[Frontend, Core] Adding stop and stop_token_ids for beam search.

[nFunctor]

An attempted fix to #9253

Update 15/10: PR updated post-merge of beam search methods.

• Removing the eos check led to no stopping in chat modes, I keep it as it is.
• Is fairly simplistic so keeps the stop strings/tokens inside (current implem with does the same with eos; as a consequence the chat mode is not clean)
• I've decided to not add lines like if not all beams since at the moment I do not really understand if they do anything.
• globally still not quite satisfied with the solution but it provides a workable minimum

Offline tests

from vllm import LLM
from vllm.sampling_params import BeamSearchParams

model_name = "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4"

llm = LLM(
    model=model_name,
    max_model_len=2000,
    enable_prefix_caching=True,
    num_scheduler_steps=1
)

prompt = "Let's count to ten: one, two, three"

params = BeamSearchParams(
    beam_width=3,
    max_tokens=100,
    temperature=0.0,
    stop=["six"]
)

print(llm.beam_search([prompt], params)[0].sequences[0].text)
# result: <|begin_of_text|>Let's count to ten: one, two, three, four, five, six


params = BeamSearchParams(
    beam_width=3,
    max_tokens=100,
    temperature=0.0,
    stop_token_ids=[llm.get_tokenizer().encode(" six")[1]] # space needed due to llama tokenizer
)

print(llm.beam_search([prompt], params)[0].sequences[0].text)
# result: <|begin_of_text|>Let's count to ten: one, two, three, four, five, six


params = BeamSearchParams(
    beam_width=3,
    max_tokens=100,
    temperature=0.0,
    ignore_eos=False
)

chat = llm.get_tokenizer().apply_chat_template([
    {"role": "user", "content": "Hello, how are you?"},
], tokenize=False, add_generation_prompt=True)

result = llm.beam_search([chat], params)

print(result[0].sequences[0].tokens[-1] == llm.get_tokenizer().eos_token_id)
# result: True

params = BeamSearchParams(
    beam_width=3,
    max_tokens=100,
    temperature=0.0,
    ignore_eos=True
)

chat = llm.get_tokenizer().apply_chat_template([
    {"role": "user", "content": "Hello, how are you?"},
], tokenize=False, add_generation_prompt=True)

result = llm.beam_search([chat], params)
print(result[0].sequences[0].tokens[-1] == llm.get_tokenizer().eos_token_id)
# result: False

Online tests

from openai import OpenAI
from transformers import AutoTokenizer

model_name = "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4"
client = OpenAI(api_key="EMPTY", base_url="http://localhost:8093/v1")
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Let's count to ten: one, two, three"
result = client.completions.create(
    model=model_name,
    prompt=prompt,
    n=3,
    max_tokens=100,
    extra_body={"use_beam_search": True},
    stop=["six"]
)

print(result.choices[0].text)
#output: ", four, five, six"

result = client.completions.create(
    model=model_name,
    prompt=prompt,
    n=3,
    max_tokens=100,
    extra_body={"use_beam_search": True, "stop_token_ids": [tokenizer.encode(" six")[1]]}
)

print(result.choices[0].text)
#output: ", four, five, six"

result = client.chat.completions.create(
    model=model_name,
    messages=[
        {"role": "user", "content": "Hello, how are you?"}
    ],
    n=3,
    max_tokens=100,
    extra_body={"use_beam_search": True, "ignore_eos": False}
)

print(result.choices[0].message.content)
# ends with <|eot_id|>

result = client.chat.completions.create(
    model=model_name,
    messages=[
        {"role": "user", "content": "Hello, how are you?"}
    ],
    n=3,
    max_tokens=100,
    extra_body={"use_beam_search": True, "ignore_eos": True}
)

print(result.choices[0].message.content)
# Usually does not end with <|eot_id|>

Original PR

The new beam search lacks any information about stop or stop_token_ids, this is my attempt to integrate it. The basic idea is to add these options to SamplingParams, generate the next token with them in mind and then do the stop check.

• I've got confused by the eos token check in the original code. It is integrated into the SamplingParams for 1-token gen but perhaps this is a false step.
• Some logging.info were removed since they seem to be excessive. Even now, the engines tend to spit out the requests at the end even with --disable-log-requests
• Please review, not used to commit to projects that vast but hopefully this helps!

@youkaichao @LunrEclipse

BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE

---

PR Checklist (Click to Expand)

Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.

PR Title and Classification

Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:

• [Bugfix] for bug fixes.
• [CI/Build] for build or continuous integration improvements.
• [Doc] for documentation fixes and improvements.
• [Model] for adding a new model or improving an existing model. Model name should appear in the title.
• [Frontend] For changes on the vLLM frontend (e.g., OpenAI API server, LLM class, etc.)
• [Kernel] for changes affecting CUDA kernels or other compute kernels.
• [Core] for changes in the core vLLM logic (e.g., LLMEngine, AsyncLLMEngine, Scheduler, etc.)
• [Hardware][Vendor] for hardware-specific changes. Vendor name should appear in the prefix (e.g., [Hardware][AMD]).
• [Misc] for PRs that do not fit the above categories. Please use this sparingly.

Note: If the PR spans more than one category, please include all relevant prefixes.

Code Quality

The PR need to meet the following code quality standards:

• We adhere to Google Python style guide and Google C++ style guide.
• Pass all linter checks. Please use format.sh to format your code.
• The code need to be well-documented to ensure future contributors can easily understand the code.
• Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.
• Please add documentation to docs/source/ if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.

Adding or changing kernels

Each custom kernel needs a schema and one or more implementations to be registered with PyTorch.

• Make sure custom ops are registered following PyTorch guidelines: Custom C++ and CUDA Operators and The Custom Operators Manual
• Custom operations that return Tensors require meta-functions. Meta-functions should be implemented and registered in python so that dynamic dims can be handled automatically. See above documents for a description of meta-functions.
• Use torch.libary.opcheck() to test the function registration and meta-function for any registered ops. See tests/kernels for examples.
• When changing the C++ signature of an existing op, the schema must be updated to reflect the changes.
• If a new custom type is needed, see the following document: Custom Class Support in PT2.

Notes for Large Changes

Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with rfc-required and might not go through the PR.

What to Expect for the Reviews

The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:

• After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.
• After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.
• After the review, the reviewer will put an action-required label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.
• Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.

Thank You

Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[njhill]

Thanks @nFunctor, this looks good but could you add some test(s) for it too?

It would probably make most sense to wait for #9296 to be merged first.

[njhill]

Was this a bug?

Better to move it a couple of lines up and use if not new_beams?

[nFunctor]

I'll be honest I got confused by the similar check in llm.py and decided to do something similar here. Not sure if it is needed, and we can do as you say, or we can leave it outside altogether.

[njhill]

OK thanks, I guess we should figure out which is correct based on the prior/expected behavior...

[nFunctor]

@njhill thanks for your input, yes, I'll wait. I've decided to clarify the eos question in the mentioned PR but maybe that's going sideways a bit too much. I'll add tests once the beam search methods were merged.

[youkaichao]

yes please merge after #9296 .

I will hand it over to @njhill for review.

[nFunctor]

@njhill hello, I've made some changes post-merge. The PR is updated to reflect that. I honestly don't know if as is this is acceptable -- we can try thinking on excluding the tokens or strings from the output, but as I said this is how it was currently done with eos token anyway.

[nFunctor]

here and below, not sure if informing the beam_search_params of ignore_eos does anything, since it did not produce the "stop" finish reason as I expected.

[nFunctor]

Ok after some more testing this is actually necessary -- otherwise the eos ignoring is not happening fully.

[nFunctor]

The stop logic seems a bit convoluted -- but I think there is an issue with eos handling coming from .generate and its analogs. If you remove ignore_eos from sampling params and do beam search with ignore_eos=True, the stop will occur at the second appearance of the eos token.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @nFunctor.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[nFunctor]

This PR is too away from the current branch so I'll close it. Not exactly sure that all the beam search issues were addressed, so maybe a new PR might be needed.