[Benchmark] guided decoding

[aarnphm]

Initial benchmark code for guided decoding to trace performance differences.

See also slack thread

tldr:

Setup

vllm version: 0.6.3.dev236+g48138a84
GPU: A100 80GB
args: --async-engine --model meta-llama/Llama-3.2-3B-Instruct --output-len [2048|4096] [--guided-decoding] [--warmup]

The below is ran with 100 prompts versus 95 prompts with 5 warmup. It seems that current implementation shows around 24% degradation in TPS when outlines is being used.

Reference:

• [Bugfix] Adds outlines performance improvement #5006
• [RFC]: Improve guided decoding (logit_processor) APIs and performance. #5423

• vllm/vllm/model_executor/layers/logits_processor.py

  Line 112 in 246598a

  logits_row = logits_processor(prompt_tokens_ids,

---

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.

🚀

[mgoin]

Could you replace most of these duplicated arguments with the engine's args, like we use in other benchmark scripts?

vllm/benchmarks/benchmark_throughput.py

Line 412 in ac49b59

parser = AsyncEngineArgs.add_cli_args(parser)

[mgoin]

Here is my suggested version. It allows for you to specify the backend on the CLI too, through --guided-decoding-backend outlines

"""Benchmark guided decoding throughput."""
import argparse
import json
import random
import time
from typing import List

import uvloop
from transformers import AutoTokenizer

from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
from vllm.entrypoints.openai.api_server import (
    build_async_engine_client_from_engine_args)
from vllm.sampling_params import GuidedDecodingParams
from vllm.utils import FlexibleArgumentParser, merge_async_iterators

SCHEMA = {
    "$schema":
    "https://json-schema.org/draft/2020-12/schema",
    "title":
    "User Profile",
    "type":
    "object",
    "properties": {
        "userId": {
            "type": "string",
            "description": "Unique identifier for the user."
        },
        "personalInfo": {
            "type": "object",
            "properties": {
                "firstName": {
                    "type": "string",
                    "description": "The user's first name."
                },
                "lastName": {
                    "type": "string",
                    "description": "The user's last name."
                },
                "age": {
                    "type": "integer",
                    "minimum": 0,
                    "description": "The user's age."
                },
                "phoneNumbers": {
                    "type":
                    "array",
                    "items": {
                        "type": "object",
                        "properties": {
                            "type": {
                                "type": "string",
                                "enum": ["home", "work", "mobile"],
                                "description": "Type of phone number."
                            },
                            "number": {
                                "type": "string",
                                "pattern": "^\\+?[1-9]\\d{1,14}$",
                                "description": "Phone number in E.164 format."
                            }
                        },
                        "required": ["type", "number"]
                    },
                    "description":
                    "List of phone numbers associated with the user."
                }
            },
            "required": ["firstName", "lastName"]
        },
        "address": {
            "type": "object",
            "properties": {
                "street": {
                    "type": "string",
                    "description": "Street address."
                },
                "city": {
                    "type": "string",
                    "description": "City name."
                },
                "state": {
                    "type": "string",
                    "description": "State or province."
                },
                "postalCode": {
                    "type": "string",
                    "pattern": "^\\d{5}(-\\d{4})?$",
                    "description": "Postal code."
                },
                "country": {
                    "type": "string",
                    "description": "Country name."
                }
            },
            "required": ["street", "city", "state", "postalCode", "country"]
        },
        "preferences": {
            "type": "object",
            "properties": {
                "newsletterSubscribed": {
                    "type":
                    "boolean",
                    "description":
                    "Indicates if the user is subscribed to the newsletter."
                },
                "favoriteCategories": {
                    "type": "array",
                    "items": {
                        "type": "string"
                    },
                    "description": "List of user's favorite categories."
                }
            },
            "required": ["newsletterSubscribed"]
        },
        "accountStatus": {
            "type": "string",
            "enum": ["active", "inactive", "suspended"],
            "description": "Current status of the user's account."
        },
        "registrationDate": {
            "type": "string",
            "format": "date-time",
            "description": "ISO 8601 formatted date-time of user registration."
        }
    },
    "required":
    ["userId", "personalInfo", "address", "accountStatus", "registrationDate"]
}

def run_vllm(
    requests: List[tuple[str, int, int]],
    engine_args: EngineArgs,
    n: int,
    guided_decoding: bool = False,
    warmup: bool = False,
) -> float:
    from vllm import LLM, SamplingParams
    llm = LLM(**vars(engine_args))

    # Add the requests to the engine.
    prompts: List[str] = []
    sampling_params: List[SamplingParams] = []
    for prompt, _, output_len in requests:
        prompts.append(prompt)
        sampling_params.append(
            SamplingParams(
                n=n,
                temperature=1.0,
                top_p=1.0,
                ignore_eos=True,
                max_tokens=output_len,
                guided_decoding=GuidedDecodingParams(
                    json=SCHEMA) if guided_decoding else None,
            ))

    start = time.perf_counter()
    llm.generate(prompts, sampling_params, use_tqdm=True)
    end = time.perf_counter()
    return end - start

async def run_vllm_async(
    requests: List[tuple[str, int, int]],
    engine_args: AsyncEngineArgs,
    n: int,
    guided_decoding: bool = False,
    warmup: bool = False,
    disable_frontend_multiprocessing: bool = False,
) -> float:
    from vllm import SamplingParams

    async with build_async_engine_client_from_engine_args(
            engine_args, disable_frontend_multiprocessing) as llm:

        # Add the requests to the engine.
        prompts: List[str] = []
        sampling_params: List[SamplingParams] = []
        if warmup:
            print("Running warmup prompt, for the first 5")
            # We setup the first 5 requests to warmup FSM
            warmup_requests = requests[:5]
            requests = requests[5:]
            for prompt, _, output_len in warmup_requests:
                prompts.append(prompt)
                sampling_params.append(
                    SamplingParams(
                        n=n,
                        temperature=1.0,
                        top_p=1.0,
                        ignore_eos=True,
                        max_tokens=output_len,
                        guided_decoding=GuidedDecodingParams(
                            json=SCHEMA) if guided_decoding else None,
                    ))
            generators = []
            for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
                generator = llm.generate(prompt, sp, request_id=f"test{i}")
                generators.append(generator)
            all_gens = merge_async_iterators(*generators)
            async for i, res in all_gens:
                pass

        prompts = []
        sampling_params = []
        for prompt, _, output_len in requests:
            prompts.append(prompt)
            sampling_params.append(
                SamplingParams(
                    n=n,
                    temperature=1.0,
                    top_p=1.0,
                    ignore_eos=True,
                    max_tokens=output_len,
                    guided_decoding=GuidedDecodingParams(
                        json=SCHEMA) if guided_decoding else None,
                ))

        generators = []
        start = time.perf_counter()
        for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
            generator = llm.generate(prompt, sp, request_id=f"test{i}")
            generators.append(generator)
        all_gens = merge_async_iterators(*generators)
        async for i, res in all_gens:
            pass
        end = time.perf_counter()
        return end - start

def main(args: argparse.Namespace):
    print(args)
    random.seed(args.seed)

    # Synthesize a prompt with the given input length.
    tokenizer = AutoTokenizer.from_pretrained(
        args.tokenizer, trust_remote_code=args.trust_remote_code)
    prompt = f"Generate an example of a user profile given the following schema: {json.dumps(SCHEMA)}"
    input_len = len(tokenizer(prompt).input_ids)
    print(f"Input length of the prompt: {input_len} tokens")
    requests = [(prompt, input_len, args.output_len)
                for _ in range(args.num_prompts)]

    if args.async_engine:
        engine_args = AsyncEngineArgs.from_cli_args(args)
        elapsed_time = uvloop.run(
            run_vllm_async(
                requests,
                engine_args,
                args.n,
                args.guided_decoding,
                args.warmup,
                args.disable_frontend_multiprocessing,
            ))
    else:
        engine_args = EngineArgs.from_cli_args(args)
        elapsed_time = run_vllm(
            requests,
            engine_args,
            args.n,
            args.guided_decoding,
            args.warmup,
        )

    total_num_tokens = sum(prompt_len + output_len
                           for _, prompt_len, output_len in requests)
    total_output_tokens = sum(output_len
                              for _, _, output_len in requests)
    print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
          f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
          f"{total_output_tokens / elapsed_time:.2f} output tokens/s")

    # Output JSON results if specified
    if args.output_json:
        results = {
            "elapsed_time": elapsed_time,
            "num_requests": len(requests),
            "total_num_tokens": total_num_tokens,
            "requests_per_second": len(requests) / elapsed_time,
            "tokens_per_second": total_num_tokens / elapsed_time,
        }
        with open(args.output_json, "w") as f:
            json.dump(results, f, indent=4)

if __name__ == "__main__":
    parser = FlexibleArgumentParser(description="Benchmark guided decoding.")
    parser = AsyncEngineArgs.add_cli_args(parser)
    
    parser.add_argument("--output-len",
                        type=int,
                        help="Output length for each request. Overrides the "
                        "output length from the dataset.")
    parser.add_argument("--n",
                        type=int,
                        default=1,
                        help="Number of generated sequences per prompt.")
    parser.add_argument("--num-prompts",
                        type=int,
                        default=10,
                        help="Number of prompts to process.")
    parser.add_argument('--output-json',
                        type=str,
                        default=None,
                        help='Path to save the throughput results in JSON format.')
    parser.add_argument("--async-engine",
                        action='store_true',
                        default=False,
                        help="Use vLLM async engine rather than LLM class.")
    parser.add_argument("--guided-decoding",
                        action='store_true',
                        default=False,
                        help="Whether to enable JSON decoding or not.")
    parser.add_argument("--disable-frontend-multiprocessing",
                        action='store_true',
                        default=False,
                        help="Disable decoupled async engine frontend.")
    parser.add_argument("--warmup",
                        action="store_true",
                        default=False,
                        help="Run warmup prompts before benchmark.")
    args = parser.parse_args()
    if args.tokenizer is None:
        args.tokenizer = args.model
    main(args)

[xuechendi]

how about adding a 'result_file_name' with --save-results in argument list for result accuracy check

result_file_name = f"{'guided' if args.guided_decoding else 'no_guided'}_{args.num_prompts}_out{args.output_len}_async{args.async_engine}_warmup{args.warmup}.json"

outputs = llm.generate(prompts, sampling_params, use_tqdm=True)
ret = []
for output in outputs:
    generated_text = output.outputs[0].text
    ret.append(generated_text)
    
# save ret list into a json
if result_file_name:
    with open(result_file_name, 'w') as f:
    json.dump(ret, f, indent=4)

[xuechendi]

@simon-mo I started to continue work on this PR: https://github.com/vllm-project/vllm/pull/10557/commits

[aarnphm]

Superceded by #10557