Skip to content

[Usage] Per-strategy max_requests constraints not respected in Sweep Mode #451

@AiKiAi-stack

Description

@AiKiAi-stack

Describe the bug
When using Sweep Mode with a custom profile that specifies a list of max_requests values under constraints, all sub-strategies appear to use the same max_requests value instead of the respective value from the list.

Purpose / Expected Behavior
I would like to assign a different max_requests value to each strategy generated during a sweep (e.g., for fine-grained control over benchmark duration or resource usage per point in the sweep).

Environment

  • GuideLLM version: 0.4.0-dev
  • Python version: 3.12
  • Backend: vLLM

To Reproduce

  1. Clone and install GuideLLM:

    git clone https://github.com/vllm-project/guidellm.git
    cd guidellm
    pip install -e .
  2. Create a profile file myProfile.json:

    {
      "profile": {
        "type_": "sweep",
        "sweep_size": 5,
        "constraints": {
          "max_requests": [10, 100, 100, 100, 100]
        }
      },
      "target": "http://localhost:8000",
      "data": [
        "prompt_tokens=512,prompt_tokens_stdev=128,prompt_tokens_min=1,prompt_tokens_max=1024,output_tokens=256,output_tokens_stdev=64,output_tokens_min=1,output_tokens_max=1024"
      ],
      "output_formats": ["benchmarks_json.html"]
    }
  3. Run the benchmark:

    guidellm benchmark run --scenario src/guidellm/benchmark/scenarios/myProfile.json

Observed Behavior
All five sweep strategies show the same max_requests value (all use 10), as shown in the output table:

Sweep strategies all using same max_requests

This suggests the per-strategy constraint in constraints.max_requests is not being applied correctly.

Additional Context

  • Is the constraints field documented for per-strategy configuration in sweep mode?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions