init v0.23.0

Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>

tests/e2e/pull_request/one_card/_310p/test_classification_310p.py

@@ -0,0 +1,41 @@
+import pytest
+import torch
+from modelscope import snapshot_download  # type: ignore[import-untyped]
+from transformers import AutoModelForSequenceClassification
+
+from tests.e2e.conftest import HfRunner, VllmRunner
+
+
+@pytest.mark.skip("Probabilistic failure, need fix")
+def test_qwen_pooling_classify_correctness() -> None:
+    model_name = snapshot_download("Howeee/Qwen2.5-1.5B-apeach")
+
+    prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is what",
+    ]
+
+    with VllmRunner(
+        model_name,
+        runner="pooling",
+        max_model_len=1024,
+        enforce_eager=True,
+        dtype="float16",
+        gpu_memory_utilization=0.6,
+    ) as vllm_runner:
+        vllm_outputs = vllm_runner.classify(prompts)
+
+    with HfRunner(
+        model_name,
+        dtype="float16",
+        model_kwargs={"attn_implementation": "eager"},
+        auto_cls=AutoModelForSequenceClassification,
+    ) as hf_runner:
+        hf_outputs = hf_runner.classify(prompts)
+
+    for hf_output, vllm_output in zip(hf_outputs, vllm_outputs):
+        hf_output = torch.tensor(hf_output)
+        vllm_output = torch.tensor(vllm_output)
+        assert torch.allclose(hf_output, vllm_output, 1e-2)

tests/e2e/pull_request/one_card/_310p/test_dense_model_310p.py

@@ -0,0 +1,173 @@
+#
+# Copyright (c) 2026 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.
+
+
+from tests.e2e.conftest import VllmRunner, wait_until_npu_memory_free
+from tests.e2e.model_utils import check_outputs_equal
+
+QWEN3_5_PREFIX_MAMBA_PROMPT = (
+    "You are reading a compact synthetic operations ledger. "
+    "Use only the rows below when answering the final question.\n"
+    + "\n".join(
+        f"Row {i}: route R{i:03d} moves cargo from zone {i % 11} to zone {(i * 7) % 13}; priority is {i % 5}."
+        for i in range(64)
+    )
+    + "\n"
+)
+
+QWEN3_5_PREFIX_MAMBA_PROMPTS = [
+    QWEN3_5_PREFIX_MAMBA_PROMPT + "Question: What route is listed in row 17? Answer briefly.",
+    QWEN3_5_PREFIX_MAMBA_PROMPT + "Question: What priority is listed in row 42? Answer briefly.",
+]
+
+
+def _generate_qwen3_5_prefix_mamba_outputs(enable_prefix_caching: bool) -> list[tuple[list[int], str]]:
+    outputs: list[tuple[list[int], str]] = []
+
+    if enable_prefix_caching:
+        with VllmRunner(
+            "Qwen/Qwen3.5-4B",
+            tensor_parallel_size=1,
+            enforce_eager=True,
+            dtype="float16",
+            max_model_len=2048,
+            max_num_batched_tokens=2048,
+            enable_prefix_caching=True,
+            mamba_cache_mode="align",
+            mamba_ssm_cache_dtype="float16",
+        ) as vllm_model:
+            for prompt in QWEN3_5_PREFIX_MAMBA_PROMPTS:
+                outputs.extend(vllm_model.generate_greedy([prompt], max_tokens=8))
+    else:
+        with VllmRunner(
+            "Qwen/Qwen3.5-4B",
+            tensor_parallel_size=1,
+            enforce_eager=True,
+            dtype="float16",
+            max_model_len=2048,
+            max_num_batched_tokens=2048,
+            enable_prefix_caching=False,
+            mamba_ssm_cache_dtype="float16",
+        ) as vllm_model:
+            for prompt in QWEN3_5_PREFIX_MAMBA_PROMPTS:
+                outputs.extend(vllm_model.generate_greedy([prompt], max_tokens=8))
+    return outputs
+
+
+def test_qwen3_dense_tp1_fp16():
+    example_prompts = [
+        "Hello, my name is",
+    ]
+    max_tokens = 5
+    with VllmRunner(
+        "Qwen/Qwen3-8B",
+        tensor_parallel_size=1,
+        enforce_eager=True,
+        dtype="float16",
+        max_model_len=16384,
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)
+
+
+@wait_until_npu_memory_free(0.7)
+def test_qwen3_dense_tp1_fp16_aclgraph():
+    example_prompts = [
+        "Hello, my name is",
+    ] * 8
+    max_tokens = 2
+    with VllmRunner(
+        "Qwen/Qwen3-8B",
+        tensor_parallel_size=1,
+        dtype="float16",
+        max_num_seqs=16,
+        max_model_len=16384,
+        gpu_memory_utilization=0.80,
+        additional_config={"ascend_compilation_config": {"fuse_norm_quant": False}},
+        compilation_config={
+            "cudagraph_mode": "FULL_DECODE_ONLY",
+            "cudagraph_capture_sizes": [1, 2, 4, 8, 16],
+        },
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)
+
+
+def test_qwen3_dense_tp1_w8a8():
+    example_prompts = [
+        "Hello, my name is",
+    ]
+    max_tokens = 5
+    with VllmRunner(
+        "vllm-ascend/Qwen3-8B-W8A8",
+        tensor_parallel_size=1,
+        enforce_eager=True,
+        dtype="float16",
+        quantization="ascend",
+        max_model_len=16384,
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)
+
+
+def test_qwen3_5_dense_tp1_fp16():
+    example_prompts = [
+        "Hello, my name is",
+    ]
+    max_tokens = 5
+    with VllmRunner(
+        "Qwen/Qwen3.5-4B",
+        tensor_parallel_size=1,
+        enforce_eager=True,
+        dtype="float16",
+        max_model_len=16384,
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)
+
+
+@wait_until_npu_memory_free(0.7)
+def test_qwen3_5_dense_prefix_mamba_cache_tp1_fp16():
+    prefix_cache_outputs = _generate_qwen3_5_prefix_mamba_outputs(enable_prefix_caching=True)
+    no_prefix_cache_outputs = _generate_qwen3_5_prefix_mamba_outputs(enable_prefix_caching=False)
+
+    assert len(prefix_cache_outputs) == len(no_prefix_cache_outputs) == len(QWEN3_5_PREFIX_MAMBA_PROMPTS)
+    check_outputs_equal(
+        outputs_0_lst=no_prefix_cache_outputs,
+        outputs_1_lst=prefix_cache_outputs,
+        name_0="no_prefix_cache_outputs",
+        name_1="prefix_cache_outputs",
+    )
+
+
+@wait_until_npu_memory_free(0.7)
+def test_qwen3_5_dense_tp1_fp16_aclgraph():
+    example_prompts = [
+        "Hello, my name is",
+    ] * 8
+    max_tokens = 2
+    with VllmRunner(
+        "Qwen/Qwen3.5-4B",
+        tensor_parallel_size=1,
+        dtype="float16",
+        max_num_seqs=16,
+        max_model_len=16384,
+        gpu_memory_utilization=0.80,
+        additional_config={"ascend_compilation_config": {"fuse_norm_quant": False}},
+        compilation_config={
+            "cudagraph_mode": "FULL_DECODE_ONLY",
+            "cudagraph_capture_sizes": [1, 2, 4, 8, 16],
+        },
+        mamba_ssm_cache_dtype="float16",
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)

tests/e2e/pull_request/one_card/_310p/test_embedding_310p.py

@@ -0,0 +1,106 @@
+#
+# 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 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)
+    with VllmRunner(
+        model_name,
+        runner="pooling",
+        max_model_len=512,
+        enforce_eager=True,
+        dtype="float16",
+        gpu_memory_utilization=0.6,
+    ) as vllm_runner:
+        vllm_outputs = vllm_runner.embed(queries)
+
+    with HfRunner(
+        model_name,
+        dtype="float16",
+        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_bge_m3_correctness():
+    queries = ["What is the capital of China?", "Explain gravity"]
+
+    model_name = snapshot_download("BAAI/bge-m3")
+    with VllmRunner(
+        model_name,
+        runner="pooling",
+        max_model_len=1024,
+        dtype="float16",
+        cudagraph_capture_sizes=[512, 1024],
+        additional_config={"ascend_compilation_config": {"fuse_norm_quant": False}},
+    ) as vllm_aclgraph_runner:
+        vllm_aclgraph_outputs = vllm_aclgraph_runner.embed(queries)
+
+    with VllmRunner(
+        model_name,
+        runner="pooling",
+        max_model_len=1024,
+        dtype="float16",
+        enforce_eager=True,
+    ) as vllm_runner:
+        vllm_eager_outputs = vllm_runner.embed(queries)
+
+    with HfRunner(
+        model_name,
+        dtype="float16",
+        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,
+    )

tests/e2e/pull_request/one_card/_310p/test_scoring_310p.py

@@ -0,0 +1,87 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import pytest
+from modelscope import snapshot_download  # type: ignore[import-untyped]
+
+from tests.e2e.conftest import HfRunner, VllmRunner
+
+CROSS_ENCODER_MODELS = [
+    "BAAI/bge-reranker-v2-m3",  # Roberta
+]
+
+
+TEXTS_1 = [
+    "What is the capital of France?",
+    "What is the capital of Germany?",
+]
+
+TEXTS_2 = [
+    "The capital of France is Paris.",
+    "The capital of Germany is Berlin.",
+]
+
+DTYPE = "float16"
+
+
+@pytest.fixture(scope="module", params=CROSS_ENCODER_MODELS)
+def model_name(request):
+    yield snapshot_download(request.param)
+
+
+def test_cross_encoder_score_1_to_1(model_name):
+    text_pair = [TEXTS_1[0], TEXTS_2[0]]
+
+    with HfRunner(model_name, dtype=DTYPE, is_cross_encoder=True) as hf_model:
+        hf_outputs = hf_model.predict([text_pair]).tolist()
+
+    with VllmRunner(
+        model_name, runner="pooling", dtype=DTYPE, max_model_len=1024, enforce_eager=True, gpu_memory_utilization=0.6
+    ) as vllm_model:
+        vllm_outputs = vllm_model.score(text_pair[0], text_pair[1])
+
+    assert len(vllm_outputs) == 1
+    assert len(hf_outputs) == 1
+
+    assert hf_outputs[0] == pytest.approx(vllm_outputs[0], rel=0.01)
+
+
+def test_cross_encoder_score_1_to_N(model_name):
+    text_pairs = [
+        [TEXTS_1[0], TEXTS_2[0]],
+        [TEXTS_1[0], TEXTS_2[1]],
+    ]
+
+    with HfRunner(model_name, dtype=DTYPE, is_cross_encoder=True) as hf_model:
+        hf_outputs = hf_model.predict(text_pairs).tolist()
+
+    with VllmRunner(
+        model_name, runner="pooling", dtype=DTYPE, max_model_len=1024, enforce_eager=True, gpu_memory_utilization=0.6
+    ) as vllm_model:
+        vllm_outputs = vllm_model.score(TEXTS_1[0], TEXTS_2)
+
+    assert len(vllm_outputs) == 2
+    assert len(hf_outputs) == 2
+
+    assert hf_outputs[0] == pytest.approx(vllm_outputs[0], rel=0.01)
+    assert hf_outputs[1] == pytest.approx(vllm_outputs[1], rel=0.01)
+
+
+def test_cross_encoder_score_N_to_N(model_name):
+    text_pairs = [
+        [TEXTS_1[0], TEXTS_2[0]],
+        [TEXTS_1[1], TEXTS_2[1]],
+    ]
+
+    with HfRunner(model_name, dtype=DTYPE, is_cross_encoder=True) as hf_model:
+        hf_outputs = hf_model.predict(text_pairs).tolist()
+
+    with VllmRunner(
+        model_name, runner="pooling", dtype=DTYPE, max_model_len=1024, enforce_eager=True, gpu_memory_utilization=0.6
+    ) as vllm_model:
+        vllm_outputs = vllm_model.score(TEXTS_1, TEXTS_2)
+
+    assert len(vllm_outputs) == 2
+    assert len(hf_outputs) == 2
+
+    assert hf_outputs[0] == pytest.approx(vllm_outputs[0], rel=0.01)
+    assert hf_outputs[1] == pytest.approx(vllm_outputs[1], rel=0.01)

tests/e2e/pull_request/one_card/_310p/test_spec_decode_mtp_310p.py

@@ -0,0 +1,35 @@
+#
+# Copyright (c) 2026 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.
+
+from tests.e2e.conftest import VllmRunner
+
+
+def test_qwen3_5_mtp_tp1_eager():
+    example_prompts = ["Hello, my name is"]
+    with VllmRunner(
+        "Qwen/Qwen3.5-4B",
+        tensor_parallel_size=1,
+        enforce_eager=True,
+        dtype="float16",
+        max_model_len=2048,
+        mamba_ssm_cache_dtype="float16",
+        speculative_config={
+            "method": "qwen3_5_mtp",
+            "num_speculative_tokens": 1,
+        },
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens=8)

tests/e2e/pull_request/one_card/_310p/test_vl_model_310p.py

@@ -0,0 +1,34 @@
+#
+# Copyright (c) 2026 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.
+
+import os
+import sys
+
+from tests.e2e.conftest import wait_until_npu_memory_free
+
+current_dir = os.path.dirname(os.path.abspath(__file__))
+full_dir = os.path.dirname(os.path.dirname(current_dir))
+sys.path.insert(0, full_dir)
+
+# ruff: noqa: E402
+from tests.e2e.pull_request.utils_310p import run_vl_model_test
+
+
+@wait_until_npu_memory_free(target_free_percentage=0.7)
+def test_qwen3_vl_8b_tp1_fp16():
+    """Qwen3-VL-8B single-card FP16 test"""
+    run_vl_model_test(model_name="Qwen/Qwen3-VL-8B-Instruct", tensor_parallel_size=1, max_tokens=5)