llm-serving-auto-benchmark（LLM serving auto benchmark 搜索）

LLM Serving Auto Benchmark

Overview

Use this skill to compare SGLang, vLLM, and TensorRT-LLM for the same model and workload.

Use a config-driven workflow:

• keep launch-only capacity choices in each framework's base_server_flags
• put the search knobs in search_space
• run the same dataset scenarios for every framework
• generate a bounded candidate list from search_space, with the baseline candidate included first
• keep failed candidates in the result file
• pick the best SLA-passing candidate after normalizing the results

For model-specific starting points, prefer the shipped configs in configs/cookbook-llm/. They reuse the SGLang auto-benchmark cookbook model set and translate it into framework-native SGLang, vLLM, and TensorRT-LLM server flags. Validate those configs before a real run:

python skills/llm-serving-auto-benchmark/scripts/validate_cookbook_configs.py \
  skills/llm-serving-auto-benchmark/configs/cookbook-llm

If you have captured target-environment --help files, add --help-dir <artifact-help-dir>. That check only loads configs, verifies the server flag names, and renders candidate commands; it does not launch model servers.

Prefer native tooling when it gives better coverage:

• SGLang: python -m sglang.auto_benchmark when available, otherwise python -m sglang.bench_serving
• vLLM: vllm bench sweep serve for server-parameter sweeps, otherwise vllm serve plus vllm bench serve
• TensorRT-LLM: trtllm-serve for the OpenAI-compatible server plus the TensorRT-LLM serving benchmark client or a common OpenAI-compatible benchmark client

TensorRT-LLM has one hard scope rule in this skill: the server backend is fixed to trtllm-serve serve --backend pytorch. Do not search TensorRT-LLM backend choice. If a request, config, or candidate asks for trt, an engine backend, or any other non-PyTorch TensorRT-LLM server backend, reject that candidate as unsupported for this skill and record the reason. This does not change the benchmark client backend; the TensorRT-LLM benchmark client still uses OpenAI-compatible modes such as --backend openai or --backend openai-chat.

Only pick a winner after each requested framework has had its main serving knobs tuned.

The parameter lists in this skill are not a compatibility contract. They are version-sensitive candidate knob families. Before every real run, record the exact framework version or git commit and verify the concrete CLI flag names with --help in the target environment.

The default search style should stay close to SGLang auto benchmark: start from a mostly pure-TP baseline, sweep a small set of high-impact runtime knobs, and cap the first pass around 10 candidates per framework. Do not search memory fractions by default.

Validation Environment

This skill is target-agnostic. It assumes any one of the following is available, and nothing more:

• a local GPU host with Docker/Podman and the target framework images pulled;
• a remote GPU host reached via ssh <host> with the framework images already running in a container there;
• a CI runner that can exec into a pre-built image for each framework.

Do not assume a specific operator host name (h100_sglang, b200_*, radixark*, rtx5090_*, etc.) inside this skill's own workflow. The concrete SSH wiring, container names, workspace paths, and HF token plumbing for a given box live in the operator-side per-host skills (for example h100, h100-sglang-diffusion, b200, rtx5090, radixark02, radixark03); this skill only requires that the caller can reach a shell inside a container with sglang, vllm, or tensorrt_llm installed.

Historical validation snapshots in references/ (for example the H100-recorded parameter audit) are evidence of which flag names and failure modes were seen in a specific image and are not a requirement that the next run happens on the same hardware.

Skill Scope

This skill is a playbook plus a config+validator toolchain, not a turn-key orchestrator. The scripts/ directory contains exactly two tools:

• validate_cookbook_configs.py: reads YAML, renders bounded candidate server commands, and checks flag names against captured --help snapshots. It never launches a model server.
• compare_benchmark_results.py: takes the normalized per-candidate JSONL and emits the markdown tables described in the Output Contract.

Launching servers, driving the workload, and writing one JSONL row per candidate are the operator's responsibility; the skill tells you how to do them, and the validator keeps your inputs honest.

The cookbook configs under configs/cookbook-llm/ and the sample runtime plan at references/example-plan.yaml use related but not identical schemas:

• Cookbook configs carry schema_version: 1, source.kind, benchmark.sla (nested), and frameworks.*.server_command; they must pass validate_cookbook_configs.py.
• example-plan.yaml is a shorter runtime plan shape with top-level sla and no server_command. It is the skeleton a caller fills in for a one-off run and is not expected to pass the cookbook validator as-is.

Either shape can feed a benchmark run; the SLA key names in references/result-schema.md are the single source of truth.

Required Inputs

Collect these before starting a long run:

• model path or Hugging Face repo id
• tokenizer path if it differs from the model
• target frameworks: any subset of sglang, vllm, tensorrt-llm
• GPU model, GPU count, and whether multi-node is allowed
• precision and quantization constraints
• endpoint shape: completions, chat completions, responses, or custom
• workload source: real traffic JSONL, ShareGPT, random synthetic, or generated shared-prefix synthetic
• dataset scenarios when synthetic traffic is used, for example chat and summarization
• SLA target: TTFT, TPOT/ITL, end-to-end latency, success rate, or goodput
• search budget: quick smoke, default search, or exhaustive search
• output directory for logs and result artifacts

Also collect a version manifest:

• framework package version and git commit when available
• container image or Python environment identifier
• --help snapshots for the server command and benchmark command
• whether each parameter in the search plan was accepted by that exact CLI

If real production traffic is the goal, use the real request distribution. A synthetic workload is fine for bring-up and first-pass comparison, but it is not enough for a production choice.

Known Gotchas

Short list of failure modes that have bitten past validation runs. Check these before starting a long sweep.

• SGLang fa3 attention backends need Hopper or newer. On A100, L40S, RTX 5090, and older GPUs, drop fa3 from the SGLang search_space and keep flashinfer (or triton when FlashInfer is unavailable).
• SGLang bench_serving has two SGLang-facing backends: --backend sglang for the native /generate endpoint and --backend sglang-oai for the OpenAI-compatible endpoint. For cross-framework comparisons, prefer sglang-oai so every framework is measured on the same request path.
• vLLM --enable-dbo only works when the target vLLM image is built with a supported all2all backend. Keep DBO out of the default candidate list unless the operator has verified the image.
• vLLM --max-num-partial-prefills > 1 is model- and runtime-gated. Keep 1 in the default pass; raise only after a preflight with the actual model.
• In the validated TensorRT-LLM 1.0.0 image, trtllm-serve serve accepts --kv_cache_free_gpu_memory_fraction; the older --free_gpu_memory_fraction exits with a CLI error. Re-check the accepted flag name via --help on the target image before a real run.
• TensorRT-LLM 1.0.0 multi-GPU PyTorch-backend servers need --ipc=host, --ulimit memlock=-1, --ulimit stack=67108864, --shm-size=16g, and NCCL_IB_DISABLE=1 (for single-node) or an equivalent NCCL setup.
• TensorRT-LLM 1.0.0 benchmark client takes --backend openai or --backend openai-chat; --backend trtllm is rejected. This is separate from the server backend, which is pinned to pytorch by this skill.
• trtllm benchmark_serving --dataset-name random silently falls back to ShareGPT sampling without --random-ids (or --download-path).
• max_seq_len / max_model_len / context_length candidates must cover max(input_len + output_len) across every scenario, including values inside search_space, not just the baseline. The validator checks this; do not bypass it.

Secrets Hygiene

• Never print HF_TOKEN, HUGGINGFACE_HUB_TOKEN, or any upstream API key into a saved artifact. Pass them through container -e VAR (unquoted on the right side so the host value is inherited) and keep them out of server_command and benchmark_command fields written to the result JSONL.
• When a framework echoes the full argv at startup, scrub the log or redact token-shaped substrings before uploading the artifact.

Fairness Rules

Use these rules throughout the benchmark:

• Run every framework on the same GPU type, GPU count, model weights, tokenizer, precision, quantization policy, prompt distribution, output length target, and sampling settings.
• Record framework version, git commit, container image, CUDA/NCCL versions, GPU driver, visible GPU ids, launch command, and benchmark command.
• Warm the server before measuring. Restart or clear state between candidate configurations when cache effects would bias the comparison.
• Compare steady-state fixed-QPS runs separately from burst throughput runs.
• Keep failed candidates in the final results with their failure reason.
• Report both raw throughput and SLA-passing throughput. The fastest failing candidate is not the best deployment command.

Workflow

1. Preflight

Verify all requested frameworks before starting a search:

python -m sglang.launch_server --help
python -m sglang.bench_serving --help
vllm serve --help
vllm serve --help=all
vllm bench serve --help
vllm bench serve --help=all
vllm bench sweep serve --help=all
trtllm-serve serve --help
python -m tensorrt_llm.serve.scripts.benchmark_serving --help

Use the framework-specific --help output in the target environment as the source of truth. Do not keep a stale launch flag just because it appears in an old note.

vLLM 0.19 and newer use grouped help. Plain vllm serve --help only shows the groups, so capture --help=all before deciding whether a search knob exists.

Save these --help outputs into the run artifact directory. If a listed search knob is missing from the current CLI, remove or translate that knob before running the benchmark. Do not silently pass unknown flags.

For TensorRT-LLM, also confirm that trtllm-serve serve --help accepts --backend pytorch. If it does not, mark TensorRT-LLM unsupported in that environment rather than falling back to a different server backend.

For each framework:

1. Launch a minimal server.
2. Confirm /v1/models or the framework-native model-info endpoint works.
3. Send one streaming request and verify TTFT can be measured.
4. Run one tiny benchmark with at least 5 requests.
5. Save the launch command, benchmark command, server log, and benchmark output.

Before any GPU-backed smoke run, check the requested GPU ids directly with nvidia-smi. If a requested GPU is already in use, stop and record that fact. Do not silently borrow a different GPU count for a performance comparison. It is fine to run a smaller one-GPU smoke only when the result is clearly labeled as a flow check rather than a fair throughput comparison.

If the target environment runs through containers, follow references/container-runbook.md. Save the image tags, pull commands, launch commands, server logs, benchmark logs, and cleanup commands in the artifact directory.

2. Normalize The Workload

Use one canonical workload for all frameworks. Recommended JSONL row shape:

{"prompt": [{"role": "user", "content": "Summarize this text."}], "output_len": 256}
{"prompt": "Write a short explanation of CUDA graphs.", "output_len": 128}

Optional fields:

{
  "prompt": [{"role": "user", "content": "Use low temperature."}],
  "output_len": 256,
  "extra_request_body": {"temperature": 0.0, "top_p": 0.95},
  "metadata": {"source": "prod-sample"}
}

When converting user data:

• inspect at least 3 rows before conversion
• preserve request-level sampling options in extra_request_body
• do not include the final assistant answer in the prompt when that answer is the target completion
• keep multimodal or tool-call payloads only if all requested frameworks support the chosen endpoint shape

For synthetic bring-up, follow the two-scenario shape used by the SGLang auto benchmark references:

dataset:
  kind: random
  num_prompts: 80
  scenario_names: [chat, summarization]
  input_len: [1000, 8000]
  output_len: [1000, 1000]

Each aligned input_len / output_len pair is one scenario. Do not take the cartesian product unless the user asks for that.

Before searching any sequence-length limit, compute the largest input_len + output_len in the dataset. SGLang context_length, vLLM max_model_len, and TensorRT-LLM max_seq_len must be at least that value for every candidate that is expected to run all scenarios.

3. Pick A Search Tier

Use the smallest tier that can answer the user's question:

• Tier 1: smoke and sanity. One baseline plus a few high-impact knobs.
• Tier 2: default. A bounded sweep over the most likely server settings.
• Tier 3: exhaustive. Only when the search space is already tight and the user accepts a long run.

Default budget:

• num_prompts: 80 for the default cross-framework comparison; num_prompts: 20 per scenario is acceptable for a smoke/flow check and must be labeled as such in the artifact (not as a performance result).
• search.max_candidates_per_framework: 10 for the first useful pass
• candidate generation: baseline first, then a bounded product or ordered candidate list from search_space
• at most 5 QPS search rounds unless the user asks for more
• stop early when every candidate in one framework is clearly OOM or fails the basic health check

Keep these in base_server_flags unless the user specifically wants a capacity or memory study:

• SGLang mem_fraction_static
• SGLang schedule_policy
• vLLM gpu_memory_utilization
• TensorRT-LLM kv_cache_free_gpu_memory_fraction

These are real knobs, but they widen the search quickly and often turn a serving comparison into a memory-limit study.

4. Tune SGLang

Prefer the SGLang auto-benchmark runner when the target checkout supports it:

python -m sglang.auto_benchmark run --config /path/to/sglang.yaml

Otherwise launch the server manually and benchmark with:

python -m sglang.bench_serving \
  --backend sglang \
  --dataset-name random \
  --random-input-len 1024 \
  --random-output-len 256 \
  --num-prompts 80 \
  --request-rate 8 \
  --output-file /path/to/sglang/results.json \
  --output-details

Version-sensitive SGLang knob families to verify:

• tp_size, pp_size, dp_size, ep_size
• attention_backend, prefill_attention_backend, decode_attention_backend
• sampling_backend
• max_running_requests, max_queued_requests
• chunked_prefill_size, prefill_max_requests, max_prefill_tokens
• max_total_tokens, page_size
• CUDA graph and piecewise CUDA graph settings
• speculative or EAGLE settings only after the non-speculative baseline is tuned

Keep mem_fraction_static and schedule_policy pinned in the default pass, matching the SGLang auto benchmark cookbook style.

For quick smoke tests, it is reasonable to disable CUDA graph and piecewise CUDA graph startup work if the goal is only to prove the framework flow. Record those flags in the artifact. Do not carry that smoke setting into a performance winner unless the user asked to tune eager-mode serving.

5. Tune vLLM

Use vLLM's sweep runner when available:

vllm bench sweep serve \
  --serve-cmd 'vllm serve <model> --port 8000' \
  --bench-cmd 'vllm bench serve --backend vllm --model <model> --port 8000 --dataset-name random --num-prompts 80' \
  --serve-params /path/to/vllm_serve_params.json \
  --bench-params /path/to/vllm_bench_params.json \
  --output-dir /path/to/vllm_results

If sweep support is unavailable, run vllm serve for each candidate and measure with vllm bench serve.

Version-sensitive vLLM knob families to verify:

• tensor, pipeline, data, decode-context, and expert parallelism
• gpu_memory_utilization
• max_num_seqs
• max_num_batched_tokens
• max_model_len
• enable_chunked_prefill, partial prefill limits, and DBO thresholds
• KV cache dtype and block size
• dtype and quantization settings
• CUDA graph capture sizes or eager-mode toggles when relevant
• prefix cache and speculative decoding settings only when the workload needs those features

vLLM should get a normal sweep, not one baseline command. See references/parameter-coverage.md for the validated flag families. The historical audit happens to use an H100 host, but the flag-family coverage is not H100-specific; confirm each flag on the target image's --help before a run.

Keep gpu_memory_utilization in the baseline for the default pass. Search it only when the question is explicitly about fitting the model or trading capacity against throughput.

Keep DBO and all2all backend settings out of the default pass unless the target vLLM environment is already set up for them. They are real tuning knobs, but a candidate can fail at startup if the required all2all backend is not available. Also preflight concurrent partial prefill before raising max_num_partial_prefills above 1; some model/runtime combinations reject it at startup.

6. Tune TensorRT-LLM

Use trtllm-serve serve as the server entrypoint when the target environment supports it:

trtllm-serve serve <model> \
  --backend pytorch \
  --tp_size <tp> \
  --pp_size <pp> \
  --kv_cache_free_gpu_memory_fraction 0.75 \
  --host 0.0.0.0 \
  --port 8000

Then benchmark the OpenAI-compatible endpoint with the TensorRT-LLM serving benchmark client or with the same OpenAI-compatible client used for the other frameworks.

For TensorRT-LLM 1.0.0, benchmark_serving --dataset-name random samples from ShareGPT unless you pass either --download-path or --random-ids. For a fast synthetic smoke test, pass --random-ids.

TensorRT-LLM flag names are especially version-sensitive. In the validated TensorRT-LLM 1.0.0 image, the KV-cache memory flag accepted by trtllm-serve serve is --kv_cache_free_gpu_memory_fraction, not --free_gpu_memory_fraction. Verify this with trtllm-serve serve --help before running a search on any GPU target.

TensorRT-LLM backend policy for this skill:

• launch the server with --backend pytorch
• keep backend: pytorch in base_server_flags
• do not add backend to search_space
• reject trt, engine-backed serving, or any other non-PyTorch TensorRT-LLM server backend as unsupported for this skill

Version-sensitive TensorRT-LLM knob families to verify:

• tp_size, pp_size, and ep_size
• max batch size, max sequence length, max number of tokens, and KV-cache budget
• inflight batching and scheduler options
• extra LLM API options YAML used by trtllm-serve with the PyTorch backend

The trtllm-serve serve CLI exposes fewer direct runtime knobs than SGLang or vLLM. Use direct flags when they exist, then use --extra_llm_api_options for PyTorch-backend settings that are not top-level CLI flags. Keep unsupported backend or engine requests in the failure table instead of translating them.

Keep kv_cache_free_gpu_memory_fraction in the baseline for the default pass. Search max_batch_size, max_num_tokens, max_seq_len, and validated PyTorch-backend config options first. The server backend remains fixed to pytorch.

7. Normalize Results

Write one JSONL row per candidate using the schema in references/result-schema.md. Then run:

python skills/llm-serving-auto-benchmark/scripts/compare_benchmark_results.py \
  --input /path/to/candidates.jsonl \
  --output /path/to/summary.md

Rank candidates in this order:

1. SLA passed
2. highest request throughput or goodput
3. highest output token throughput
4. lower p99 TTFT
5. lower p99 TPOT/ITL
6. lower GPU count or simpler deployment if performance is close

Output Contract

Return a compact report with:

• workload and SLA used
• hardware and framework versions
• for each framework, one table listing the best deployment command for each dataset scenario and all relevant performance metrics
• one cross-framework comparison table for the selected best command per framework and scenario, including the command, so the deployment choice is clear for each dataset
• failed or excluded candidates with reasons. Explain that this table is an record of tried configs that were not selected: candidates that failed, were skipped by policy, or completed but missed the SLA.
• exact launch command and benchmark command for each winner
• artifact paths: canonical workload, raw results JSONL, normalized JSONL, CSV or markdown summary, and server logs needed to debug winners or failures
• a caveat if the workload was synthetic, if any framework did not complete a fair search, or if any framework needed framework-specific parameter substitutions

Use references/framework-matrix.md when you need command templates or source links for each framework. Use references/example-plan.yaml as the starting point for a full cross-framework run plan. Use references/version-notes.md to understand which source snapshots informed this skill and what has or has not been smoke-tested.