Space-filling sweeps (Sobol, Latin Hypercube)

When grid is too dense and adaptive search is too goal-directed, quasi-Monte-Carlo (QMC) sweeps draw a fixed budget of variations that cover the parameter space evenly. Use them for characterization — getting a representative scatter you can plot — rather than optimization.

When to use which

Default to Sobol unless you have all-discrete dims or want exact per-bin balance.

Use case 1 — Perf surface across realistic input shapes

1
sweep:
2
  type: sobol
3
  samples: 64
4
  seed: 42
5
  dimensions:
6
    - {path: concurrency,                  lo: 1,   hi: 256,   scale: log, kind: int}
7
    - {path: datasets.default.prompts.isl, lo: 128, hi: 32768, scale: log, kind: int}
8
    - {path: datasets.default.prompts.osl, lo: 16,  hi: 4096,  scale: log, kind: int}
9
    - {path: rate,                         lo: 1,   hi: 100,   scale: log, kind: real}
10
multi_run:
11
  num_runs: 3

64 samples × 3 trials = 192 cells. A 5⁴ grid would be 625 cells — you save ~70% with comparable scatter coverage.

Use case 2 — A/B build comparison on a fixed workload distribution

Run identical YAML on build A, then build B. The same seed makes Sobol pick identical 32 points both runs, giving you paired comparisons — much better variance than independent sweeps.

1
sweep:
2
  type: sobol
3
  samples: 32
4
  seed: 42
5
  scramble: true
6
  dimensions:
7
    - {path: concurrency,                  lo: 1,   hi: 128,  scale: log, kind: int}
8
    - {path: datasets.default.prompts.isl, lo: 256, hi: 8192, scale: log, kind: int}
9
    - {path: datasets.default.prompts.osl, lo: 64,  hi: 2048, scale: log, kind: int}

Use case 3 — Latin Hypercube on discrete dims

LHS guarantees each model appears samples / len(choices) times — its niche over Sobol when all dims are discrete.

1
sweep:
2
  type: latin_hypercube
3
  samples: 12
4
  seed: 7
5
  dimensions:
6
    - path: model
7
      choices:
8
        - meta-llama/Llama-3.1-8B-Instruct
9
        - mistralai/Mixtral-8x7B-Instruct-v0.1
10
        - Qwen/Qwen2-72B
11
    - path: concurrency
12
      choices: [1, 4, 16, 64]
13
    - path: datasets.default.prompts.isl
14
      choices: [256, 1024, 4096]

Reproducibility

Every QMC run writes <artifacts>/sweep_aggregate/sampling_design.json containing the mapped sample values (samples_mapped), the dimension specs, the sampler type/seed, and sampler options (scramble for Sobol, optimization for LHS). Re-runs with the same seed and scipy version produce byte-identical points. Pin scipy in pyproject.toml if you need cross-machine reproducibility over time.

Edge cases

See also

• Bayesian-Optimization Outer Loop — when you want optimization, not characterization
• Bayesian Optimization — 1D SLA saturation — single-dim ramp recipe (max-concurrency-under-sla / max-goodput-under-slo)
• Search Recipes — high-level recipes that build sweeps from intent