LLM Serving Auto Benchmark

LLM Serving Auto Benchmark

Overview

Use this skill to compare LLM serving frameworks such as SGLang, vLLM, TensorRT-LLM, and TokenSpeed 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 define a framework-neutral LLM serving cookbook model set and translate each entry into framework-native SGLang, vLLM, and TensorRT-LLM, and TokenSpeed 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/*.yaml

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: current public cookbooks use sglang serve; python -m sglang.launch_server remains accepted. Prefer 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
• TokenSpeed: tokenspeed serve for the OpenAI-compatible server plus tokenspeed bench serve or the same OpenAI-compatible benchmark client used for the other frameworks

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.

Framework selection is caller-controlled. If the caller explicitly supplies a framework list, benchmark only those enabled frameworks. Do not silently add TensorRT-LLM or TokenSpeed just because cookbook configs or history docs exist; record omitted frameworks as user-excluded, not unsupported.

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 is framework-neutral: 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 inside this skill's own workflow. The concrete SSH wiring, container names, workspace paths, and HF token plumbing for a given box live in operator-side per-host skills; this skill only requires that the caller can reach a shell inside a container with the requested framework installed.

Reference files are optional and version-sensitive. Treat historical flag notes as evidence from one image, not as a compatibility guarantee for the next run.

Additional H100 validation on 2026-05-01 used two 2-card models with a bounded search of two SGLang memory-fraction candidates and two vLLM memory-utilization candidates. The workload was random input 512, output 64, 8 prompts, and 2 warmup requests, only to prove the search and summary path can finish quickly.

Additional B200 smoke validation on 2026-06-27 used GPUC5A6 (cirrascale-gpuc5a6) in container sglang_bbuf, artifact root /data/bbuf/ai_infra_skills_pr72_20260627. The target image had SGLang 0.5.13.post1 installed, but no vllm, trtllm-serve, or tokenspeed CLI in that container, so only SGLang was model-smoked and the missing frameworks were recorded as environment gaps, not as unsupported frameworks.

The 2026-08-23 source refresh did not recapture those rows. The assigned B200 host (lmsys@40.142.96.44) closed SSH on port 22, so there is no new GPU smoke, MiniMax-M3 revalidation, or four-framework CLI help capture from this date. Treat every 2026-08-23 flag or recipe note as a source check, not as a replacement for the 2026-06-27 B200 evidence.

The same B200 refresh ran the cookbook validator against captured help snapshots. Missing-command help captures such as trtllm-serve_missing.txt are now ignored unless at least one real --flag is present, preventing a missing framework binary from being misreported as hundreds of unsupported framework flags.

Skill Scope

This skill is a playbook plus a config+validator toolchain, not a turn-key orchestrator. The operator still launches servers, drives workloads, and writes one normalized JSONL row per candidate.

The scripts/ directory contains exactly two tools:

• validate_cookbook_configs.py: load cookbook YAML, render bounded candidate server commands, and check flag names against captured --help snapshots without launching servers.
• compare_benchmark_results.py: turn normalized per-candidate JSONL into the markdown and optional CSV tables described in the Output Contract.

Cookbook configs under configs/cookbook-llm/ must pass the validator. The shorter references/example-plan.yaml is a one-off runtime-plan skeleton and is not expected to pass as-is. Use references/result-schema.md as the single source of truth for SLA key names.

Required Inputs

Collect these before a long run:

• model and tokenizer path, target frameworks, GPU model/count, multi-node allowance, precision, and quantization constraints
• endpoint shape, workload source, dataset scenarios, SLA target, search budget, and artifact output directory
• version manifest: framework package version or git commit, container/Python environment, --help snapshots, and whether each search parameter 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.

Record each scenario's input/output length distribution in the normalized result rows. This is now part of the profiler handoff contract: if SGLang is slower and sglang-sota-humanize-loop invokes llm-torch-profiler-analysis, the profiler workload must reuse the slow SGLang benchmark scenario lengths instead of falling back to its generic prefill 4090->1 and decode 1->2048 defaults.

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 still exposes --long-prefill-token-threshold; verify the exact flag against the target image before searching it.
• SGLang current mainline was checked on 2026-08-23 at eec794bce0808ae26cc1dcb84a56b65d2df82af5 (released tags v0.5.17 on 2026-08-08 and v0.5.18 on 2026-08-22). Record these as source notes, not as GPU-smoked winners:
  • public cookbooks now emit sglang serve; keep capturing python -m sglang.launch_server --help until the target image drops it
  • v0.5.17 adds opt-in --enable-session-radix-cache plus /close_session, experimental --dwdp-size MoE prefill, and a weight-cache daemon
  • v0.5.18 moves compiled-kernel caches under SGLANG_CACHE_DIR (first launch after upgrade recompiles once) and adds --startup-weight-load-mode overlap
  • v0.5.18 CUDA images require torch 2.13.0 / triton 3.7.1; treat older torch 2.9/2.11 images as stale before scoring them. --torchao-config is removed. NVFP4 + flashinfer_trtllm MoE deferred finalize is on by default for the DeepSeek-V3 family
  • FlashInfer MNNVL pure-allreduce is auto-on for DeepSeek-V3/V3.2/V4; elsewhere it is --enable-flashinfer-pure-allreduce
  • v0.5.18 known issues: Kimi K3 MLA gate→QKV-A GEMM fusion landed then reverted (#33623 / #34642); AMD GLM-5.2 fused shared-expert append also reverted (#31323 / #35105); v0.5.17 gRPC parallel request lifecycle tracking was reverted (#34160)
• vLLM current mainline was checked on 2026-08-23 at bbe8b23e1a2b32a96240b27f63255170d09ef144 (released tags v0.27.0 on 2026-08-10 and v0.27.1 on 2026-08-11). It includes the earlier PR #46735 CUDA-graph Triton / NVFP4-emulation MoE fix, plus the Kimi K3 stack that landed in 0.27.0. If a target image predates either, treat Triton-MoE graph-capture failures, eager fallback, or missing Kimi K3 loaders as an image/runtime issue before scoring it against SGLang.
• The same vLLM refresh includes PR #44800 (VLLM_GPU_SYNC_CHECK). For sync-heavy profiler rows, record whether the target image exposes this debug knob before labeling the gap as kernel-local.
• vLLM PR #42669 extends FlashAttention-4 SM100 support to head dimension 256, while PR #49982 fixes MLA padding, grouped top-k routing, and routed scaling in the Transformers modeling backend. Treat images predating either change as stale when those exact paths affect a row; neither merge is itself benchmark evidence.
• The same vLLM head lists Qwen/Qwen3.8-27B in the model registry, but the public HF config.json uses model_type=qwen3_5. Do not invent a separate vLLM Qwen3.8 implementation tree; compare that checkpoint against the existing Qwen3.5 loader. The four-framework cookbook still leaves vLLM disabled until a target-image smoke exists.
• TensorRT-LLM mainline was checked on 2026-08-23 at da38c1d2e0dffd073b7dfb6d69e15ee7b45d84a9. Keep kv_cache_free_gpu_memory_fraction in shipped configs until the target trtllm-serve serve --help proves a shorter alias is accepted. Same-day mainline also includes Qwen3.5/3.8 wave-2 (#17700), Qwen3.8-27B FP8 VLM quant-config cleanup (#17786), Kimi K3 MLA decode backend selection (#17800), and Kimi K3 NVFP4 MegaMoE SiTU (#17865). Record those as source notes only; they do not enable a four-framework cookbook lane without target-image smoke.
• TensorRT-LLM 1.2+ and current 1.3.0rc line have removed the TensorRT engine backend. PyTorch is the sole server backend, which matches this skill's existing trtllm-serve serve --backend pytorch pin. Images that still advertise --backend tensorrt are stale; do not search that backend.
• TensorRT-LLM current mainline includes PR #11685 and PR #15546, which affect KV block eviction and KV block-offset host staging. If a target image predates them, record stale-runtime risk when cache pressure, block-offset races, or prefix/KV residency affect benchmark rows.
• TensorRT-LLM PR #16805 fixes disaggregated draft-token adoption and sequence-length accounting; PR #16763 unifies phase-1 CUDA graph cleanup before final KV-cache allocation. Record the image SHA when disaggregated speculative output or startup memory differs across runs.
• The historical TensorRT-LLM 1.0.0 multi-GPU PyTorch-backend validation used --ipc=host, --ulimit memlock=-1, --ulimit stack=67108864, --shm-size=16g, and NCCL_IB_DISABLE=1 (for single-node) or an equivalent NCCL setup. Keep these as a starting point, not as a version-independent requirement.
• TensorRT-LLM current mainline still exposes benchmark client backends openai and openai-chat, not trtllm. 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).
• TokenSpeed is a fast-moving engine. Current mainline checked on 2026-08-23 at lightseekorg/tokenspeed@2706143a8669d50a8f56466b9d340b86922b8f2d exposes tokenspeed serve, tokenspeed bench, tokenspeed env, and tokenspeed version. Its server command is tokenspeed serve <model>, not a python -m tokenspeed entrypoint.
• TokenSpeed PR #821 documents Kimi K3 FlatKV, KDA/MLA, and B300/AMD deployment contracts. Keep it as source guidance only: the recipe contains platform-specific sidecars, checkpoint-layout requirements, and explicit output-quality and validation caveats, so it does not enable a generic benchmark lane without target-image and model smoke evidence. After 2026-07-28, TokenSpeed also merged scheduler 0.1.9 (#1208), Kimi K3 unrouted-decode projection routing (#1200), and an MLA FP8 KV packing fallback (#1199). Record the target SHA; do not treat those merges as a benchmark winner.
• TokenSpeed's SGLang/vLLM-compatible parameter names are not always identical in meaning. Prefer --max-model-len, --max-num-seqs, --chunked-prefill-size, --max-prefill-tokens, --max-total-tokens, --tensor-parallel-size, --attn-tp-size, --moe-tp-size, --enable-expert-parallel, --attention-backend, --moe-backend, --kv-cache-dtype, and speculative flags only after confirming the target tokenspeed serve --help output.
• TokenSpeed has an agentic benchmark path in-tree. When the workload is multi-turn or tool-heavy, add a TokenSpeed-native tokenspeed bench serve or EvalScope-style run beside the common OpenAI-compatible client and record both result files in the same normalized row set.
• 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 SGLang plus all requested comparison frameworks before starting a search. Run only the commands for the requested framework set:

sglang serve --help
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
tokenspeed serve --help
tokenspeed bench --help
tokenspeed bench serve --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 TokenSpeed, confirm both the server and benchmark entrypoints because some installations alias the binary as ts. Record the exact binary used in server_command.

For each framework, launch a minimal server, confirm /v1/models or the native model-info endpoint, send one streaming request, run one tiny benchmark with at least 5 requests, then 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 and save image tags, pull commands, launch/benchmark logs, and cleanup commands.

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, use the shipped two-scenario shape:

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. Name each scenario and keep the aligned pair in the artifacts. For custom datasets, compute or record representative input_len and output_len buckets, at least p50 and p95 when possible, so later profiler runs can match the slow bucket rather than profiling an unrelated synthetic shape.

Before searching any sequence-length limit, compute the largest input_len + output_len in the dataset. SGLang context_length, vLLM max_model_len, TensorRT-LLM max_seq_len, and TokenSpeed max_model_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
• TokenSpeed gpu_memory_utilization

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 shared cookbook config 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, long_prefill_token_threshold, 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/framework-reference.md for native command templates and cross-framework knob families. 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. Verify long_prefill_token_threshold on the concrete vllm serve --help=all surface before adding it to a search.

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.

In the historical TensorRT-LLM 1.0.0 validation image, benchmark_serving --dataset-name random sampled from ShareGPT unless either --download-path or --random-ids was passed. For a fast synthetic smoke test, pass --random-ids, then confirm the behavior on the target TensorRT-LLM image.

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 was --kv_cache_free_gpu_memory_fraction, not --free_gpu_memory_fraction. Current mainline was rechecked at da38c1d2e0dffd073b7dfb6d69e15ee7b45d84a9 on 2026-08-23. Always verify flags 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. Tune TokenSpeed

Use TokenSpeed as a first-class comparison framework, especially for agentic or multi-turn workloads where it may be the strongest non-SGLang baseline:

tokenspeed serve <model> \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size <tp> \
  --gpu-memory-utilization 0.90 \
  --max-model-len 12288 \
  --max-num-seqs 64 \
  --chunked-prefill-size 8192 \
  --kv-cache-dtype auto \
  --trust-remote-code

Benchmark either with TokenSpeed's native online serving benchmark:

tokenspeed bench serve \
  --base-url http://127.0.0.1:8000 \
  --model <model> \
  --dataset-name random \
  --random-input-len 1024 \
  --random-output-len 256 \
  --num-prompts 80

or with the same OpenAI-compatible client used for the other frameworks. When TokenSpeed is a likely leader and profiler handoff will be needed, the native benchmark can also arm torch profiling for the same request shape:

tokenspeed bench serve \
  --base-url http://127.0.0.1:8000 \
  --model <model> \
  --dataset-name random \
  --random-input-len 1024 \
  --random-output-len 256 \
  --num-prompts 80 \
  --profile \
  --profile-num-steps 5 \
  --extra-body '{"output_dir":"/data/bbuf/profiles/tokenspeed","activities":["CPU","GPU"],"with_stack":true,"profile_id":"ts-bench"}'

If output_dir is not supplied, TokenSpeed writes under TOKENSPEED_PROFILER_DIR, defaulting to /tmp.

For large MoE or agentic checkpoints, validate these TokenSpeed knobs before adding them to the search:

• tensor_parallel_size, attn_tp_size, moe_tp_size, and enable_expert_parallel
• attention_backend, drafter_attention_backend, moe_backend, and draft_moe_backend
• max_num_seqs, chunked_prefill_size, max_prefill_tokens, and max_total_tokens
• kv_cache_dtype, quantization, and prefix-cache controls
• speculative flags such as speculative_algorithm, speculative_num_steps, and speculative_num_draft_tokens
• comm_fusion_max_num_tokens, enable_allreduce_fusion, and related communication-fusion flags only after a target-image smoke run

If a TokenSpeed-native agentic config exists for the same model family, run it as an additional workload lane rather than replacing the common cross-framework scenario. Normalize its result rows with the same schema and mark workload.kind accordingly.

Keep gpu_memory_utilization pinned in the default pass. Search it only when the user explicitly wants a capacity study.

8. 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 p50 TTFT
5. lower p50 TPOT/ITL
6. lower GPU count or simpler deployment if performance is close

Keep the SLA gate itself unchanged. In the cookbook configs and normalized result schema, TTFT SLA still uses max_p50_ttft_ms and TPOT SLA still uses max_p50_tpot_ms; only the default cross-candidate comparison order switches to p50 TTFT and p50 TPOT.

Output Contract

Return a compact report with workload/SLA, hardware and framework versions, best deployment-command tables per framework/scenario, one cross-framework comparison table, exact launch and benchmark commands for winners, and artifact paths for workload, raw/normalized results, CSV or markdown summary, and server logs.

When SGLang is not the winner, include a profiler handoff note with the slow SGLang scenario name and the exact input/output lengths or percentile bucket to pass to llm-torch-profiler-analysis.

When a candidate uses speculative decoding, prefix cache, offload, or an agentic workload, record the optional normalized fields for accept length, pre-scheduler time, cache hit rate, and memory residency. The summary script will display those columns when present.

Include failed or excluded candidates with reasons. Explain that this table is a record of tried configs that were not selected: candidates that failed, were skipped by policy, or completed but missed the SLA. Add caveats for synthetic workloads, incomplete fair searches, or framework-specific parameter substitutions.

Use references/framework-reference.md when you need command templates, source links, or knob-family mappings. Use references/example-plan.yaml as the starting point for a full cross-framework run plan.