[Bugfix] Correct prompt lengths for timed_traces benchmark

[s3woz]

Purpose

#39795 introduced timed_traces support for vllm bench serve. The traces look as follows:

{"timestamp": 0, "input_length": 6758, "output_length": 500, "hash_ids": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]}
{"timestamp": 0, "input_length": 7322, "output_length": 490, "hash_ids": [0, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27]}
{"timestamp": 0, "input_length": 7236, "output_length": 794, "hash_ids": [0, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41]}
{"timestamp": 0, "input_length": 2290, "output_length": 316, "hash_ids": [0, 42, 43, 44, 45]}

Current implementation creates prompts with these input_lengths, and then runs:

prompt = tokenizer.decode(prompt_ids)
[...]
samples.append(
    SampleRequest(
        prompt=prompt,
        prompt_len=prompt_len,

vLLM engine on the server side runs:

prompt_idx = tokenizer.encode(prompt)

Unfortunately, tokenizer logic is not idempotent, so whereas the client generates requests of length 6758, 7322, 7236, 2290, etc., the server receives requests of different length 7253, 7844, 7679, 2445, etc.

Solution: Avoid calling .decode and .encode
The OpenAI API and vLLM engine support direct tokens id passing, see:
vllm/entrypoints/openai/completion/protocol.py

class CompletionRequest(OpenAIBaseModel):
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/completions/create
    model: str | None = None
    prompt: (
        list[Annotated[int, Field(ge=0)]]
        | list[list[Annotated[int, Field(ge=0)]]]
        | str
        | list[str]
        | None
    ) = None

This PR avoids the .encode() to directly pass token ids, maintaining correct prompt lengths on the server side.

Test Plan

Just run any of the traces sequentially (--max-concurrency 1) :

vllm bench serve \
 --model Qwen/Qwen3.5-0.8B \
 --dataset-name timed_trace \
 --num-prompts 50 \
 --dataset-path conversation_trace_optimal_store_and_prefix_match.jsonl \
 --base-url http://localhost:12305 \
 --ignore-eos \
 --self-timed \
 --timed-trace-label-timestamp timestamp \
 --timed-trace-chunk-hash-size 512 \
 --timed-trace-label-hash-ids hash_ids \
 --timed-trace-label-input-length input_length \
 --timed-trace-label-output-length output_length \
 --timed-trace-sec-multiplier 0.001 --max-concurrency 1

• vLLM serve in debug mode:

vllm serve --port 12305 --enable-prefix-caching --model Qwen/Qwen3.5-0.8B

and observe request.num_tokens in vllm/v1/core/sched/scheduler.py::Scheduler.schedule

Test Result

Json file input_lengths: 6758, 7322, 7236, 2290, ...

Main:
For consecutive scheduling rounds, request.num_tokens=7253, 7844, 7679, 2445,...

This PR:
For consecutive scheduling rounds, request.num_tokens=6758, 7322, 7236, 2290, ...
-- as expected

@animeshtrivedi @tdoublep

[tdoublep]

LGTM

[tdoublep]

/ci run

[github-actions]

✅ @s3woz, CI is now available for this PR.

• /ci run starts a CI build.
• /ci retry retries failed jobs in the CI build for the current PR head. If the current head has no CI build, it starts a new CI build for the current head containing only jobs that failed in the latest earlier CI build for this PR.
• /ci cancel cancels scheduled or running CI builds for this PR branch.

[github-actions]

✅ Triggered Buildkite CI #83725 for commit d675770f521f.

[tdoublep]

/ci run

[github-actions]

✅ Triggered Buildkite CI #83790 for commit 9ff854152ae4.