add qwen3

vllm-v0.6.2/tests/entrypoints/llm/test_accuracy.py

@@ -0,0 +1,56 @@
+"""
+This file test accuracy of the vLLM server via LMEval.
+It uses local-completions, which interacts with vLLM
+through the OAI API with N concurrent connections.
+This simulates real work usage of the API and makes
+sure that the zmq frontend mp RPC message passing and
+AsyncLLMEngine are working correctly.
+"""
+
+import lm_eval
+import pytest
+
+from vllm.platforms import current_platform
+
+MODEL_NAME = "Qwen/Qwen2-1.5B-Instruct"
+NUM_CONCURRENT = 500
+TASK = "gsm8k"
+FILTER = "exact_match,strict-match"
+RTOL = 0.03
+EXPECTED_VALUE = 0.58
+
+
+def run_test():
+    """Run the end to end accuracy test."""
+
+    model_args = f"pretrained={MODEL_NAME},max_model_len=2048"
+
+    results = lm_eval.simple_evaluate(
+        model="vllm",
+        model_args=model_args,
+        tasks="gsm8k",
+        batch_size="auto",
+    )
+
+    measured_value = results["results"][TASK][FILTER]
+    assert (measured_value - RTOL < EXPECTED_VALUE
+            and measured_value + RTOL > EXPECTED_VALUE
+            ), f"Expected: {EXPECTED_VALUE} |  Measured: {measured_value}"
+
+
+@pytest.mark.skipif(not current_platform.is_cuda(),
+                    reason="V1 is currently only supported on CUDA.")
+def test_lm_eval_accuracy_v1_engine(monkeypatch):
+    """Run with the V1 Engine."""
+
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_USE_V1", "1")
+        run_test()
+
+
+def test_lm_eval_accuracy_v0_engine(monkeypatch):
+    """Run with the V0 Engine."""
+
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_USE_V1", "0")
+        run_test()

vllm-v0.6.2/tests/entrypoints/llm/test_chat.py

@@ -0,0 +1,93 @@
+from typing import List
+
+import pytest
+
+from vllm import LLM
+
+from ..openai.test_vision import TEST_IMAGE_URLS
+
+
+def test_chat():
+    llm = LLM(model="meta-llama/Llama-3.2-1B-Instruct")
+
+    prompt1 = "Explain the concept of entropy."
+    messages = [
+        {
+            "role": "system",
+            "content": "You are a helpful assistant"
+        },
+        {
+            "role": "user",
+            "content": prompt1
+        },
+    ]
+    outputs = llm.chat(messages)
+    assert len(outputs) == 1
+
+
+def test_multi_chat():
+    llm = LLM(model="meta-llama/Llama-3.2-1B-Instruct")
+
+    prompt1 = "Explain the concept of entropy."
+    prompt2 = "Explain what among us is."
+
+    conversation1 = [
+        {
+            "role": "system",
+            "content": "You are a helpful assistant"
+        },
+        {
+            "role": "user",
+            "content": prompt1
+        },
+    ]
+
+    conversation2 = [
+        {
+            "role": "system",
+            "content": "You are a helpful assistant"
+        },
+        {
+            "role": "user",
+            "content": prompt2
+        },
+    ]
+
+    messages = [conversation1, conversation2]
+
+    outputs = llm.chat(messages)
+    assert len(outputs) == 2
+
+
+@pytest.mark.skip("Not support Phi vision model yet.")
+@pytest.mark.parametrize("image_urls",
+                         [[TEST_IMAGE_URLS[0], TEST_IMAGE_URLS[1]]])
+def test_chat_multi_image(image_urls: List[str]):
+    llm = LLM(
+        model="microsoft/Phi-3.5-vision-instruct",
+        dtype="bfloat16",
+        max_model_len=4096,
+        max_num_seqs=5,
+        enforce_eager=True,
+        trust_remote_code=True,
+        limit_mm_per_prompt={"image": 2},
+    )
+
+    messages = [{
+        "role":
+        "user",
+        "content": [
+            *({
+                "type": "image_url",
+                "image_url": {
+                    "url": image_url
+                }
+            } for image_url in image_urls),
+            {
+                "type": "text",
+                "text": "What's in this image?"
+            },
+        ],
+    }]
+    outputs = llm.chat(messages)
+    assert len(outputs) >= 0

vllm-v0.6.2/tests/entrypoints/llm/test_encode.py

@@ -0,0 +1,107 @@
+import weakref
+from typing import List
+
+import pytest
+
+from vllm import LLM, EmbeddingRequestOutput, PoolingParams
+from vllm.distributed import cleanup_dist_env_and_memory
+
+MODEL_NAME = "intfloat/e5-mistral-7b-instruct"
+
+PROMPTS = [
+    "Hello, my name is",
+    "The president of the United States is",
+    "The capital of France is",
+    "The future of AI is",
+]
+
+TOKEN_IDS = [
+    # Using ID={0, 1, 2, 3} results in NaN values,
+    # so we add this offset of 1000
+    [1000],
+    [1000, 1001],
+    [1000, 1002, 1001],
+    [1000, 1003, 1001, 1002],
+]
+
+
+@pytest.fixture(scope="module")
+def llm():
+    # pytest caches the fixture so we use weakref.proxy to
+    # enable garbage collection
+    llm = LLM(model=MODEL_NAME,
+              max_num_batched_tokens=32768,
+              tensor_parallel_size=1,
+              gpu_memory_utilization=0.75,
+              enforce_eager=True)
+
+    with llm.deprecate_legacy_api():
+        yield weakref.proxy(llm)
+
+        del llm
+
+    cleanup_dist_env_and_memory()
+
+
+def assert_outputs_equal(o1: List[EmbeddingRequestOutput],
+                         o2: List[EmbeddingRequestOutput]):
+    assert [o.outputs for o in o1] == [o.outputs for o in o2]
+
+
+@pytest.mark.skip_global_cleanup
+@pytest.mark.parametrize('prompt_token_ids', TOKEN_IDS)
+def test_v1_v2_api_consistency_single_prompt_tokens(llm: LLM,
+                                                    prompt_token_ids):
+    pooling_params = PoolingParams()
+
+    with pytest.warns(DeprecationWarning, match="'prompt_token_ids'"):
+        v1_output = llm.encode(prompt_token_ids=prompt_token_ids,
+                               pooling_params=pooling_params)
+
+    v2_output = llm.encode({"prompt_token_ids": prompt_token_ids},
+                           pooling_params=pooling_params)
+    assert_outputs_equal(v1_output, v2_output)
+
+
+@pytest.mark.skip_global_cleanup
+def test_v1_v2_api_consistency_multi_prompt_tokens(llm: LLM):
+    pooling_params = PoolingParams()
+
+    with pytest.warns(DeprecationWarning, match="'prompt_token_ids'"):
+        v1_output = llm.encode(prompt_token_ids=TOKEN_IDS,
+                               pooling_params=pooling_params)
+
+    v2_output = llm.encode(
+        [{
+            "prompt_token_ids": p
+        } for p in TOKEN_IDS],
+        pooling_params=pooling_params,
+    )
+    assert_outputs_equal(v1_output, v2_output)
+
+
+@pytest.mark.skip_global_cleanup
+def test_multiple_pooling_params(llm: LLM):
+    pooling_params = [
+        PoolingParams(),
+        PoolingParams(),
+        PoolingParams(),
+        PoolingParams(),
+    ]
+
+    # Multiple PoolingParams should be matched with each prompt
+    outputs = llm.encode(PROMPTS, pooling_params=pooling_params)
+    assert len(PROMPTS) == len(outputs)
+
+    # Exception raised, if the size of params does not match the size of prompts
+    with pytest.raises(ValueError):
+        outputs = llm.encode(PROMPTS, pooling_params=pooling_params[:3])
+
+    # Single PoolingParams should be applied to every prompt
+    single_pooling_params = PoolingParams()
+    outputs = llm.encode(PROMPTS, pooling_params=single_pooling_params)
+    assert len(PROMPTS) == len(outputs)
+
+    # pooling_params is None, default params should be applied
+    outputs = llm.encode(PROMPTS, pooling_params=None)
+    assert len(PROMPTS) == len(outputs)

vllm-v0.6.2/tests/entrypoints/llm/test_generate.py

@@ -0,0 +1,104 @@
+import weakref
+from typing import List
+
+import pytest
+
+from vllm import LLM, RequestOutput, SamplingParams
+from vllm.distributed import cleanup_dist_env_and_memory
+
+MODEL_NAME = "facebook/opt-125m"
+
+PROMPTS = [
+    "Hello, my name is",
+    "The president of the United States is",
+    "The capital of France is",
+    "The future of AI is",
+]
+
+TOKEN_IDS = [
+    [0],
+    [0, 1],
+    [0, 2, 1],
+    [0, 3, 1, 2],
+]
+
+
+@pytest.fixture(scope="module")
+def llm():
+    # pytest caches the fixture so we use weakref.proxy to
+    # enable garbage collection
+    llm = LLM(model=MODEL_NAME,
+              max_num_batched_tokens=4096,
+              tensor_parallel_size=1,
+              gpu_memory_utilization=0.10,
+              enforce_eager=True)
+
+    with llm.deprecate_legacy_api():
+        yield weakref.proxy(llm)
+
+        del llm
+
+    cleanup_dist_env_and_memory()
+
+
+def assert_outputs_equal(o1: List[RequestOutput], o2: List[RequestOutput]):
+    assert [o.outputs for o in o1] == [o.outputs for o in o2]
+
+
+@pytest.mark.skip_global_cleanup
+@pytest.mark.parametrize('prompt_token_ids', TOKEN_IDS)
+def test_v1_v2_api_consistency_single_prompt_tokens(llm: LLM,
+                                                    prompt_token_ids):
+    sampling_params = SamplingParams(temperature=0.0, top_p=1.0)
+
+    with pytest.warns(DeprecationWarning, match="'prompt_token_ids'"):
+        v1_output = llm.generate(prompt_token_ids=prompt_token_ids,
+                                 sampling_params=sampling_params)
+
+    v2_output = llm.generate({"prompt_token_ids": prompt_token_ids},
+                             sampling_params=sampling_params)
+    assert_outputs_equal(v1_output, v2_output)
+
+
+@pytest.mark.skip_global_cleanup
+def test_v1_v2_api_consistency_multi_prompt_tokens(llm: LLM):
+    sampling_params = SamplingParams(temperature=0.0, top_p=1.0)
+
+    with pytest.warns(DeprecationWarning, match="'prompt_token_ids'"):
+        v1_output = llm.generate(prompt_token_ids=TOKEN_IDS,
+                                 sampling_params=sampling_params)
+
+    v2_output = llm.generate(
+        [{
+            "prompt_token_ids": p
+        } for p in TOKEN_IDS],
+        sampling_params=sampling_params,
+    )
+    assert_outputs_equal(v1_output, v2_output)
+
+
+@pytest.mark.skip_global_cleanup
+def test_multiple_sampling_params(llm: LLM):
+    sampling_params = [
+        SamplingParams(temperature=0.01, top_p=0.95),
+        SamplingParams(temperature=0.3, top_p=0.95),
+        SamplingParams(temperature=0.7, top_p=0.95),
+        SamplingParams(temperature=0.99, top_p=0.95),
+    ]
+
+    # Multiple SamplingParams should be matched with each prompt
+    outputs = llm.generate(PROMPTS, sampling_params=sampling_params)
+    assert len(PROMPTS) == len(outputs)
+
+    # Exception raised, if the size of params does not match the size of prompts
+    with pytest.raises(ValueError):
+        outputs = llm.generate(PROMPTS, sampling_params=sampling_params[:3])
+
+    # Single SamplingParams should be applied to every prompt
+    single_sampling_params = SamplingParams(temperature=0.3, top_p=0.95)
+    outputs = llm.generate(PROMPTS, sampling_params=single_sampling_params)
+    assert len(PROMPTS) == len(outputs)
+
+    # sampling_params is None, default params should be applied
+    outputs = llm.generate(PROMPTS, sampling_params=None)
+    assert len(PROMPTS) == len(outputs)

vllm-v0.6.2/tests/entrypoints/llm/test_generate_multiple_loras.py

@@ -0,0 +1,66 @@
+import weakref
+
+import pytest
+# downloading lora to test lora requests
+from huggingface_hub import snapshot_download
+
+from vllm import LLM
+from vllm.distributed import cleanup_dist_env_and_memory
+from vllm.lora.request import LoRARequest
+
+MODEL_NAME = "HuggingFaceH4/zephyr-7b-beta"
+
+PROMPTS = [
+    "Hello, my name is",
+    "The president of the United States is",
+    "The capital of France is",
+    "The future of AI is",
+]
+
+LORA_NAME = "typeof/zephyr-7b-beta-lora"
+
+
+@pytest.fixture(scope="module")
+def llm():
+    # pytest caches the fixture so we use weakref.proxy to
+    # enable garbage collection
+    llm = LLM(model=MODEL_NAME,
+              tensor_parallel_size=1,
+              max_model_len=8192,
+              enable_lora=True,
+              max_loras=4,
+              max_lora_rank=64,
+              max_num_seqs=128,
+              enforce_eager=True)
+
+    with llm.deprecate_legacy_api():
+        yield weakref.proxy(llm)
+
+        del llm
+
+    cleanup_dist_env_and_memory()
+
+
+@pytest.fixture(scope="module")
+def zephyr_lora_files():
+    return snapshot_download(repo_id=LORA_NAME)
+
+
+@pytest.mark.skip_global_cleanup
+def test_multiple_lora_requests(llm: LLM, zephyr_lora_files):
+    lora_request = [
+        LoRARequest(LORA_NAME + str(idx), idx + 1, zephyr_lora_files)
+        for idx in range(len(PROMPTS))
+    ]
+    # Multiple SamplingParams should be matched with each prompt
+    outputs = llm.generate(PROMPTS, lora_request=lora_request)
+    assert len(PROMPTS) == len(outputs)
+
+    # Exception raised, if the size of params does not match the size of prompts
+    with pytest.raises(ValueError):
+        outputs = llm.generate(PROMPTS, lora_request=lora_request[:1])
+
+    # Single LoRARequest should be applied to every prompt
+    single_lora_request = lora_request[0]
+    outputs = llm.generate(PROMPTS, lora_request=single_lora_request)
+    assert len(PROMPTS) == len(outputs)

vllm-v0.6.2/tests/entrypoints/llm/test_guided_generate.py

@@ -0,0 +1,161 @@
+import json
+import re
+import weakref
+
+import jsonschema
+import pytest
+
+from vllm.distributed import cleanup_dist_env_and_memory
+from vllm.entrypoints.llm import LLM
+from vllm.outputs import RequestOutput
+from vllm.sampling_params import GuidedDecodingParams, SamplingParams
+
+MODEL_NAME = "HuggingFaceH4/zephyr-7b-beta"
+
+
+@pytest.fixture(scope="module")
+def llm():
+    # pytest caches the fixture so we use weakref.proxy to
+    # enable garbage collection
+    llm = LLM(model=MODEL_NAME, max_model_len=1024)
+
+    with llm.deprecate_legacy_api():
+        yield weakref.proxy(llm)
+        del llm
+    cleanup_dist_env_and_memory()
+
+
+@pytest.mark.skip_global_cleanup
+def test_guided_regex(sample_regex, llm):
+    sampling_params = SamplingParams(
+        temperature=0.8,
+        top_p=0.95,
+        guided_decoding=GuidedDecodingParams(regex=sample_regex))
+    outputs = llm.generate(prompts=[
+        f"Give an example IPv4 address with this regex: {sample_regex}"
+    ] * 2,
+                           sampling_params=sampling_params,
+                           use_tqdm=True)
+
+    assert outputs is not None
+    for output in outputs:
+        assert output is not None
+        assert isinstance(output, RequestOutput)
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        print(generated_text)
+        assert generated_text is not None
+        assert re.fullmatch(sample_regex, generated_text) is not None
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+
+
+@pytest.mark.skip_global_cleanup
+def test_guided_json_completion(sample_json_schema, llm):
+    sampling_params = SamplingParams(
+        temperature=1.0,
+        max_tokens=1000,
+        guided_decoding=GuidedDecodingParams(json=sample_json_schema))
+    outputs = llm.generate(prompts=[
+        f"Give an example JSON for an employee profile "
+        f"that fits this schema: {sample_json_schema}"
+    ] * 2,
+                           sampling_params=sampling_params,
+                           use_tqdm=True)
+
+    assert outputs is not None
+
+    for output in outputs:
+        assert output is not None
+        assert isinstance(output, RequestOutput)
+        prompt = output.prompt
+
+        generated_text = output.outputs[0].text
+        assert generated_text is not None
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+        output_json = json.loads(generated_text)
+        jsonschema.validate(instance=output_json, schema=sample_json_schema)
+
+
+@pytest.mark.skip_global_cleanup
+def test_guided_choice_completion(sample_guided_choice, llm):
+    sampling_params = SamplingParams(
+        temperature=0.8,
+        top_p=0.95,
+        guided_decoding=GuidedDecodingParams(choice=sample_guided_choice))
+    outputs = llm.generate(
+        prompts="The best language for type-safe systems programming is ",
+        sampling_params=sampling_params,
+        use_tqdm=True)
+
+    assert outputs is not None
+    for output in outputs:
+        assert output is not None
+        assert isinstance(output, RequestOutput)
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        print(generated_text)
+        assert generated_text is not None
+        assert generated_text in sample_guided_choice
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+
+
+@pytest.mark.skip_global_cleanup
+def test_guided_grammar(sample_sql_statements, llm):
+
+    sampling_params = SamplingParams(
+        temperature=0.8,
+        top_p=0.95,
+        max_tokens=1000,
+        guided_decoding=GuidedDecodingParams(grammar=sample_sql_statements))
+    outputs = llm.generate(
+        prompts=("Generate a sql state that select col_1 from "
+                 "table_1 where it is equals to 1"),
+        sampling_params=sampling_params,
+        use_tqdm=True,
+    )
+
+    assert outputs is not None
+    for output in outputs:
+        assert output is not None
+        assert isinstance(output, RequestOutput)
+        prompt = output.prompt
+
+        generated_text = output.outputs[0].text
+        assert generated_text is not None
+        # use Lark to parse the output, and make sure it's a valid parse tree
+        from lark import Lark
+        parser = Lark(sample_sql_statements)
+        parser.parse(generated_text)
+
+        # remove spaces for comparison b/c we removed them in the grammar
+        ground_truth = "SELECT col_1 from table_1 where col_1 = 1".replace(
+            " ", "")
+
+        assert generated_text.strip() == ground_truth
+
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+
+
+@pytest.mark.skip_global_cleanup
+def test_guided_options_request_deprecation_warning(sample_regex, llm):
+    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+    with pytest.warns(DeprecationWarning, match="guided_options_request"):
+        llm.generate(prompts="This should fail",
+                     sampling_params=sampling_params,
+                     use_tqdm=True,
+                     guided_options_request=dict(guided_regex=sample_regex))
+
+
+@pytest.mark.skip_global_cleanup
+def test_validation_against_both_guided_decoding_options(sample_regex, llm):
+    sampling_params = SamplingParams(
+        temperature=0.8,
+        top_p=0.95,
+        guided_decoding=GuidedDecodingParams(regex=sample_regex))
+
+    with pytest.raises(ValueError, match="Cannot set both"):
+        llm.generate(prompts="This should fail",
+                     sampling_params=sampling_params,
+                     use_tqdm=True,
+                     guided_options_request=dict(guided_regex=sample_regex))

vllm-v0.6.2/tests/entrypoints/llm/test_init.py

@@ -0,0 +1,22 @@
+import pytest
+
+from vllm import LLM
+
+from ...utils import error_on_warning
+
+MODEL_NAME = "facebook/opt-125m"
+
+
+def test_pos_args_deprecated():
+    with error_on_warning(DeprecationWarning):
+        LLM(model=MODEL_NAME, tokenizer=MODEL_NAME)
+
+    with error_on_warning(DeprecationWarning):
+        LLM(MODEL_NAME, tokenizer=MODEL_NAME)
+
+    with pytest.warns(DeprecationWarning, match="'tokenizer'"):
+        LLM(MODEL_NAME, MODEL_NAME)
+
+    with pytest.warns(DeprecationWarning,
+                      match="'tokenizer', 'tokenizer_mode'"):
+        LLM(MODEL_NAME, MODEL_NAME, "auto")

vllm-v0.6.2/tests/entrypoints/llm/test_lazy_outlines.py

@@ -0,0 +1,55 @@
+import sys
+
+from vllm import LLM, SamplingParams
+from vllm.distributed import cleanup_dist_env_and_memory
+
+
+def test_lazy_outlines(sample_regex):
+    """If users don't use guided decoding, outlines should not be imported.
+    """
+    prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
+    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+    # Create an LLM without guided decoding as a baseline.
+    llm = LLM(model="facebook/opt-125m",
+              enforce_eager=True,
+              gpu_memory_utilization=0.3)
+    outputs = llm.generate(prompts, sampling_params)
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+
+    # make sure outlines is not imported
+    assert 'outlines' not in sys.modules
+
+    # Destroy the LLM object and free up the GPU memory.
+    del llm
+    cleanup_dist_env_and_memory()
+
+    # Create an LLM with guided decoding enabled.
+    llm = LLM(model="facebook/opt-125m",
+              enforce_eager=True,
+              guided_decoding_backend="lm-format-enforcer",
+              gpu_memory_utilization=0.6)
+    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+    outputs = llm.generate(
+        prompts=[
+            f"Give an example IPv4 address with this regex: {sample_regex}"
+        ] * 2,
+        sampling_params=sampling_params,
+        use_tqdm=True,
+        guided_options_request=dict(guided_regex=sample_regex))
+
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+
+    # make sure outlines is not imported
+    assert 'outlines' not in sys.modules

vllm-v0.6.2/tests/entrypoints/llm/test_prompt_validation.py

@@ -0,0 +1,25 @@
+import pytest
+
+from vllm import LLM
+
+
+@pytest.fixture(autouse=True)
+def v1(run_with_both_engines):
+    # Simple autouse wrapper to run both engines for each test
+    # This can be promoted up to conftest.py to run for every
+    # test in a package
+    pass
+
+
+@pytest.mark.skip_v1
+def test_empty_prompt():
+    llm = LLM(model="gpt2", enforce_eager=True)
+    with pytest.raises(ValueError, match='Prompt cannot be empty'):
+        llm.generate([""])
+
+
+@pytest.mark.skip_v1
+def test_out_of_vocab_token():
+    llm = LLM(model="gpt2", enforce_eager=True)
+    with pytest.raises(ValueError, match='out of vocabulary'):
+        llm.generate({"prompt_token_ids": [999999]})