szl-engine-bench

An honest, standard-library-only benchmark client for OpenAI-compatible streaming endpoints exposed by vLLM, SGLang, llama.cpp, MLX, TGI, and Transformers Serve.

Version 0.3.0 reports separate distributions for time to first token/chunk (TTFT), inter-chunk latency (ITL), end-to-end latency, and the fraction of runs meeting a declared TTFT service-level objective. It does not collapse an engine comparison into one winner.

The distinction is deliberate. SGLang's serving guide reports TTFT, TPOT, ITL, and throughput as separate online-serving metrics, while vLLM Bench separately exposes TTFT/ITL percentiles and SLO goodput. Cache shape, request rate, concurrency, model, and hardware can all change the result; a benchmark from a different workload is not evidence for this one.

• SGLang benchmark and profiling guide (source snapshot)
• vLLM Bench metrics (source snapshot)

States and measurement semantics

• MEASURED: every requested run produced at least two non-empty streaming response chunks and valid monotonic arrival times.
• BLOCKED: the engine endpoint was not configured, or fewer than two engines produced measurements for comparison.
• INVALID: configuration, stream shape, or fairness inputs were unsuitable for the requested metric.
• FAILED: the configured endpoint or streaming protocol actually failed.

ITL is measured between non-empty SSE response chunks. An SSE chunk is not assumed to equal one model token. chunks_per_s is always labeled as such; the request asks the server for streamed usage with stream_options.include_usage=true, and tok_per_s is UNAVAILABLE unless the endpoint actually supplies an integer usage.completion_tokens value. This avoids turning packet counts into fake token throughput.

The comparison keeps the v0.1.0 fairness gates: measured engines must use the same model string and run count for low-level numerical diagnostics. Version 0.3.0 additionally requires the declared identity contract below for CLI comparisons; the old model-name check alone is explicitly unqualified. goodput_at_slo is the auditable fraction of runs whose TTFT is less than or equal to --slo-ttft-ms; the boundary is inclusive. It is not claimed to be a maximum sustainable request rate.

Install and verify

python -m pip install -e . pytest
python -m pytest tests -q
python -m compileall -q engine_bench.py benchmark_manifest.py tests
python engine_bench.py --engines vllm --runs 1

The last command is a fail-closed smoke test: without a declaration it emits a BLOCKED result and a valid receipt, even when an endpoint is configured. It issues no network request. No model or GPU is needed for the test suite.

The deterministic burst fixture from issue #2 can be reproduced without an endpoint:

python -c "import engine_bench as e; print(e.run_stats([.05,.06,.07,.4,.45,.5,.55,.6], 8))"

Its TTFT is 50 ms. The inter-chunk gaps are 10, 10, 330, 50, 50, 50, and 50 ms, so linear interpolation yields ITL p50/p95/p99 of 50, 246, and 313.2 ms. This is a unit fixture, not a hardware benchmark.

Exact experiment declarations (v0.3.0)

The default CLI requires --manifest comparison.json before contacting any configured engine. The complete selected cohort is checked first: a missing or changed endpoint, extra engine, mismatched request configuration, invalid digest, or different hardware declaration cannot become a successful partial comparison.

The manifest uses the closed szl.engine-comparison/v1 contract. Its template is examples/comparison.synthetic.json. Every identity in that file is a synthetic test fixture, not a qualified model, hardware profile, or live endpoint. Populate real operator-reviewed declarations before measuring a deployment. The loader never downloads or executes a manifest reference and rejects duplicate JSON keys, non-finite values, oversized inputs, unknown fields, control characters, mutable revision names, and newline-suffixed hashes. SHA checks use full-string matching, not permissive end anchors.

benchmark_manifest.endpoint_digest() computes the endpoint digest. File digests refer to exact bytes; they are not a cryptographic demonstration that a remote server loaded those bytes. Engine names must match the supported registry. The client runs serially, performs no warmup/reset, and requires concurrency: 1 and cache_state: UNCONTROLLED; comparing cold and warm caches is not silently claimed. It freezes the preflight endpoints and refuses redirects for bound requests.

Every sample commits the exact serialized request bytes. A run binding includes the experiment, model/artifact subject, workload, per-engine build/configuration, hardware, endpoint, and request hashes. compare_bound() rejects altered bindings, missing samples, different wire requests, and omitted/failed cohort members. No failed engine is discarded to obtain a better-looking comparison.

Identity evidence remains OPERATOR_DECLARED_NOT_RUNTIME_ATTESTED. A matching manifest is not server identity attestation. Comparison reports keep runtime_identity: NOT_ATTESTED, quality_evaluation: NOT_PERFORMED, production_admission: NOT_EVALUATED, and uncontrolled cache state. They measure client-observed latency only and never decide model promotion or an engine winner. Different hardware should be evaluated as a separately designed system experiment, not misrepresented as an engine-only improvement.

An incomplete or invalid explicit --manifest invocation exits 2 after emitting its available receipt. The default offline smoke command retains its historical zero exit. Library run_engine(), compare(), and compare_engines() remain available for numerical diagnostics, labeled UNBOUND_METRICS_ONLY in comparisons. The CLI requires --unbound-diagnostics to explicitly opt into that legacy path; it cannot be combined with --manifest and is not release admission evidence.

Measure configured engines

PowerShell:

$env:VLLM_ENDPOINT = "http://127.0.0.1:8000"
$env:SGLANG_ENDPOINT = "http://127.0.0.1:30000"
python engine_bench.py --engines vllm sglang --model YOUR_MODEL --runs 10 --max-tokens 64 --slo-ttft-ms 200 --manifest comparison.json

POSIX shells:

export VLLM_ENDPOINT=http://127.0.0.1:8000
export SGLANG_ENDPOINT=http://127.0.0.1:30000
python engine_bench.py --engines vllm sglang --model YOUR_MODEL --runs 10 --max-tokens 64 --slo-ttft-ms 200 --manifest comparison.json

Supported variables are VLLM_ENDPOINT, SGLANG_ENDPOINT, LLAMACPP_ENDPOINT, MLX_ENDPOINT, TGI_ENDPOINT, and TRANSFORMERS_ENDPOINT. Each must be an absolute HTTP(S) base URL without embedded credentials, query parameters, fragments, or an invalid TCP port. Duplicate names in --engines are rejected as INVALID before any request, so receipt maps cannot silently overwrite a result. The client posts to /v1/completions.

Receipts

Every CLI invocation emits one canonical-JSON SHA-256 chain receipt using the repository's prev_hash / self_hash schema and the all-zero genesis hash. The receipt anchors:

• a canonical SHA-256 of every complete engine result;
• a canonical SHA-256 of each measured engine's nested ITL gap arrays;
• the ordered engine selection, benchmark configuration, and a prompt hash; and
• the comparison verdict and every engine state; and
• the canonical manifest digest and each exact wire-request/run binding (schema version 3).

Each run sample also includes its raw itl_gaps_ms array and matching itl_gaps_sha256, so the hash can be independently recomputed. UNSIGNED_HONEST proves integrity and order only; it does not claim signer identity or external attestation.

Verification scope

The test suite covers strict declarations, complete-cohort comparisons, actual loopback HTTP request bytes/sampling fields, redirect refusal, preflight zero- request behavior, endpoint freezing, bad JSON/digests, failure retention, and receipt commitments, in addition to the original streaming and numerical tests. Loopback test responses are fixtures: no trained model, remote production system, GPU performance, energy measurement, or quality result is represented by CI.

Apache-2.0 · Doctrine v11 · SZL Holdings