init v0.23.0

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

tests/e2e/pull_request/two_card/lora/test_ilama_lora_tp2.py

@@ -0,0 +1,24 @@
+import pytest
+
+from tests.e2e.conftest import VllmRunner
+from tests.e2e.pull_request.one_card.lora.test_ilama_lora import EXPECTED_LORA_OUTPUT, MODEL_PATH, do_sample
+
+
+@pytest.mark.parametrize("distributed_executor_backend", ["mp"])
+def test_ilama_lora_tp2(distributed_executor_backend, ilama_lora_files):
+    with VllmRunner(
+        MODEL_PATH,
+        enable_lora=True,
+        max_loras=4,
+        dtype="half",
+        max_model_len=1024,
+        max_num_seqs=16,
+        tensor_parallel_size=2,
+        cudagraph_capture_sizes=[1, 2, 4, 8],
+        distributed_executor_backend=distributed_executor_backend,
+        enforce_eager=True,
+    ) as vllm_model:
+        output = do_sample(vllm_model.model, ilama_lora_files, lora_id=2)
+
+    for i in range(len(EXPECTED_LORA_OUTPUT)):
+        assert output[i] == EXPECTED_LORA_OUTPUT[i]

tests/e2e/pull_request/two_card/lora/test_llama32_lora_tp2.py

@@ -0,0 +1,30 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import pytest
+
+from tests.e2e.conftest import VllmRunner, wait_until_npu_memory_free
+from tests.e2e.pull_request.one_card.lora.test_llama32_lora import generate_and_test
+from vllm_ascend.utils import enable_custom_op
+
+enable_custom_op()
+
+# For hk region, we need to use the model from hf to avoid the network issue
+MODEL_PATH = "vllm-ascend/Llama-3.2-3B-Instruct"
+
+
+@pytest.mark.parametrize("fully_sharded_loras", [False, True])
+@wait_until_npu_memory_free()
+def test_llama_lora_tp2(llama32_lora_files, fully_sharded_loras):
+    with VllmRunner(
+        MODEL_PATH,
+        enable_lora=True,
+        # also test odd max_num_seqs
+        max_num_seqs=7,
+        max_model_len=1024,
+        max_loras=4,
+        tensor_parallel_size=2,
+        fully_sharded_loras=fully_sharded_loras,
+        compilation_config={"cudagraph_mode": "PIECEWISE"},
+    ) as vllm_model:
+        llm = vllm_model.model
+        generate_and_test(llm, llama32_lora_files)

tests/e2e/pull_request/two_card/lora/test_qwen3moe_lora_tp.py

@@ -0,0 +1,68 @@
+import vllm
+from vllm.lora.request import LoRARequest
+
+MODEL_PATH = "Qwen/Qwen3-30B-A3B"
+
+PROMPT_TEMPLATE = """<|im_start|>user
+I want you to act as a SQL terminal in front of an example database, you need only to return the sql command to me.Below is an instruction that describes a task, Write a response that appropriately completes the request.
+"
+##Instruction:
+candidate_poll contains tables such as candidate, people. Table candidate has columns such as Candidate_ID, People_ID, Poll_Source, Date, Support_rate, Consider_rate, Oppose_rate, Unsure_rate. Candidate_ID is the primary key.
+Table people has columns such as People_ID, Sex, Name, Date_of_Birth, Height, Weight. People_ID is the primary key.
+The People_ID of candidate is the foreign key of People_ID of people.
+
+
+###Input:
+{context}
+
+###Response:<|im_end|>
+<|im_start|>assistant"""  # noqa: E501
+
+EXPECTED_LORA_OUTPUT = [
+    "<think>\n\n</think>\n\nSELECT count(*) FROM candidate",
+    "<think>\n\n</think>\n\nSELECT count(*) FROM candidate",
+    "<think>\n\n</think>\n\nSELECT poll_source FROM candidate GROUP BY poll_source ORDER BY count(*) DESC LIMIT 1",  # noqa: E501
+    "<think>\n\n</think>\n\nSELECT poll_source FROM candidate GROUP BY poll_source ORDER BY count(*) DESC LIMIT 1",  # noqa: E501
+]
+
+
+def generate_and_test(llm: vllm.LLM, lora_path: str, lora_id: int) -> None:
+    prompts = [
+        PROMPT_TEMPLATE.format(context="How many candidates are there?"),
+        PROMPT_TEMPLATE.format(context="Count the number of candidates."),
+        PROMPT_TEMPLATE.format(
+            context="Which poll resource provided the most number of candidate information?"  # noqa: E501
+        ),
+        PROMPT_TEMPLATE.format(context="Return the poll resource associated with the most candidates."),
+    ]
+    sampling_params = vllm.SamplingParams(temperature=0, max_tokens=64)
+    outputs = llm.generate(
+        prompts,
+        sampling_params,
+        lora_request=LoRARequest(str(lora_id), lora_id, lora_path) if lora_id else None,
+    )
+    # Print the outputs.
+    generated_texts: list[str] = []
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text.strip()
+        generated_texts.append(generated_text)
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+
+    for i in range(len(EXPECTED_LORA_OUTPUT)):
+        assert generated_texts[i].startswith(EXPECTED_LORA_OUTPUT[i])
+
+
+def test_qwen3moe_lora(qwen3moe_lora_files):
+    llm = vllm.LLM(
+        MODEL_PATH,
+        max_model_len=1024,
+        enable_lora=True,
+        max_loras=4,
+        enforce_eager=True,
+        trust_remote_code=True,
+        enable_chunked_prefill=True,
+        tensor_parallel_size=2,
+    )
+
+    generate_and_test(llm, qwen3moe_lora_files, lora_id=1)