init v0.23.0

Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
2026-08-27 15:11:51 +08:00
parent b582a8e7d1
commit 7f8a1b1f7a
2849 changed files with 712887 additions and 22001 deletions
--- a/tests/e2e/pull_request/one_card/pooling/test_embedding.py
+++ b/tests/e2e/pull_request/one_card/pooling/test_embedding.py
@@ -0,0 +1,135 @@
+#
+# 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/tests/basic_correctness/test_basic_correctness.py
+#
+import huggingface_hub
+import pytest
+from modelscope import snapshot_download  # type: ignore[import-untyped]
+
+from tests.e2e.conftest import HfRunner, VllmRunner
+from tests.e2e.utils import check_embeddings_close
+
+MODELS = [
+    "Qwen/Qwen3-Embedding-0.6B",  # lasttoken
+    "intfloat/multilingual-e5-small",  # mean_tokens
+]
+
+
+@pytest.mark.parametrize("model", MODELS)
+def test_embed_models_correctness(model: str):
+    queries = ["What is the capital of China?", "Explain gravity"]
+
+    model_name = snapshot_download(
+        model,
+        local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
+    )
+    with VllmRunner(
+        model_name,
+        runner="pooling",
+        max_model_len=None,
+        cudagraph_capture_sizes=[4],
+    ) as vllm_runner:
+        vllm_outputs = vllm_runner.embed(queries)
+
+    with HfRunner(
+        model_name,
+        dtype="float32",
+        is_sentence_transformer=True,
+    ) as hf_runner:
+        hf_outputs = hf_runner.encode(queries)
+
+    check_embeddings_close(
+        embeddings_0_lst=hf_outputs,
+        embeddings_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+        tol=1e-2,
+    )
+
+
+def test_causal_embed_models_using_prefix_caching_correctness():
+    # This test is to verify the correctness of prefix caching for embedding models.
+    # We compare the outputs of vLLM with and without prefix caching enabled, and check if they are close enough.
+    # We set the input query to be very long to make sure prefix caching is triggered.
+    queries = ["What is the capital of China?" * 256, "Explain gravity"]
+
+    model_name = snapshot_download(
+        "Qwen/Qwen3-Embedding-0.6B",
+        local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
+    )
+    with VllmRunner(
+        model_name,
+        runner="pooling",
+        max_model_len=None,
+        cudagraph_capture_sizes=[4],
+        enable_prefix_caching=True,
+    ) as vllm_runner_using_caching:
+        vllm_outputs_without_caching = vllm_runner_using_caching.embed(queries)
+        vllm_outputs_with_caching = vllm_runner_using_caching.embed(queries)
+
+    check_embeddings_close(
+        embeddings_0_lst=vllm_outputs_without_caching,
+        embeddings_1_lst=vllm_outputs_with_caching,
+        name_0="without_caching",
+        name_1="with_caching",
+        tol=1e-2,
+    )
+
+
+def test_bge_m3_correctness():
+    queries = ["What is the capital of China?", "Explain gravity"]
+
+    model_name = snapshot_download(
+        "BAAI/bge-m3",
+        local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
+    )
+    with VllmRunner(
+        model_name,
+        runner="pooling",
+        cudagraph_capture_sizes=[4],
+    ) as vllm_aclgraph_runner:
+        vllm_aclgraph_outputs = vllm_aclgraph_runner.embed(queries)
+
+    with VllmRunner(
+        model_name,
+        runner="pooling",
+        enforce_eager=True,
+    ) as vllm_runner:
+        vllm_eager_outputs = vllm_runner.embed(queries)
+
+    with HfRunner(
+        model_name,
+        dtype="float32",
+        is_sentence_transformer=True,
+    ) as hf_runner:
+        hf_outputs = hf_runner.encode(queries)
+
+    check_embeddings_close(
+        embeddings_0_lst=hf_outputs,
+        embeddings_1_lst=vllm_eager_outputs,
+        name_0="hf",
+        name_1="vllm",
+        tol=1e-2,
+    )
+
+    check_embeddings_close(
+        embeddings_0_lst=vllm_eager_outputs,
+        embeddings_1_lst=vllm_aclgraph_outputs,
+        name_0="eager",
+        name_1="aclgraph",
+        tol=1e-2,
+    )