 # testing utils
 awscli

tests/conftest.py

@@ -8,6 +8,8 @@ from collections import UserList
 from enum import Enum
 from typing import (Any, Callable, Dict, List, Optional, Tuple, Type,
                     TypedDict, TypeVar, Union)
+import pytest
+import pytest_html
 
 import numpy as np
 import pytest
@@ -898,3 +900,21 @@ def dummy_opt_path():
         with open(json_path, "w") as f:
             json.dump(config, f)
     return _dummy_path
+
+# 定义一个 pytest 钩子，在测试后生成报告
+@pytest.hookimpl(tryfirst=True, hookwrapper=True)
+def pytest_runtest_makereport(item, call):
+    # 获取测试结果
+    outcome = yield
+    result = outcome.get_result()
+
+    # 如果测试失败并且有浏览器实例，添加截图
+    if result.when == "call" and result.failed:
+        if hasattr(item, "funcargs") and "browser" in item.funcargs:
+            browser = item.funcargs["browser"]
+            screenshot_path = "screenshot.png"  # 设置截图路径
+            browser.save_screenshot(screenshot_path)
+
+            # 如果测试结果有 extra 属性，则添加截图
+            if hasattr(result, "extra"):
+                result.extra.append(pytest_html.extras.image(screenshot_path))

tests/kernels/test_attention_selector.py

@@ -12,7 +12,7 @@ from vllm.utils import STR_FLASH_ATTN_VAL, STR_INVALID_VAL
 #     "name", ["TORCH_SDPA", "ROCM_FLASH", "XFORMERS", "FLASHINFER", "OPENVINO"])
 # @pytest.mark.parametrize("device", ["cpu", "openvino", "hip", "cuda"])
 @pytest.mark.parametrize(
-    "name", ["TORCH_SDPA", "ROCM_FLASH", "XFORMERS", "FLASHINFER"])
+    "name", ["ROCM_FLASH"])
 @pytest.mark.parametrize("device", ["cpu", "hip", "cuda"])
 def test_env(name: str, device: str, monkeypatch):
     """Test that the attention selector can be set via environment variable.

tests/kernels/test_int8_quant.py

@@ -5,6 +5,7 @@ from tests.kernels.quant_utils import ref_dynamic_per_token_quant
 from tests.kernels.utils import opcheck
 from vllm._custom_ops import scaled_int8_quant
 from vllm.utils import seed_everything
+from vllm.utils import is_hip
 
 DTYPES = [torch.half, torch.bfloat16, torch.float]
 HIDDEN_SIZES = [16, 67, 768, 2048, 5120, 5137, 8192,
@@ -37,7 +38,8 @@ def opcheck_int8_quant_dynamic(output, input, symmetric=True):
         opcheck(torch.ops._C.dynamic_scaled_int8_quant,
                 (output, input, scale, azp))
 
-
+@pytest.mark.skipif(is_hip(),
+                    reason="Currently, there is not supported on ROCm.")
 @pytest.mark.parametrize("num_tokens", NUM_TOKENS)
 @pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
 @pytest.mark.parametrize("dtype", DTYPES)
@@ -60,7 +62,8 @@ def test_dynamic_scaled_int8_quant(num_tokens: int, hidden_size: int,
 
     opcheck_int8_quant_dynamic(ops_out, x)
 
-
+@pytest.mark.skipif(is_hip(),
+                    reason="Currently, there is not supported on ROCm.")
 @pytest.mark.parametrize("num_tokens", NUM_TOKENS)
 @pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
 @pytest.mark.parametrize("dtype", DTYPES)

tests/tokenization/test_tokenizer.py

@@ -5,9 +5,15 @@ from transformers import PreTrainedTokenizerBase
 from vllm.transformers_utils.tokenizer import get_tokenizer
 from ..utils import models_path_prefix
 
+# TOKENIZER_NAMES = [
+#     os.path.join(models_path_prefix, "facebook/opt-125m"),
+#     os.path.join(models_path_prefix, "gpt2"),
+# ]
+
+# export HF_ENDPOINT=https://hf-mirror.com
 TOKENIZER_NAMES = [
-    os.path.join(models_path_prefix, "facebook/opt-125m"),
-    os.path.join(models_path_prefix, "gpt2"),
+    "facebook/opt-125m",
+    "gpt2",
 ]
 
 

vllm/triton_utils/custom_cache_manager.py

@@ -45,7 +45,7 @@ class CustomCacheManager(FileCacheManager):
             self.cache_dir = os.getenv("TRITON_CACHE_DIR",
                                        "").strip() or default_cache_dir()
             if self.cache_dir:
-                self.cache_dir = f"{self.cache_dir}_{os.getpid()}"
+                # self.cache_dir = f"{self.cache_dir}_{os.getpid()}"
                 self.cache_dir = os.path.join(self.cache_dir, self.key)
                 self.lock_path = os.path.join(self.cache_dir, "lock")
                 os.makedirs(self.cache_dir, exist_ok=True)