InferScale-Sim / tests /test_measurements.py
ArchitSharma's picture
Finalize InferScale research consolidation
9916edb
Raw History Blame Contribute Delete
2.21 kB
import json
from inferscale.measurements import calibrate_measurements, import_measurements
BASE = {
"model": "Qwen2.5-3B",
"accelerator": "L4",
"quantization": "int8",
"scheduler": "continuous_fcfs",
"arrival_process": "poisson",
"duration_s": 20,
"request_rate_rps": 2,
"prompt_tokens_mean": 256,
"output_tokens_mean": 64,
"prompt_tokens_cv": 0,
"output_tokens_cv": 0,
}
def test_sglang_jsonl_import():
content = json.dumps(
{
"backend": "sglang",
"request_rate": 2,
"random_input_len": 256,
"random_output_len": 64,
"request_throughput": 1.9,
"p95_ttft_ms": 210.0,
"p95_e2e_latency_ms": 2200.0,
}
)
result = import_measurements(content, "auto", BASE)
assert result["case_count"] == 1
case = result["cases"][0]
assert case["measurement_metadata"]["source"] == "sglang"
assert case["measured"]["p95_ttft_ms"] == 210.0
def test_vllm_percentile_pair_import():
content = json.dumps(
{
"label": "vllm",
"request_rate": 2,
"input_len": 256,
"output_len": 64,
"request_goodput": 1.8,
"percentiles_ttft_ms": [[95, 190.0], [99, 240.0]],
"percentiles_e2el_ms": [[95, 2050.0], [99, 2300.0]],
}
)
result = import_measurements(content, "auto", BASE)
measured = result["cases"][0]["measured"]
assert measured["p95_ttft_ms"] == 190.0
assert measured["p99_e2e_ms"] == 2300.0
def test_calibration_uses_heldout_when_enough_cases():
cases = []
for idx in range(4):
cases.append(
{
"name": f"case-{idx}",
"config": BASE | {"request_rate_rps": 1.0 + idx * 0.25, "seed": 10 + idx},
"measured": {"p95_ttft_ms": 220.0 + idx * 10, "p95_e2e_ms": 2200.0 + idx * 50},
}
)
result = calibrate_measurements(cases, holdout_fraction=0.25, seed=3)
assert result["validation_mode"] == "held-out"
assert result["train_count"] == 3
assert result["holdout_count"] == 1
assert result["fitted_scales"]["prefill_time_scale"] > 0