Sglang Sota Humanize Loop

SGLang SOTA Humanize Loop

Overview

Use this skill when the user names a model and wants the SGLang serving path to autonomously keep improving until it matches or beats the best reproducible result from the requested comparison framework set in the same target environment.

This workflow has two durable parts:

1. A fixed baseline phase that must be completed once before any code patching.
2. One Humanize RLCR loop that owns gap decision, profiling, required layer/kernel deep dive, SGLang patching, optional NCU evidence, and real-model revalidation.

Do not split the campaign into a pre-loop profiling phase plus a later patch loop. After the fixed benchmark exists, Phase 2 gap decisions, Phase 3 profiling, llm-pipeline-analysis, kernel evidence, and code changes all belong inside the same model-level RLCR loop.

Runtime Roots

This skill can run from Claude Code, Codex, or another compatible skill runtime. Resolve companion roots in this order:

1. Prefer installed Claude Code skills under ~/.claude/skills when running in Claude Code.
2. Prefer installed Codex skills under ${CODEX_HOME:-~/.codex}/skills when running in Codex.
3. Fall back to checked-out repositories when the skills are symlinked or kept local for development.

Example local paths:

Humanize runtime: ${CODEX_HOME:-~/.codex}/skills/humanize
ncu-report-skill: ${CODEX_HOME:-~/.codex}/skills/ncu-report-skill/SKILL.md
Model PR history knowledge: <repo>/model-pr-optimization-history

For Claude Code installs, the equivalent defaults are typically:

Humanize runtime: ~/.claude/skills/humanize
ncu-report-skill: ~/.claude/skills/ncu-report-skill/SKILL.md
Model PR history knowledge: ~/.claude/skills/model-pr-history-knowledge

If the Humanize runtime is missing, locate a plugin or skill directory containing scripts/setup-rlcr-loop.sh. If ncu-report-skill is unavailable, kernel edits may still proceed from torch-profiler/source evidence, but record the missing NCU evidence path as a blocker when a kernel change would normally need Nsight Compute diagnostics.

Companion Skills

Read these before a real run:

• ../llm-serving-auto-benchmark/SKILL.md
• ../llm-torch-profiler-analysis/SKILL.md
• ../llm-pipeline-analysis/SKILL.md
• ../../model-pr-optimization-history/SKILL.md
• the matching host or operator skill for SSH, container, GPU, and artifact conventions

Read ncu-report-skill/SKILL.md only when the active RLCR round is writing or evaluating a CUDA, Triton, CuTe, CUTLASS, TileLang, or torch.compile kernel path and Nsight Compute evidence is needed.

Contract

Given a model-level SGLang SOTA request, do not ask the user to run separate benchmark, profiler, gen-plan, refine-plan, or Humanize setup commands. Do the setup yourself.

Ask the user only if the model, target GPU environment, or precision/quantization policy is missing and cannot be inferred from local configs or the active host skill.

Keep only the fixed benchmark phase outside the RLCR patch loop. Once the fixed cross-framework benchmark and model PR history notes exist, start Humanize. The RLCR loop itself must decide whether a gap still exists, collect current profiler evidence, run layer pipeline analysis, patch SGLang, and revalidate.

Treat the model optimization campaign as the durable unit, not one terminal session. The campaign is recoverable from the run artifact root, checkpoint files, benchmark/profile artifacts, NCU digests, and ledgers.

Phase 0: Inputs And Run Directory

Collect or infer:

• model id or checkpoint path, tokenizer, precision, quantization, trust policy, and max context length
• target SGLang checkout to patch
• GPU type/count, visible GPU ids, container or remote shell, CUDA/NCCL versions, and whether multi-node is allowed
• target framework: always SGLang
• comparison framework set: if the user explicitly names or excludes comparison frameworks, honor that exact set; otherwise default to vLLM, TensorRT-LLM, and TokenSpeed when available
• model-family history slug inferred from the model id, checkpoint, or hot framework source path when possible
• artifact root

Create one run directory:

runs/YYYYMMDD_<model_slug>_sota_humanize/
  manifest.md
  help/
  benchmark/
    fairness/
  profiles/
  analysis/
    root-cause.md
    layer-pipeline.md
    fairness-diagnostics.md
  history/
    model-pr-history-notes.md
  kernel/
    ncu-digests/
  patches/
  humanize/
    model-loop-checkpoint.md
  final_report.md

Never save Hugging Face tokens or other secrets in artifacts.

Framework Selection

Treat SGLang and the comparison frameworks as separate concepts:

• SGLang is always the target framework because this loop patches SGLang.
• comparison_frameworks contains only competitors, using canonical names vllm, tensorrt_llm, and tokenspeed.
• If the user says "only compare with vLLM", "comparison_frameworks: [vllm]", or "do not consider TensorRT-LLM/TokenSpeed", run only SGLang plus vLLM.
• Do not silently add TensorRT-LLM or TokenSpeed because a config or doc exists.
• Record selected comparison frameworks and user-excluded frameworks in manifest.md and benchmark/fairness/diagnostics.md.
• Only read competitor history, launch competitor benchmarks, and profile competitor traces for the selected comparison framework set.

Phase 0.5: Model PR History Knowledge Gate

Before the fixed benchmark and before any patch planning, query and read model-pr-optimization-history for the target model family.

Rules:

• If the slug is unclear, run scripts/query.py "<model id or family>" from the knowledge root and choose the closest model-family history.
• Read the SGLang history for that family whenever it exists.
• Read competitor history too when a selected comparison framework later becomes the leading competitor or its source/trace suggests a missing SGLang fast path.
• Write history/model-pr-history-notes.md with the paths read, PR numbers, immutable source heads, PR state (open, merged, or closed-unmerged), source files, symbols, validation evidence, known limitations, validation risks, and the concrete decision each item influences.
• Treat open PRs as candidate baselines only. Record their full head SHAs and never describe them as behavior already present in the recorded source head.
• Treat these notes as source and PR memory that helps choose a better SGLang patch, not measured proof by itself.

If the knowledge root is unavailable, record the blocker in the same notes file and continue with benchmark/profile evidence.

Phase 1: Fixed Fair Benchmark Gate

This phase is mandatory and happens exactly once before Humanize starts.

Use llm-serving-auto-benchmark as the source of truth for candidate generation, result schema, workload, and comparison.

Hard requirements:

• Search SGLang and every selected comparison framework's best deployment command when that framework is supported in the target environment.
• Do not compare tuned SGLang against competitor defaults. Every framework gets its own bounded search.
• Use the same model weights, tokenizer, precision, quantization, GPU type/count, GPU ids, endpoint path, sampling settings, and SLA.
• Record package version or git commit plus server/benchmark --help snapshots for SGLang and every selected comparison framework. The 2026-08-23 immutable source heads checked for this skill are SGLang eec794bce0808ae26cc1dcb84a56b65d2df82af5, vLLM bbe8b23e1a2b32a96240b27f63255170d09ef144, TensorRT-LLM da38c1d2e0dffd073b7dfb6d69e15ee7b45d84a9, and lightseekorg/tokenspeed@2706143a8669d50a8f56466b9d340b86922b8f2d; still prefer target-image --help over these source notes and re-check open PRs before every long SOTA run.
• Use the target image's current commands: sglang serve or python -m sglang.launch_server, vllm serve, trtllm-serve serve --backend pytorch, and, when selected, tokenspeed serve <model> (or its verified image alias).
• Use the default two dataset scenarios from llm-serving-auto-benchmark unless the user explicitly provides a production workload:
  • dataset kind random, num_prompts: 80
  • chat: random input 1000, output 1000
  • summarization: random input 8000, output 1000
  • treat the two input/output pairs as aligned scenarios, not a cartesian product
• Do not replace those scenarios with an easier smoke dataset for the real SOTA decision. Smoke runs are allowed only when labeled as flow checks.
• For selected TensorRT-LLM comparisons, keep trtllm-serve serve --backend pytorch; reject non-PyTorch TensorRT-LLM server backends for this skill.
• For selected TokenSpeed comparisons, use tokenspeed serve <model> or the target image's exact alias such as ts serve; record the exact binary spelling and validate tokenspeed serve --help plus tokenspeed bench serve --help.
• If TokenSpeed is selected and a TokenSpeed-native agentic workload exists for the model family, run it as an additional lane beside the common cross-framework workload. Do not let it replace the common scenario unless the user explicitly chose that workload as the SOTA target before benchmarking.
• Keep failed, skipped, and SLA-failing candidates in the benchmark artifact.
• Record optional fairness fields in normalized rows whenever available: speculative accept length, pre-scheduler/scheduler time, cache hit rate, offload or memory residency, endpoint, and request-shape notes.

Before finishing or reporting a blocker, stop only the server, benchmark, profiler, downloader, and log-tail processes started by this run. Remove only this run's explicit model snapshot, lock, mirror directory, and run-specific cache entries. Record before/after process and GPU state; never kill another user's process or delete shared model caches.

Write:

• benchmark/candidates.jsonl
• benchmark/summary.md
• benchmark/winning-commands.md
• benchmark/fairness/diagnostics.md
• framework help outputs under help/
• the exact launch and benchmark commands for every winner

Do not choose a code patch outside RLCR. The fixed winner table is the baseline input to the model loop.

Phase 2: Build The Humanize Plan

Create a Humanize plan inside the SGLang checkout that will be patched:

.humanize/sglang-sota-agent/refined-plan.md

Use references/refined-plan-template.md as the skeleton and fill it with the actual model, workload, benchmark winners, artifact root, model PR history notes, and target SGLang checkout.

The plan must require every RLCR round to:

• preserve the fixed benchmark workload and SLA
• preserve and consult history/model-pr-history-notes.md before choosing model-specific SGLang source paths
• preserve and consult benchmark/fairness/diagnostics.md before treating a gap as a code bottleneck
• run the gap decision inside the loop before patching
• run llm-torch-profiler-analysis inside the loop when SGLang is behind or when the previous patch changed the profiled path
• run llm-pipeline-analysis inside the loop after profiler triage and before choosing a source path, representative layer, or kernel target
• patch SGLang code, not just benchmark parameters
• use ncu-report-skill inside the same loop when a kernel edit needs Nsight Compute evidence
• re-run real model benchmark/profile after each accepted patch
• continue through multiple minimal patches when one patch only closes part of the gap
• record every attempt, failed idea, partial win, rejected source idea, and final selected patch in artifacts

Phase 3: Start RLCR

Before starting Humanize from the SGLang checkout:

• Ensure the SGLang checkout is a git repository with at least one commit and a clean working tree, excluding only gitignored Humanize runtime state.
• Ensure .humanize* is gitignored so RLCR state, round summaries, and local checkpoints cannot be staged accidentally.
• Ensure the intended review base branch is present locally. Pass --base-branch <branch> if Humanize's auto-detection would be ambiguous.
• Do not start a new loop if any existing .humanize/rlcr/*/state.md is active in the SGLang checkout. Resume, finish, or cancel the old model loop first.

From the SGLang checkout, run:

"$HUMANIZE_RUNTIME_ROOT/scripts/setup-rlcr-loop.sh" \
  .humanize/sglang-sota-agent/refined-plan.md --yolo

If HUMANIZE_RUNTIME_ROOT is not already set by the client/plugin environment, resolve it to the installed Humanize runtime first. In Codex, this is often ${CODEX_HOME:-~/.codex}/skills/humanize; in Claude Code it is often ~/.claude/skills/humanize or a plugin-provided Humanize runtime. If setup exits non-zero, stop and report the error. Do not bypass the gate.

After setup succeeds:

1. Find the active state file with find .humanize/rlcr -maxdepth 2 -name state.md -print.
2. Verify the state file exists, contains current_round: 0, and contains ask_codex_question: false when using --yolo.
3. Verify .humanize/rlcr/<timestamp>/round-0-prompt.md exists and includes the generated Round Contract Setup instructions.
4. Read .humanize/rlcr/<timestamp>/round-0-prompt.md.
5. Execute the current round.
6. Commit SGLang changes.
7. Write the required Humanize round summary.
8. Stop normally so the native Humanize Stop hook can review.

If no active state file exists, or if round-0-prompt.md is missing, stop and report that RLCR did not start correctly. Do not continue into SGLang patch work outside the Humanize loop. If the hook blocks exit, follow the generated next-round prompt exactly.

Inside Each RLCR Round

Gap Decision

At the start of every round, compute current SGLang's gap against the best SLA-passing selected comparison framework for each fixed scenario.

Use 1% as the default stable noise threshold. If the current result is within +/-1%, rerun the winning commands enough times to decide whether the gap is stable before choosing a patch.

Patch only when SGLang is slower than the best selected comparison framework by more than 1%, fails SLA while a selected competitor passes, or has a profiled bottleneck that explains the remaining gap under the fixed workload.

Before patching, update analysis/fairness-diagnostics.md from the current benchmark rows and the fixed benchmark/fairness/diagnostics.md. Check:

• same GPU count, visible GPU ids, model weights, tokenizer, precision, quantization, endpoint, sampling settings, and request shape
• speculative decoding status, algorithm, number of steps, and mean accept length for every framework that enables a drafter/MTP/EAGLE-style path
• pre-scheduler, scheduler, queue, tokenizer, and detokenizer time when exposed by server logs or benchmark output
• prefix/KV cache status and hit rate, especially for repeated prompts or multi-turn workloads
• GPU memory utilization, KV dtype, offload, and model weight residency
• TokenSpeed-specific serving knobs, when TokenSpeed is selected, such as attn_tp_size, moe_tp_size, enable_expert_parallel, backend choices, and communication-fusion flags

If the current leader wins because it used a non-equivalent workload, non-equivalent cache/speculative setting, different resident weights, or a framework-specific endpoint shortcut, fix the benchmark normalization first and rerun. Do not patch SGLang code until the gap survives this fairness gate.

If SGLang is already best or tied within the stable threshold, write the final report and stop under the normal Humanize review path.

Required Profiling

When SGLang is behind, profile the current best SGLang command and the leading competitor command with llm-torch-profiler-analysis.

Rules:

• Always profile SGLang when it is behind.
• Always profile at least the current best selected comparison framework.
• If multiple competitors are more than 1% ahead of SGLang in a stable result, profile each ahead competitor that can produce a trace. For TokenSpeed, analyze an existing torch-profiler trace or run-directory trace emitted by the deployment; do not fabricate a live profiler endpoint when the target image only supports offline traces.
• Use the slow benchmark scenario lengths, not the profiler defaults:
  • prefill profile: slow input length -> 1 output token
  • decode profile: 1 input token -> slow output length
• For mixed or production datasets, use the slowest representative p50 or p95 bucket already recorded by the benchmark artifact.
• Capture or analyze separate prefill and decode evidence when the framework supports it.

For every profiled framework, save the same three tables:

• kernel table
• overlap-opportunity table
• fuse-pattern table

Then write or update analysis/root-cause.md with the current cross-framework comparison: which stage is slower, which table rows explain it, and which SGLang source paths or kernel families are plausible patch targets.

Do not patch SGLang until this report exists for the current gap.

Layer Pipeline Deep Dive

Run llm-pipeline-analysis inside every RLCR round after profiler triage and before choosing a patch target.

The report must identify:

• the chosen forward pass and why it is representative
• the relevant layer types, especially for heterogeneous layers such as MoE, hash layers, or compress_ratios
• representative layers for the patch target
• top hot kernels in those representative layers
• any Perfetto time ranges needed for inspection

Use the profiled SGLang trace and the served model config. Write analysis/layer-pipeline.md with the chosen forward pass, layer-type timing table, representative layers, top hot kernels, and any Perfetto ranges used for inspection. Do not choose a SGLang patch before this report exists for the current round.

Kernel Evidence Assist

Use ncu-report-skill only when the active RLCR round is writing a concrete kernel or small kernel-family patch and torch-profiler evidence is not enough to choose or validate the next edit.

Eligibility gate:

• SGLang is still more than 1% behind the best selected comparison framework for the fixed benchmark scenario after the required repeat/profiler checks.
• The slow stage has a concrete SGLang kernel or tightly scoped kernel family in the kernel table with at least 1% cumulative GPU-time share. Do not spend kernel-specialist effort on a lone kernel below 1% share unless a shared implementation affects an aggregated family above 1%.
• llm-pipeline-analysis has identified the representative layer/forward pass and top hot kernels for the current round.
• The proposed kernel target has a clear correctness reference, representative shapes/dtypes/layouts from the model run, and a path to wire the candidate into the active SGLang serving code.

For each eligible kernel target:

1. Read ncu-report-skill/SKILL.md and follow its Nsight Compute workflow for harness construction, ncu collection, report parsing, stall diagnosis, and evidence-backed next-edit selection.
2. Store NCU outputs under kernel/ncu-digests/<version>/ or the host's equivalent artifact root. Each digest must compare baseline vs candidate and end with exactly one concrete next edit.
3. Patch the SGLang kernel or call path directly in the SGLang checkout, with focused correctness and microbench coverage when available.
4. Wire the candidate into the active model-serving path that produced the original profiler row.
5. Re-run the same real-model benchmark and profiler after the candidate is correct. A microbench or NCU win alone is not success.

If no focused harness exists, build the smallest harness that preserves the model-derived shapes/dtypes/layouts. If NCU cannot run on the host, record the blocker in the digest path and keep the next edit grounded in the available torch-profiler, layer-pipeline, and source evidence.

Do not start any standalone .humanize/rlcr session for a kernel target. Kernel work stays inside the active model RLCR loop.

Model-Loop Checkpoint

After every accepted round, update humanize/model-loop-checkpoint.md with:

• original model, tokenizer, precision, quantization, hardware, workload, SLA, artifact root, and benchmark winner commands
• current SGLang branch, commit, patches applied, tests run, and current best SGLang benchmark row
• remaining gap, profiler rows, model PR history notes, layer-pipeline notes, NCU digest paths, rejected source ideas, and the next planned SGLang patch

This checkpoint is for campaign recovery inside the same model-level workflow. It records enough context to resume the campaign without losing benchmark/profile lineage.

Loop Ledgers

Keep these files under the run artifact root or the SGLang checkout, depending on the host convention:

humanize/attempt-ledger.md
humanize/optimization-ledger.md
humanize/source-idea-ledger.md
humanize/lineage.jsonl
humanize/profile-digests/

Every patch attempt gets an attempt row. Only correct patches with measured improvement get optimization rows. Source ideas must include profiler rows, layer-pipeline evidence, NCU report paths when used, and code provenance so later rounds can avoid re-reading the same source. Model PR history evidence should be recorded beside SGLang, selected comparison frameworks, and NCU source ideas when it influenced the patch.

After two consecutive rounds with less than 1% geomean improvement over the prior best SGLang result, expand code-first research before editing again. Prefer code and PR evidence from SGLang, selected comparison frameworks, and relevant kernel source guides before prose-only articles.

Stop Conditions

Stop only when one of these is true:

• SGLang beats the best SLA-passing selected comparison framework result on the fixed workload.
• SGLang is tied within the stable 1% threshold after repeat runs.
• The remaining gap is proven external to SGLang, such as unavailable hardware support, missing framework dependency, unsupported TensorRT-LLM PyTorch backend, unsupported TokenSpeed build, or model weights that cannot be loaded fairly.
• Profile evidence shows the remaining hot path is already near the relevant hardware or algorithmic limit and no low-risk SGLang patch remains.

The final report must include the fixed benchmark table, post-patch benchmark table, all winner commands, model PR history paths, profile paths, layer-pipeline paths when used, NCU digest paths when used, SGLang changed files, tests, and whether SGLang reached target-environment SOTA.