[SKILL] Upgrade sglang profile and auto_benchmark skills (#24250)

.claude/skills/llm-serving-auto-benchmark/SKILL.md

@@ -0,0 +1,527 @@
+---
+name: llm-serving-auto-benchmark
+description: Framework-independent LLM serving benchmark skill for comparing SGLang, vLLM, TensorRT-LLM, or another serving framework. Use when a user wants to find the best deployment command for one model across multiple serving frameworks under the same workload, GPU budget, and latency SLA.
+---
+
+# LLM Serving Auto Benchmark
+
+## Overview
+
+Use this skill to compare LLM serving frameworks such as 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 define a framework-neutral LLM serving cookbook
+model set and translate each entry into framework-native SGLang, vLLM, and
+TensorRT-LLM server flags. Validate those configs before a real run:
+
+```bash
+python .claude/skills/llm-serving-auto-benchmark/scripts/validate_cookbook_configs.py \
+  .claude/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 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 (`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.
+
+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.
+
+| Model | GPUs | Best SGLang | Best vLLM | Artifact root |
+| --- | --- | --- | --- | --- |
+| `Qwen/Qwen3-8B` | 2x H100, TP=2 | `sglang_mem086`, 21.64 req/s, 1385.05 output tok/s, mean TTFT 70.54 ms | `vllm_mem080`, 22.88 req/s, 1464.25 output tok/s, mean TTFT 60.56 ms | `/data/bbuf/validate/core_skill_validation_20260501/qwen3_8b/auto_benchmark` |
+| `mistralai/Mistral-7B-Instruct-v0.3` | 2x H100, TP=2 | `sglang_mem080`, 24.09 req/s, 1541.92 output tok/s, mean TTFT 61.47 ms | `vllm_mem090`, 24.76 req/s, 1584.54 output tok/s, mean TTFT 58.63 ms | `/data/bbuf/validate/core_skill_validation_20260501/mistral_7b_instruct_v03/auto_benchmark` |
+
+## 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](references/example-plan.yaml) is a
+one-off runtime-plan skeleton and is not expected to pass as-is. Use
+[references/result-schema.md](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-performance` 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 `--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.
+- The historical TensorRT-LLM 1.0.0 validation image accepted
+  `--kv_cache_free_gpu_memory_fraction`; the older `--free_gpu_memory_fraction`
+  exited with a CLI error. TensorRT-LLM was refreshed to 1.2.1 stable and
+  1.3.0 release candidates by 2026-04-28, so re-check the accepted flag name
+  via `--help` on the target image before a real run.
+- 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.
+- The historical TensorRT-LLM 1.0.0 benchmark client took `--backend openai` or
+  `--backend openai-chat`; `--backend trtllm` was 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:
+
+```bash
+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, 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](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:
+
+```json
+{"prompt": [{"role": "user", "content": "Summarize this text."}], "output_len": 256}
+{"prompt": "Write a short explanation of CUDA graphs.", "output_len": 128}
+```
+
+Optional fields:
+
+```json
+{
+  "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:
+
+```yaml
+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`, 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:
+
+```bash
+python -m sglang.auto_benchmark run --config /path/to/sglang.yaml
+```
+
+Otherwise launch the server manually and benchmark with:
+
+```bash
+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:
+
+```bash
+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/framework-reference.md](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.
+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:
+
+```bash
+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`. TensorRT-LLM 1.2.1 is the latest stable GitHub
+release as of 2026-04-28, with 1.3.0 release candidates also published; verify
+the current flag 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](references/result-schema.md). Then run:
+
+```bash
+python .claude/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 mean TTFT
+5. lower mean 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_p99_ttft_ms` and TPOT SLA still uses
+`max_p99_tpot_ms`; only the default cross-candidate comparison order switches
+to mean TTFT and mean 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`.
+
+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](references/framework-reference.md) when
+you need command templates, source links, or knob-family mappings. Use
+[references/example-plan.yaml](references/example-plan.yaml) as the starting
+point for a full cross-framework run plan.

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/README.md

@@ -0,0 +1,17 @@
+# Cookbook LLM Configs
+
+These configs define a framework-neutral LLM serving cookbook model set and translate each model into a three-framework run plan for SGLang, vLLM, and TensorRT-LLM.
+
+Scope:
+- SGLang can preserve source-recipe `base_flags` and `search_space` where applicable; if a sequence limit is smaller than the default synthetic scenario, the config raises that limit so the shipped workload can run.
+- vLLM uses framework-native `vllm serve` flags. The translation keeps the same model, tokenizer, dataset shape, GPU count, and high-impact batching/prefix-cache knobs; it does not copy SGLang-only parser or scheduler flags.
+- TensorRT-LLM uses `trtllm-serve serve` with `backend: pytorch` fixed in `base_server_flags`. Backend choice is never searched.
+- The two default random scenarios remain aligned pairs: `chat` uses `1000 -> 1000`, and `summarization` uses `8000 -> 1000`.
+
+Before a real run, capture the target framework `--help` output and validate the configs:
+
+```bash
+python .claude/skills/llm-serving-auto-benchmark/scripts/validate_cookbook_configs.py   .claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm
+```
+
+With captured help files, add `--help-dir <artifact-help-dir>` to check the concrete flag names against that environment. This check only loads configs and renders candidate commands; it does not launch model servers.

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/deepseek-math-v2.yaml

@@ -0,0 +1,130 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: deepseek-math-v2.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: deepseek-ai/DeepSeek-Math-V2
+  tokenizer: deepseek-ai/DeepSeek-Math-V2
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: deepseek-ai/DeepSeek-Math-V2
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/deepseek-math-v2
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      model_path: deepseek-ai/DeepSeek-Math-V2
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - flashinfer
+      decode_attention_backend:
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+      - 8
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4
+      - 8

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/deepseek-r1-0528.yaml

@@ -0,0 +1,133 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: deepseek-r1-0528.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: deepseek-ai/DeepSeek-R1-0528
+  tokenizer: deepseek-ai/DeepSeek-R1-0528
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: deepseek-ai/DeepSeek-R1-0528
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/deepseek-r1-0528
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      enable_symm_mem: true
+      model_path: deepseek-ai/DeepSeek-R1-0528
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+      - 8
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4
+      - 8

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/deepseek-v3.1.yaml

@@ -0,0 +1,132 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: deepseek-v3.1.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: deepseek-ai/DeepSeek-V3.1
+  tokenizer: deepseek-ai/DeepSeek-V3.1
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: deepseek-ai/DeepSeek-V3.1
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/deepseek-v3.1
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      model_path: deepseek-ai/DeepSeek-V3.1
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+      - 8
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4
+      - 8

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/deepseek-v3.2.yaml

@@ -0,0 +1,132 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: deepseek-v3.2.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: deepseek-ai/DeepSeek-V3.2
+  tokenizer: deepseek-ai/DeepSeek-V3.2
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: deepseek-ai/DeepSeek-V3.2
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/deepseek-v3.2
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      model_path: deepseek-ai/DeepSeek-V3.2
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+      - 8
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4
+      - 8

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/deepseek-v3.yaml

@@ -0,0 +1,133 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: deepseek-v3.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: deepseek-ai/DeepSeek-V3
+  tokenizer: deepseek-ai/DeepSeek-V3
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: deepseek-ai/DeepSeek-V3
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/deepseek-v3
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      enable_symm_mem: true
+      model_path: deepseek-ai/DeepSeek-V3
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+      - 8
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4
+      - 8

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/devstral-small-2-24b-instruct-2512.yaml

@@ -0,0 +1,123 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: devstral-small-2-24b-instruct-2512.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: mistralai/Devstral-Small-2-24B-Instruct-2512
+  tokenizer: mistralai/Devstral-Small-2-24B-Instruct-2512
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 1
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: mistralai/Devstral-Small-2-24B-Instruct-2512
+  max_concurrency:
+  - null
+  - 16
+  - 32
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 16.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/devstral-small-2-24b-instruct-2512
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      model_path: mistralai/Devstral-Small-2-24B-Instruct-2512
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 1
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 1
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/ernie-4.5-21b-a3b-pt.yaml

@@ -0,0 +1,117 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: ernie-4.5-21b-a3b-pt.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: baidu/ERNIE-4.5-21B-A3B-PT
+  tokenizer: baidu/ERNIE-4.5-21B-A3B-PT
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 1
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: baidu/ERNIE-4.5-21B-A3B-PT
+  max_concurrency:
+  - null
+  - 16
+  - 32
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 16.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/ernie-4.5-21b-a3b-pt
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      model_path: baidu/ERNIE-4.5-21B-A3B-PT
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 1
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 1
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/glm-4.5.yaml

@@ -0,0 +1,122 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: glm-4.5.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: zai-org/GLM-4.5
+  tokenizer: zai-org/GLM-4.5
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 4
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: zai-org/GLM-4.5
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 8.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/glm-4.5
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 4
+      context_length: 9000
+      model_path: zai-org/GLM-4.5
+      trust_remote_code: true
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 4
+      trust_remote_code: true
+      gpu_memory_utilization: 0.9
+      max_model_len: 9000
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 4
+      pp_size: 1
+      trust_remote_code: true
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 9000
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 9000
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/glm-4.6.yaml

@@ -0,0 +1,135 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: glm-4.6.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: zai-org/GLM-4.6
+  tokenizer: zai-org/GLM-4.6
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: zai-org/GLM-4.6
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/glm-4.6
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      model_path: zai-org/GLM-4.6
+      trust_remote_code: true
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+      - 8
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      trust_remote_code: true
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      trust_remote_code: true
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4
+      - 8

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/glm-4.7-flash.yaml

@@ -0,0 +1,126 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: glm-4.7-flash.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: zai-org/GLM-4.7-Flash
+  tokenizer: zai-org/GLM-4.7-Flash
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 1
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: zai-org/GLM-4.7-Flash
+  max_concurrency:
+  - null
+  - 16
+  - 32
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 16.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/glm-4.7-flash
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      model_path: zai-org/GLM-4.7-Flash
+      trust_remote_code: true
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 1
+      trust_remote_code: true
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 1
+      pp_size: 1
+      trust_remote_code: true
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/glm-4.7.yaml

@@ -0,0 +1,130 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: glm-4.7.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: zai-org/GLM-4.7
+  tokenizer: zai-org/GLM-4.7
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 4
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: zai-org/GLM-4.7
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 8.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/glm-4.7
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 4
+      context_length: 9000
+      model_path: zai-org/GLM-4.7
+      trust_remote_code: true
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+      ep_size:
+      - 1
+      - 2
+      - 4
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 4
+      trust_remote_code: true
+      gpu_memory_utilization: 0.9
+      max_model_len: 9000
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 4
+      pp_size: 1
+      trust_remote_code: true
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 9000
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 9000
+      - 16384
+      ep_size:
+      - 1
+      - 2
+      - 4

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/glm-5-fp8.yaml

@@ -0,0 +1,132 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: glm-5-fp8.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: zai-org/GLM-5-FP8
+  tokenizer: zai-org/GLM-5-FP8
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: zai-org/GLM-5-FP8
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/glm-5-fp8
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      model_path: zai-org/GLM-5-FP8
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+      - 8
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4
+      - 8

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/glyph.yaml

@@ -0,0 +1,126 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: glyph.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: zai-org/Glyph
+  tokenizer: zai-org/Glyph
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 4
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: zai-org/Glyph
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 8.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/glyph
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 4
+      reasoning_parser: glm45
+      tool_call_parser: glm45
+      model_path: zai-org/Glyph
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 4
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 4
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/gpt-oss-120b.yaml

@@ -0,0 +1,132 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: gpt-oss-120b.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: openai/gpt-oss-120b
+  tokenizer: openai/gpt-oss-120b
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: openai/gpt-oss-120b
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/gpt-oss-120b
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      model_path: openai/gpt-oss-120b
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+      - 8
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4
+      - 8

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/intern-s1.yaml

@@ -0,0 +1,135 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: intern-s1.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: internlm/Intern-S1
+  tokenizer: internlm/Intern-S1
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: internlm/Intern-S1
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/intern-s1
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      trust_remote_code: true
+      model_path: internlm/Intern-S1
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+      - 8
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4
+      - 8

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/kimi-k2-instruct.yaml

@@ -0,0 +1,133 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: kimi-k2-instruct.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: moonshotai/Kimi-K2-Instruct
+  tokenizer: moonshotai/Kimi-K2-Instruct
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: moonshotai/Kimi-K2-Instruct
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/kimi-k2-instruct
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      trust_remote_code: true
+      model_path: moonshotai/Kimi-K2-Instruct
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/kimi-k2.5.yaml

@@ -0,0 +1,127 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: kimi-k2.5.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: moonshotai/Kimi-K2.5
+  tokenizer: moonshotai/Kimi-K2.5
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: moonshotai/Kimi-K2.5
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/kimi-k2.5
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      trust_remote_code: true
+      model_path: moonshotai/Kimi-K2.5
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/kimi-linear-48b-a3b-instruct.yaml

@@ -0,0 +1,121 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: kimi-linear-48b-a3b-instruct.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: moonshotai/Kimi-Linear-48B-A3B-Instruct
+  tokenizer: moonshotai/Kimi-Linear-48B-A3B-Instruct
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 4
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: moonshotai/Kimi-Linear-48B-A3B-Instruct
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 8.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/kimi-linear-48b-a3b-instruct
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 4
+      trust_remote_code: true
+      model_path: moonshotai/Kimi-Linear-48B-A3B-Instruct
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 4
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 4
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/ling-2.5-1t.yaml

@@ -0,0 +1,134 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: ling-2.5-1t.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: inclusionAI/Ling-2.5-1T
+  tokenizer: inclusionAI/Ling-2.5-1T
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: true
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: inclusionAI/Ling-2.5-1T
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 2.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/ling-2.5-1t
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      pp_size: 2
+      nnodes: 2
+      trust_remote_code: true
+      tool_call_parser: qwen
+      model_path: inclusionAI/Ling-2.5-1T
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      pp_size:
+      - 1
+      - 2
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      pipeline_parallel_size: 2
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 2
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/llada2-1-mini.yaml

@@ -0,0 +1,130 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: llada2-1-mini.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: inclusionAI/LLaDA2.1-mini
+  tokenizer: inclusionAI/LLaDA2.1-mini
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 1
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: inclusionAI/LLaDA2.1-mini
+  max_concurrency:
+  - 1
+  - 2
+  - 4
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/llada2-1-mini
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 1
+      dllm_algorithm: JointThreshold
+      trust_remote_code: true
+      max_running_requests: 1
+      attention_backend: flashinfer
+      model_path: inclusionAI/LLaDA2.1-mini
+      mem_fraction_static: 0.77
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 1
+      - 2
+      - 4
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 1
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 1
+      - 2
+      - 4
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 1
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 1
+      - 2
+      - 4
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/llama-3.1-70b-instruct.yaml

@@ -0,0 +1,124 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: llama-3.1-70b-instruct.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: meta-llama/Llama-3.1-70B-Instruct
+  tokenizer: meta-llama/Llama-3.1-70B-Instruct
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 4
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: meta-llama/Llama-3.1-70B-Instruct
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 12.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/llama-3.1-70b-instruct
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 4
+      model_path: meta-llama/Llama-3.1-70B-Instruct
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 4
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 4
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/llama-3.3-70b-instruct.yaml

@@ -0,0 +1,118 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: llama-3.3-70b-instruct.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: meta-llama/Llama-3.3-70B-Instruct
+  tokenizer: meta-llama/Llama-3.3-70B-Instruct
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 1
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: meta-llama/Llama-3.3-70B-Instruct
+  max_concurrency:
+  - null
+  - 16
+  - 32
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 16.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/llama-3.3-70b-instruct
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tool_call_parser: llama3
+      model_path: meta-llama/Llama-3.3-70B-Instruct
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 1
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 1
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/llama-4-maverick-17b-128e-instruct-fp8.yaml

@@ -0,0 +1,122 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: llama-4-maverick-17b-128e-instruct-fp8.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8
+  tokenizer: meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8
+  max_concurrency:
+  - null
+  - 2
+  - 4
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 2.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/llama-4-maverick-17b-128e-instruct-fp8
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      context_length: 1000000
+      trust_remote_code: true
+      enable_multimodal: true
+      model_path: meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 4
+      - 8
+      - 12
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 1000000
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 4
+      - 8
+      - 12
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 1000000
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 4
+      - 8
+      - 12
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 1000000

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/llama-4-scout-17b-16e-instruct.yaml

@@ -0,0 +1,129 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: llama-4-scout-17b-16e-instruct.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: meta-llama/Llama-4-Scout-17B-16E-Instruct
+  tokenizer: meta-llama/Llama-4-Scout-17B-16E-Instruct
+  precision: bfloat16
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: meta-llama/Llama-4-Scout-17B-16E-Instruct
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/llama-4-scout-17b-16e-instruct
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      enable_multimodal: true
+      context_length: 65536
+      dtype: bfloat16
+      trust_remote_code: true
+      model_path: meta-llama/Llama-4-Scout-17B-16E-Instruct
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 8
+      - 16
+      - 24
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 65536
+      dtype: bfloat16
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 8
+      - 16
+      - 24
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 65536
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 8
+      - 16
+      - 24
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 65536

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/mimo-v2-flash.yaml

@@ -0,0 +1,133 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: mimo-v2-flash.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: XiaomiMiMo/MiMo-V2-Flash
+  tokenizer: XiaomiMiMo/MiMo-V2-Flash
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: XiaomiMiMo/MiMo-V2-Flash
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/mimo-v2-flash
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      trust_remote_code: true
+      max_running_requests: 128
+      chunked_prefill_size: 16384
+      model_loader_extra_config: '{"enable_multithread_load": "true","num_threads": 64}'
+      attention_backend: fa3
+      reasoning_parser: qwen3
+      tool_call_parser: mimo
+      model_path: XiaomiMiMo/MiMo-V2-Flash
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/minimax-m2.1.yaml

@@ -0,0 +1,121 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: minimax-m2.1.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: MiniMaxAI/MiniMax-M2.1
+  tokenizer: MiniMaxAI/MiniMax-M2.1
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 4
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: MiniMaxAI/MiniMax-M2.1
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 8.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/minimax-m2.1
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 4
+      trust_remote_code: true
+      model_path: MiniMaxAI/MiniMax-M2.1
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 4
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 4
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/minimax-m2.5.yaml

@@ -0,0 +1,133 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: minimax-m2.5.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: MiniMaxAI/MiniMax-M2.5
+  tokenizer: MiniMaxAI/MiniMax-M2.5
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 4
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: MiniMaxAI/MiniMax-M2.5
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/minimax-m2.5
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 4
+      trust_remote_code: true
+      model_path: MiniMaxAI/MiniMax-M2.5
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+      ep_size:
+      - 1
+      - 4
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 4
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 4
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/ministral-3-8b-instruct-2512.yaml

@@ -0,0 +1,121 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: ministral-3-8b-instruct-2512.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: mistralai/Ministral-3-8B-Instruct-2512
+  tokenizer: mistralai/Ministral-3-8B-Instruct-2512
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 1
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: mistralai/Ministral-3-8B-Instruct-2512
+  max_concurrency:
+  - null
+  - 16
+  - 32
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 16.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/ministral-3-8b-instruct-2512
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      trust_remote_code: true
+      tool_call_parser: mistral
+      model_path: mistralai/Ministral-3-8B-Instruct-2512
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 1
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 1
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/mistral-small-4-119b-2603.yaml

@@ -0,0 +1,124 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: mistral-small-4-119b-2603.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: mistralai/Mistral-Small-4-119B-2603
+  tokenizer: mistralai/Mistral-Small-4-119B-2603
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 2
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: mistralai/Mistral-Small-4-119B-2603
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 6.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/mistral-small-4-119b-2603
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 2
+      model_path: mistralai/Mistral-Small-4-119B-2603
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 2
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 2
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/nemotron-3-nano-30b-a3b-bf16.yaml

@@ -0,0 +1,128 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: nemotron-3-nano-30b-a3b-bf16.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
+  tokenizer: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 1
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
+  max_concurrency:
+  - null
+  - 16
+  - 32
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 16.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/nemotron-3-nano-30b-a3b-bf16
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 1
+      trust_remote_code: true
+      kv_cache_dtype: fp8_e4m3
+      model_path: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 1
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: fp8_e4m3
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 1
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/nemotron-3-super-120b-a12b-bf16.yaml

@@ -0,0 +1,128 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: nemotron-3-super-120b-a12b-bf16.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
+  tokenizer: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 4
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 6.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/nemotron-3-super-120b-a12b-bf16
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 4
+      trust_remote_code: true
+      kv_cache_dtype: fp8_e4m3
+      model_path: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 4
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: fp8_e4m3
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 4
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/qwen3-235b-a22b.yaml

@@ -0,0 +1,132 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: qwen3-235b-a22b.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: Qwen/Qwen3-235B-A22B
+  tokenizer: Qwen/Qwen3-235B-A22B
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: Qwen/Qwen3-235B-A22B
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/qwen3-235b-a22b
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      model_path: Qwen/Qwen3-235B-A22B
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+      - 8
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4
+      - 8

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/qwen3-coder-480b-a35b-instruct.yaml

@@ -0,0 +1,131 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: qwen3-coder-480b-a35b-instruct.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: Qwen/Qwen3-Coder-480B-A35B-Instruct
+  tokenizer: Qwen/Qwen3-Coder-480B-A35B-Instruct
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: Qwen/Qwen3-Coder-480B-A35B-Instruct
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/qwen3-coder-480b-a35b-instruct
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      ep_size: 2
+      moe_runner_backend: triton
+      model_path: Qwen/Qwen3-Coder-480B-A35B-Instruct
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - flashinfer
+      decode_attention_backend:
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 2
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      ep_size: 2
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 2

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/qwen3-coder-next.yaml

@@ -0,0 +1,124 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: qwen3-coder-next.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: Qwen/Qwen3-Coder-Next
+  tokenizer: Qwen/Qwen3-Coder-Next
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 2
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: Qwen/Qwen3-Coder-Next
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 12.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/qwen3-coder-next
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 2
+      model_path: Qwen/Qwen3-Coder-Next
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 2
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 2
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/qwen3-next-80b-a3b-instruct.yaml

@@ -0,0 +1,130 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: qwen3-next-80b-a3b-instruct.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: Qwen/Qwen3-Next-80B-A3B-Instruct
+  tokenizer: Qwen/Qwen3-Next-80B-A3B-Instruct
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 2
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: Qwen/Qwen3-Next-80B-A3B-Instruct
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 12.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/qwen3-next-80b-a3b-instruct
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 2
+      model_path: Qwen/Qwen3-Next-80B-A3B-Instruct
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+      ep_size:
+      - 1
+      - 2
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 2
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 2
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 2

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/qwen35-397b-a17b-fp8.yaml

@@ -0,0 +1,132 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: qwen35-397b-a17b-fp8.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: Qwen/Qwen3.5-397B-A17B-FP8
+  tokenizer: Qwen/Qwen3.5-397B-A17B-FP8
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 4
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: Qwen/Qwen3.5-397B-A17B-FP8
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 4.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/qwen35-397b-a17b-fp8
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 4
+      model_path: Qwen/Qwen3.5-397B-A17B-FP8
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+      ep_size:
+      - 1
+      - 4
+      - 8
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 4
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 4
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4
+      - 8

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/ring-2.5-1t.yaml

@@ -0,0 +1,124 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: ring-2.5-1t.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: inclusionAI/Ring-2.5-1T
+  tokenizer: inclusionAI/Ring-2.5-1T
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 8
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: inclusionAI/Ring-2.5-1T
+  max_concurrency:
+  - null
+  - 4
+  - 8
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 2.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/ring-2.5-1t
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 8
+      model_path: inclusionAI/Ring-2.5-1T
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 32
+      - 48
+      - 64
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 8
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+    search_space:
+      max_num_seqs:
+      - 32
+      - 48
+      - 64
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 8
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+    search_space:
+      max_batch_size:
+      - 32
+      - 48
+      - 64
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384

.claude/skills/llm-serving-auto-benchmark/configs/cookbook-llm/step-3.5-flash.yaml

@@ -0,0 +1,133 @@
+schema_version: 1
+source:
+  kind: llm_serving_cookbook
+  source_recipe_file: step-3.5-flash.yaml
+  translation: SGLang flags preserve the source recipe where applicable, with sequence limits raised when needed for the default dataset; vLLM and TensorRT-LLM use framework-native serving flags for the same model and dataset shape.
+model:
+  name: stepfun-ai/Step-3.5-Flash
+  tokenizer: stepfun-ai/Step-3.5-Flash
+  precision: auto
+  quantization: model default
+hardware:
+  gpu_count: 4
+  multi_node: false
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names:
+  - chat
+  - summarization
+  input_len:
+  - 1000
+  - 8000
+  output_len:
+  - 1000
+  - 1000
+benchmark:
+  endpoint: /v1/completions
+  backend: openai-compatible
+  tokenizer: stepfun-ai/Step-3.5-Flash
+  max_concurrency:
+  - null
+  - 8
+  - 16
+  extra_request_body:
+    temperature: 0.0
+  qps:
+    lower: 0.25
+    upper: 8.0
+    tolerance: 0.1
+  sla:
+    max_p99_ttft_ms: 1500
+    max_p99_tpot_ms: 30
+    min_success_rate: 0.99
+  output_dir: ./auto_benchmark_results/cookbook-llm/step-3.5-flash
+search:
+  tier: 2
+  max_candidates_per_framework: 8
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+frameworks:
+  sglang:
+    enabled: true
+    server_command: python -m sglang.launch_server
+    base_server_flags:
+      tp_size: 4
+      trust_remote_code: true
+      model_path: stepfun-ai/Step-3.5-Flash
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+    search_space:
+      prefill_attention_backend:
+      - fa3
+      - flashinfer
+      decode_attention_backend:
+      - fa3
+      - flashinfer
+      chunked_prefill_size:
+      - 4096
+      - 8192
+      max_running_requests:
+      - 64
+      - 96
+      - 128
+      ep_size:
+      - 1
+      - 4
+  vllm:
+    enabled: true
+    server_command: vllm serve
+    config_source: framework_generic_translation
+    base_server_flags:
+      tensor_parallel_size: 4
+      gpu_memory_utilization: 0.9
+      max_model_len: 12288
+      dtype: auto
+      enable_chunked_prefill: true
+      kv_cache_dtype: auto
+      trust_remote_code: true
+    search_space:
+      max_num_seqs:
+      - 64
+      - 96
+      - 128
+      max_num_batched_tokens:
+      - 8192
+      - 16384
+      max_num_partial_prefills:
+      - 1
+      max_long_partial_prefills:
+      - 1
+      long_prefill_token_threshold:
+      - 0
+      - 4096
+      enable_prefix_caching:
+      - true
+      block_size:
+      - 16
+  tensorrt_llm:
+    enabled: true
+    server_command: trtllm-serve serve
+    backend_policy: fixed_pytorch
+    config_source: framework_generic_translation
+    base_server_flags:
+      backend: pytorch
+      tp_size: 4
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      max_seq_len: 12288
+      trust_remote_code: true
+    search_space:
+      max_batch_size:
+      - 64
+      - 96
+      - 128
+      max_num_tokens:
+      - 8192
+      - 16384
+      max_seq_len:
+      - 12288
+      - 16384
+      ep_size:
+      - 1
+      - 4

.claude/skills/llm-serving-auto-benchmark/references/container-runbook.md

@@ -0,0 +1,321 @@
+# Container Runbook
+
+Use this runbook when the benchmark environment is container-based. It records
+the exact image, command, help output, server log, benchmark log, and cleanup
+step for each framework.
+
+This runbook is target-agnostic. Every `docker run` / `docker exec` command
+works on a local box, an SSH-reachable remote GPU host, or a CI runner; the
+per-host skills (for example `h100`, `b200`, `rtx5090`, `radixark02`,
+`radixark03`) only add the SSH wrapper, container name, and workspace path
+for a specific operator box. Substitute those values where you see
+`$SGLANG_CONTAINER`, `$SGLANG_WORKSPACE`, and similar; nothing below assumes
+an H100.
+
+## Common Setup
+
+Pull the images that will be used:
+
+```bash
+docker pull lmsysorg/sglang:dev
+docker pull vllm/vllm-openai:latest
+docker pull nvcr.io/nvidia/tensorrt-llm/release:latest
+```
+
+Use quoted Docker GPU device lists:
+
+```bash
+GPU_ARG='"device=6,7"'
+docker run --gpus "$GPU_ARG" ...
+```
+
+The unquoted form `--gpus device=6,7` can be parsed incorrectly by Docker.
+
+Mount the shared Hugging Face cache and pass tokens through environment variables
+when gated models are used:
+
+```bash
+-v /data/.cache:/root/.cache \
+-e HF_TOKEN \
+-e HUGGINGFACE_HUB_TOKEN
+```
+
+Do not print token values into logs.
+
+Set the run variables once and pass them into containers that need them:
+
+```bash
+export MODEL=TinyLlama/TinyLlama-1.1B-Chat-v1.0
+export TP=1
+export PP=1
+export PORT=8000
+export RUN_DIR=/tmp/llm-serving-auto-benchmark
+mkdir -p "$RUN_DIR"
+```
+
+For synthetic validation, use two aligned scenarios rather than one tiny request
+shape:
+
+```bash
+# chat-like
+RANDOM_INPUT_LEN=1000
+RANDOM_OUTPUT_LEN=1000
+
+# summarization-like
+RANDOM_INPUT_LEN=8000
+RANDOM_OUTPUT_LEN=1000
+```
+
+For a fast smoke on larger models, 20 prompts per scenario is a reasonable
+minimum. Do not treat that as a performance result.
+
+Set each framework's sequence-length limit to cover the largest scenario. For
+the example above, use at least 9000 tokens for SGLang `--context-length`, vLLM
+`--max-model-len`, and TensorRT-LLM `--max_seq_len`.
+
+Before launching a server, save the help output:
+
+```bash
+python -m sglang.launch_server --help > artifacts/help/sglang_launch_server.txt
+python -m sglang.bench_serving --help > artifacts/help/sglang_bench_serving.txt
+vllm serve --help=all > artifacts/help/vllm_serve_all.txt
+vllm bench serve --help=all > artifacts/help/vllm_bench_serve_all.txt
+vllm bench sweep serve --help=all > artifacts/help/vllm_bench_sweep_serve_all.txt
+trtllm-serve serve --help > artifacts/help/trtllm_serve.txt
+python -m tensorrt_llm.serve.scripts.benchmark_serving --help \
+  > artifacts/help/trtllm_benchmark_serving.txt
+```
+
+## SGLang
+
+If a prepared GPU host already has a long-running SGLang container (local or
+reached via ssh; name is operator-specific), reuse it via `docker exec`
+instead of creating a new container. The per-host skills — `h100`,
+`h100-sglang-diffusion`, `b200`, `rtx5090`, `radixark02`, `radixark03`,
+and similar — provide the concrete container name and workspace path for
+that box; this runbook assumes the operator substitutes them:
+
+```bash
+docker exec \
+  -e MODEL \
+  -e TP \
+  -e PORT \
+  "$SGLANG_CONTAINER" bash -lc "
+cd \"\$SGLANG_WORKSPACE\"
+python -m sglang.launch_server \\
+  --model-path \"\$MODEL\" \\
+  --tp-size \"\$TP\" \\
+  --host 0.0.0.0 \\
+  --port \"\$PORT\"
+"
+```
+
+For a fresh container:
+
+```bash
+docker run -d --name llmbench-sglang \
+  --gpus "$GPU_ARG" \
+  --network host \
+  --ipc=host \
+  -v /data/.cache:/root/.cache \
+  -e MODEL \
+  -e TP \
+  -e PORT \
+  -e HF_TOKEN \
+  -e HUGGINGFACE_HUB_TOKEN \
+  --entrypoint bash \
+  lmsysorg/sglang:dev -lc '
+python -m sglang.launch_server \
+  --model-path "$MODEL" \
+  --tp-size "$TP" \
+  --host 0.0.0.0 \
+  --port "$PORT"
+'
+```
+
+Then run either SGLang auto benchmark:
+
+```bash
+python -m sglang.auto_benchmark run --config /path/to/sglang.yaml
+```
+
+or a tiny OpenAI-compatible smoke benchmark:
+
+```bash
+python -m sglang.bench_serving \
+  --backend sglang-oai \
+  --host 127.0.0.1 \
+  --port "$PORT" \
+  --dataset-name random \
+  --random-input-len 32 \
+  --random-output-len 8 \
+  --num-prompts 4 \
+  --request-rate 1 \
+  --max-concurrency 2 \
+  --output-file "$RUN_DIR/sglang/results.json" \
+  --output-details
+```
+
+## vLLM
+
+Server template:
+
+```bash
+docker run -d --name llmbench-vllm \
+  --gpus "$GPU_ARG" \
+  --network host \
+  --ipc=host \
+  -v /data/.cache:/root/.cache \
+  -e MODEL \
+  -e TP \
+  -e PORT \
+  -e HF_TOKEN \
+  -e HUGGINGFACE_HUB_TOKEN \
+  --entrypoint bash \
+  vllm/vllm-openai:latest -lc '
+vllm serve "$MODEL" \
+  --host 0.0.0.0 \
+  --port "$PORT" \
+  --tensor-parallel-size "$TP" \
+  --dtype auto \
+  --gpu-memory-utilization 0.90 \
+  --max-model-len 4096 \
+  --max-num-seqs 64 \
+  --max-num-batched-tokens 8192 \
+  --enable-chunked-prefill \
+  --kv-cache-dtype auto \
+  --enable-prefix-caching \
+  --trust-remote-code
+'
+```
+
+Benchmark template:
+
+```bash
+docker run --rm \
+  --network host \
+  -v /data/.cache:/root/.cache \
+  -v "$RUN_DIR:/artifacts" \
+  -e MODEL \
+  -e PORT \
+  --entrypoint bash \
+  vllm/vllm-openai:latest -lc '
+vllm bench serve \
+  --backend vllm \
+  --base-url "http://127.0.0.1:$PORT" \
+  --model "$MODEL" \
+  --dataset-name random \
+  --random-input-len 1024 \
+  --random-output-len 256 \
+  --num-prompts 80 \
+  --request-rate 8 \
+  --max-concurrency 64 \
+  --save-result \
+  --result-dir /artifacts/vllm \
+  --result-filename results.json
+'
+```
+
+Use `vllm bench sweep serve` when the target image supports it and the search
+can be described with serve/bench parameter JSON files.
+
+## TensorRT-LLM
+
+This skill only supports the TensorRT-LLM PyTorch server backend. Keep
+`--backend pytorch` in every `trtllm-serve serve` command. Do not switch the
+server to `--backend trt`, an engine path, or any other backend; mark that
+candidate unsupported instead.
+
+For single-node multi-GPU TensorRT-LLM containers, keep the IPC, ulimit, shared
+memory, and NCCL settings below. In a multi-GPU PyTorch-backend validation
+run (captured on an H100 host; the rule is not H100-specific), the server
+entered `PyTorchConfig` but failed NCCL allreduce without these container
+options; the same model and candidate list passed after adding them. Expect
+the same requirement on any single-node multi-GPU target.
+
+Server template:
+
+```bash
+docker run -d --name llmbench-trtllm \
+  --gpus "$GPU_ARG" \
+  --ipc=host \
+  --ulimit memlock=-1 \
+  --ulimit stack=67108864 \
+  --shm-size=16g \
+  --network host \
+  -v /data/.cache:/root/.cache \
+  -e MODEL \
+  -e TP \
+  -e PP \
+  -e PORT \
+  -e HF_TOKEN \
+  -e HUGGINGFACE_HUB_TOKEN \
+  -e NCCL_IB_DISABLE=1 \
+  --entrypoint bash \
+  nvcr.io/nvidia/tensorrt-llm/release:latest -lc '
+trtllm-serve serve "$MODEL" \
+  --host 0.0.0.0 \
+  --port "$PORT" \
+  --backend pytorch \
+  --tp_size "$TP" \
+  --pp_size "$PP" \
+  --max_batch_size 64 \
+  --max_num_tokens 8192 \
+  --max_seq_len 4096 \
+  --kv_cache_free_gpu_memory_fraction 0.75 \
+  --trust_remote_code
+'
+```
+
+Benchmark template:
+
+```bash
+docker run --rm \
+  --network host \
+  -v /data/.cache:/root/.cache \
+  -v "$RUN_DIR:/artifacts" \
+  -e MODEL \
+  -e PORT \
+  --entrypoint bash \
+  nvcr.io/nvidia/tensorrt-llm/release:latest -lc '
+python -m tensorrt_llm.serve.scripts.benchmark_serving \
+  --backend openai \
+  --host 127.0.0.1 \
+  --port "$PORT" \
+  --endpoint /v1/completions \
+  --model "$MODEL" \
+  --dataset-name random \
+  --random-input-len 1024 \
+  --random-output-len 256 \
+  --random-ids \
+  --num-prompts 80 \
+  --request-rate 8 \
+  --max-concurrency 64 \
+  --save-result \
+  --result-dir /artifacts/trtllm \
+  --result-filename results.json
+'
+```
+
+For TensorRT-LLM 1.0.0, the serving benchmark client `--backend` choices are
+`openai` and `openai-chat`. Do not pass `--backend trtllm`. This client flag is
+separate from the server backend pinned above.
+
+## Cleanup
+
+Use unique container names per run and clean up by name:
+
+```bash
+docker rm -f llmbench-sglang llmbench-vllm llmbench-trtllm
+```
+
+If a port remains bound after container cleanup, inspect it before killing
+anything:
+
+```bash
+ss -ltnp | grep ':8000'
+ps -eo pid,ppid,user,etime,cmd | grep '<model-or-port>'
+```
+
+Only kill raw PIDs when the command line proves they belong to the current
+validation run.

.claude/skills/llm-serving-auto-benchmark/references/example-plan.yaml

@@ -0,0 +1,133 @@
+# Example run plan for the llm-serving-auto-benchmark skill. Baseline flags stay
+# in base_server_flags, search knobs stay in search_space, and aligned dataset
+# length pairs define scenarios.
+#
+# Note: this is the runtime plan shape (top-level `sla`, no `schema_version` or
+# `server_command`). Cookbook configs in configs/cookbook-llm/ use the extended
+# schema enforced by scripts/validate_cookbook_configs.py; do not run the
+# validator against this file as-is.
+
+model:
+  name: Qwen/Qwen3-32B
+  tokenizer: Qwen/Qwen3-32B
+  precision: bf16
+  quantization: none
+
+version_manifest:
+  sglang:
+    container_image: lmsysorg/sglang:dev
+    package_version: null
+    git_commit: null
+    server_help: artifacts/help/sglang_launch_server.txt
+    benchmark_help: artifacts/help/sglang_bench_serving.txt
+  vllm:
+    container_image: vllm/vllm-openai:latest
+    package_version: null
+    git_commit: null
+    server_help: artifacts/help/vllm_serve_all.txt
+    benchmark_help: artifacts/help/vllm_bench_serve_all.txt
+    sweep_help: artifacts/help/vllm_bench_sweep_serve_all.txt
+  tensorrt_llm:
+    container_image: nvcr.io/nvidia/tensorrt-llm/release:latest
+    package_version: null
+    git_commit: null
+    server_help: artifacts/help/trtllm_serve.txt
+    benchmark_help: artifacts/help/trtllm_benchmark_serving.txt
+
+hardware:
+  # Example values; replace with the actual target GPU (A100, H100, H200,
+  # B200, MI300, RTX 5090, etc.). gpu_model is recorded for fairness audit,
+  # not used as a scheduling hint.
+  gpu_model: NVIDIA H100 80GB HBM3
+  gpu_count: 4
+  multi_node: false
+
+dataset:
+  kind: random
+  num_prompts: 80
+  scenario_names: [chat, summarization]
+  input_len: [1000, 8000]
+  output_len: [1000, 1000]
+  canonical_jsonl: null
+
+benchmark:
+  endpoint: /v1/chat/completions
+  backend: auto
+  request_rates: null
+  max_concurrency: [null, 16, 32]
+  qps:
+    lower: 1.0
+    upper: 12.0
+    tolerance: 0.1
+    max_rounds: 5
+  extra_request_body:
+    temperature: 0.0
+
+sla:
+  max_p99_ttft_ms: 2000
+  max_p99_tpot_ms: 80
+  min_success_rate: 0.99
+
+search:
+  tier: 2
+  max_candidates_per_framework: 10
+  candidate_generation: baseline_first_bounded_product
+  resume: true
+  output_dir: /bench/results/llm-serving-auto-benchmark
+
+frameworks:
+  sglang:
+    enabled: true
+    base_server_flags:
+      tp_size: 4
+      trust_remote_code: true
+      mem_fraction_static: 0.82
+      schedule_policy: lpm
+      context_length: 12288
+    search_space:
+      # Verify these names against `python -m sglang.launch_server --help`.
+      prefill_attention_backend: [fa3, flashinfer]
+      decode_attention_backend: [fa3, flashinfer]
+      chunked_prefill_size: [8192, 16384]
+      max_running_requests: [64, 128]
+
+  vllm:
+    enabled: true
+    base_server_flags:
+      tensor_parallel_size: 4
+      trust_remote_code: true
+      gpu_memory_utilization: 0.90
+      max_model_len: 12288
+      dtype: auto
+    search_space:
+      # Verify these names against `vllm serve --help=all`.
+      max_num_seqs: [64, 128]
+      max_num_batched_tokens: [8192, 16384]
+      enable_chunked_prefill: [true]
+      # Raise above 1 only after the target model/runtime supports concurrent partial prefill.
+      max_num_partial_prefills: [1]
+      max_long_partial_prefills: [1]
+      long_prefill_token_threshold: [0, 4096]
+      enable_prefix_caching: [true]
+      kv_cache_dtype: [auto]
+      block_size: [16]
+
+  tensorrt_llm:
+    enabled: true
+    backend_policy: fixed_pytorch
+    base_server_flags:
+      backend: pytorch
+      tp_size: 4
+      pp_size: 1
+      kv_cache_free_gpu_memory_fraction: 0.75
+      trust_remote_code: true
+    search_space:
+      # Verify these names against `trtllm-serve serve --help`.
+      # Do not add backend choices here; TensorRT-LLM is fixed to the PyTorch backend.
+      max_batch_size: [64, 128]
+      max_num_tokens: [8192, 16384]
+      max_seq_len: [12288, 16384]
+      # Uncomment and point at concrete config files to sweep PyTorch-backend
+      # options via --extra_llm_api_options. A single [null] value contributes
+      # no dimension to the search.
+      # extra_llm_api_options: [null, /path/to/trt_llm_config_A.yaml]

.claude/skills/llm-serving-auto-benchmark/references/framework-reference.md

@@ -0,0 +1,113 @@
+# Framework Reference
+
+Use this file when choosing native framework commands or translating tuning
+knobs across SGLang, vLLM, and TensorRT-LLM. Always verify the concrete CLI in
+the target container with `--help` before a long run.
+
+## Native Entry Points
+
+| Framework | Server | Benchmark | Notes |
+| --- | --- | --- | --- |
+| SGLang | `python -m sglang.launch_server` | `python -m sglang.auto_benchmark` or `python -m sglang.bench_serving` | Use `auto_benchmark` when available for server-flag search. Use `bench_serving` for direct native or OpenAI-compatible endpoint checks. |
+| vLLM | `vllm serve` | `vllm bench sweep serve` or `vllm bench serve` | Prefer `bench sweep serve` when sweeping server and benchmark parameter JSON files. |
+| TensorRT-LLM | `trtllm-serve serve --backend pytorch` | TensorRT-LLM serving benchmark client or a common OpenAI-compatible client | This skill does not cover engine-backed serving or non-PyTorch server backends. |
+
+Common source docs:
+
+- SGLang bench serving: <https://docs.sglang.ai/developer_guide/bench_serving.html>
+- vLLM benchmark sweeps: <https://docs.vllm.ai/en/latest/benchmarking/sweeps/>
+- vLLM `bench sweep serve`: <https://docs.vllm.ai/en/latest/cli/bench/sweep/serve.html>
+- TensorRT-LLM `trtllm-serve`: <https://nvidia.github.io/TensorRT-LLM/commands/trtllm-serve/trtllm-serve.html>
+- TensorRT-LLM deployment guide: <https://nvidia.github.io/TensorRT-LLM/deployment-guide/index.html>
+
+## Command Templates
+
+### SGLang
+
+```bash
+python -m sglang.launch_server \
+  --model-path <model> \
+  --tp-size <tp> \
+  --port 30000
+
+python -m sglang.bench_serving \
+  --backend sglang-oai \
+  --host 127.0.0.1 \
+  --port 30000 \
+  --dataset-name random \
+  --random-input-len 1024 \
+  --random-output-len 256 \
+  --num-prompts 80 \
+  --request-rate 8
+```
+
+Use `--backend sglang` for SGLang-native `/generate` checks. Use
+`--backend sglang-oai` when comparing against vLLM or TensorRT-LLM through an
+OpenAI-compatible path.
+
+### vLLM
+
+```bash
+vllm serve <model> \
+  --host 0.0.0.0 \
+  --port 8000 \
+  --tensor-parallel-size <tp> \
+  --gpu-memory-utilization 0.90 \
+  --max-model-len 4096 \
+  --max-num-seqs 64 \
+  --max-num-batched-tokens 8192 \
+  --enable-chunked-prefill
+
+vllm bench serve \
+  --backend vllm \
+  --base-url http://127.0.0.1:8000 \
+  --model <model> \
+  --dataset-name random \
+  --random-input-len 1024 \
+  --random-output-len 256 \
+  --num-prompts 80
+```
+
+### TensorRT-LLM
+
+```bash
+trtllm-serve serve <model> \
+  --backend pytorch \
+  --tp_size <tp> \
+  --kv_cache_free_gpu_memory_fraction 0.75 \
+  --host 0.0.0.0 \
+  --port 8000
+```
+
+Benchmark the OpenAI-compatible endpoint with the TensorRT-LLM serving benchmark
+client or the same OpenAI-compatible client used for the other frameworks. Keep
+server backend choice fixed to `pytorch`.
+
+## Knob Family Mapping
+
+Do not copy flag names across frameworks. Compare knob families, then translate
+to the target CLI.
+
+| Family | SGLang | vLLM | TensorRT-LLM |
+| --- | --- | --- | --- |
+| Parallelism | `--tp-size`, `--pp-size`, `--dp-size`, `--ep-size`, `--expert-parallel-size` | `--tensor-parallel-size`, `--pipeline-parallel-size`, `--data-parallel-size`, `--enable-expert-parallel` | `--tp_size`, `--pp_size`, `--ep_size`, `--gpus_per_node`, `--cluster_size` |
+| Memory and KV cache | `--mem-fraction-static`, `--max-total-tokens`, `--kv-cache-dtype`, `--page-size`, `--cpu-offload-gb` | `--gpu-memory-utilization`, `--kv-cache-memory-bytes`, `--kv-cache-dtype`, `--block-size`, `--cpu-offload-gb` | `--kv_cache_free_gpu_memory_fraction`, plus `--max_num_tokens`, `--max_seq_len`, `--max_batch_size` |
+| Batching and scheduler | `--max-running-requests`, `--schedule-policy`, `--chunked-prefill-size`, `--max-prefill-tokens`, `--prefill-max-requests` | `--max-num-seqs`, `--max-num-batched-tokens`, `--enable-chunked-prefill`, partial-prefill and DBO flags | `--max_batch_size`, `--max_num_tokens`, `--max_seq_len`; extra scheduler knobs may require `--extra_llm_api_options` |
+| Attention/backend | `--attention-backend`, `--prefill-attention-backend`, `--decode-attention-backend`, `--sampling-backend` | `--attention-backend`, `--gdn-prefill-backend`, `--mm-encoder-attn-backend` | `--backend pytorch` is fixed; do not search backend choice |
+| CUDA graph and compile | `--disable-cuda-graph`, `--cuda-graph-bs`, `--cuda-graph-max-bs`, `--disable-piecewise-cuda-graph`, `--enable-torch-compile` | `--enforce-eager`, `--compilation-config`, `--cudagraph-capture-sizes`, `--max-cudagraph-capture-size` | use direct flags or `--extra_llm_api_options`; record resolved PyTorch config from logs |
+| Prefix/speculative | `--disable-radix-cache`, `--disable-chunked-prefix-cache`, speculative decoding flags | `--enable-prefix-caching`, `--speculative-config` | only use PyTorch-backend options accepted by the target image |
+| Dtype, quantization, loading | `--dtype`, `--quantization`, `--load-format`, `--model-loader-extra-config`, `--trust-remote-code` | `--dtype`, `--quantization`, `--load-format`, `--model-loader-extra-config`, `--trust-remote-code`, `--hf-token` | `--trust_remote_code`, `--tokenizer`; engine build and non-PyTorch quantization flows are out of scope |
+
+## Version Rules
+
+Framework CLIs move quickly. For every real run:
+
+1. Record the framework package version, git commit, image tag, and help files.
+2. Validate concrete flags with
+   `scripts/validate_cookbook_configs.py --help-dir <artifact-help-dir>`.
+3. Move renamed or removed flags out of the run plan before benchmarking.
+4. Record which frameworks were model-smoked and which only passed preflight.
+
+Historical validation from April 2026 used SGLang `0.5.10rc0`, vLLM `0.19.1`,
+and TensorRT-LLM `1.0.0`. Treat those notes as old evidence, not as current
+compatibility guarantees.

.claude/skills/llm-serving-auto-benchmark/references/result-schema.md

@@ -0,0 +1,161 @@
+# Result Schema
+
+Write one JSON object per candidate. Keep failed candidates in the same file so
+the final summary explains what was tried.
+
+## SLA Key Convention
+
+One canonical naming across this skill. Config files and normalized result rows
+must agree.
+
+| Key | Where | Type |
+| --- | --- | --- |
+| `max_p99_ttft_ms` | both | float, milliseconds, p99 |
+| `max_p99_tpot_ms` | both | float, milliseconds, p99 |
+| `min_success_rate` | both | float in [0, 1] |
+| `passed` | result only | bool; recomputed after the run |
+
+Do not use `max_ttft_ms` or `max_tpot_ms` without the `p99_` prefix; those names
+hide whether the target is a mean or a tail. Older cookbook configs used mean
+latency targets by accident and have been migrated to the p99 names above.
+
+The config-level SLA block lives under `benchmark.sla` (cookbook configs) or at
+the top level (example plan). Either location is acceptable, but the key names
+must match this table.
+
+## JSONL Row
+
+The values below (`gpu_model`, `gpu_count`, file paths, numeric metrics, etc.)
+are illustrative. Replace them with the actual target hardware and measured
+values; this schema is not tied to H100.
+
+```json
+{
+  "framework": "sglang",
+  "framework_version": "0.5.0",
+  "framework_commit": "abcdef0",
+  "candidate_id": "sglang-tp8-flashinfer",
+  "model": "meta-llama/Llama-3.1-70B-Instruct",
+  "status": "ok",
+  "failure_reason": "",
+  "hardware": {
+    "gpu_model": "NVIDIA H100 80GB HBM3",
+    "gpu_count": 8,
+    "visible_devices": "0,1,2,3,4,5,6,7"
+  },
+  "workload": {
+    "kind": "custom",
+    "scenario": "chat",
+    "dataset_path": "/bench/workload.autobench.jsonl",
+    "input_len": 2048,
+    "output_len": 512,
+    "input_len_p50": 1800,
+    "input_len_p95": 4096,
+    "output_len_p50": 384,
+    "output_len_p95": 1024,
+    "num_prompts": 1000,
+    "request_rate": 16,
+    "max_concurrency": 256,
+    "endpoint": "/v1/chat/completions"
+  },
+  "sla": {
+    "max_p99_ttft_ms": 2000,
+    "max_p99_tpot_ms": 80,
+    "min_success_rate": 0.99,
+    "passed": true
+  },
+  "metrics": {
+    "request_throughput": 15.8,
+    "output_token_throughput": 12500.0,
+    "total_token_throughput": 42000.0,
+    "mean_ttft_ms": 430.0,
+    "p99_ttft_ms": 1550.0,
+    "mean_tpot_ms": 26.0,
+    "p99_tpot_ms": 72.0,
+    "mean_e2e_ms": 8200.0,
+    "p99_e2e_ms": 19000.0,
+    "success_rate": 0.995
+  },
+  "server_command": "python -m sglang.launch_server ...",
+  "benchmark_command": "python -m sglang.bench_serving ...",
+  "validated_cli_flags": {
+    "server": ["tp_size", "attention_backend"],
+    "benchmark": ["dataset_name", "request_rate", "max_concurrency"]
+  },
+  "artifacts": {
+    "server_log": "/bench/sglang/server.log",
+    "raw_result": "/bench/sglang/results.jsonl",
+    "server_help": "/bench/sglang/help_launch_server.txt",
+    "benchmark_help": "/bench/sglang/help_bench_serving.txt"
+  }
+}
+```
+
+`input_len` and `output_len` are the representative scenario lengths used for
+synthetic workloads or a named bucket. For custom production-like datasets,
+also include p50/p95 buckets when available. These fields let
+`sglang-sota-performance` pass the slow benchmark shape directly into
+`llm-torch-profiler-analysis`:
+
+- prefill profile: `--prefill-input-len <slow input len>` and
+  `--prefill-output-len 1`
+- decode profile: `--decode-input-len 1` and
+  `--decode-output-len <slow output len>`
+
+## Status Values
+
+- `ok`: benchmark finished and metrics are trustworthy
+- `failed`: command failed for a known non-OOM reason
+- `oom`: model or candidate exhausted GPU/host memory
+- `timeout`: server or benchmark timed out
+- `skipped`: intentionally not run, with a reason in `failure_reason`
+
+## Ranking Rule
+
+The default ranking is:
+
+1. `status == "ok"`
+2. `sla.passed == true`
+3. higher `metrics.request_throughput`
+4. higher `metrics.output_token_throughput`
+5. lower `metrics.mean_ttft_ms`
+6. lower `metrics.mean_tpot_ms`
+7. lower `hardware.gpu_count`
+
+If the user cares more about token throughput than request throughput, swap
+steps 3 and 4 and state that in the final report.
+
+This ranking rule does not change the SLA gate. Keep `sla.max_p99_ttft_ms` and
+`sla.max_p99_tpot_ms` as the tail-latency constraints; use mean TTFT and mean
+TPOT only for default winner selection among rows that have already passed SLA.
+
+Missing metric semantics:
+
+- If `metrics.mean_ttft_ms` is absent from a row, the ranking script treats it
+  as the worst possible value, so that row falls below any candidate with a
+  real mean-TTFT measurement. Do not write `0` as a placeholder for "no
+  measurement"; leave the field out or set it to `null`.
+- If `metrics.mean_tpot_ms` is absent from a row, the ranking script treats it
+  as the worst possible value, so that row falls below any candidate with a
+  real mean-TPOT measurement. Do not write `0` as a placeholder for "no
+  measurement"; leave the field out or set it to `null`.
+- If `metrics.request_throughput` or `metrics.output_token_throughput` is
+  missing, the row ranks below any candidate with a real measurement in those
+  keys. A failed candidate that still produced partial metrics should keep the
+  metrics it did produce.
+
+## Final Report Tables
+
+The markdown summary must include these sections:
+
+1. `Best Commands By Framework`: one table per framework. Each table has one row
+   per workload scenario and includes the best candidate, SLA result, throughput,
+   latency metrics, GPU count, exact server command, and artifacts.
+2. `Cross-Framework Best Comparison`: one table that compares the best SGLang,
+   vLLM, and TensorRT-LLM command for each scenario. Sort each scenario by the
+   ranking rule above so the best deployment choice is first.
+3. `Failed Or SLA-Failing Candidates`: include this table when any candidate
+   failed, was skipped, or completed without passing SLA. This table records
+   tried configs that were not selected. Keep each reason concrete enough to
+   tell whether the candidate needs a retry, lower concurrency, a parameter fix,
+   or no further action.

.claude/skills/llm-serving-auto-benchmark/scripts/compare_benchmark_results.py

@@ -0,0 +1,308 @@
+#!/usr/bin/env python3
+"""Summarize normalized cross-framework benchmark JSONL results."""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+from pathlib import Path
+from typing import Any
+
+
+def _get(row: dict[str, Any], path: str, default: Any = None) -> Any:
+    current: Any = row
+    for part in path.split("."):
+        if not isinstance(current, dict) or part not in current:
+            return default
+        current = current[part]
+    return current
+
+
+def _float(row: dict[str, Any], path: str, default: float = 0.0) -> float:
+    value = _get(row, path, default)
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        return default
+
+
+def _bool(row: dict[str, Any], path: str, default: bool = False) -> bool:
+    value = _get(row, path, default)
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        return value.lower() in {"1", "true", "yes", "y"}
+    return bool(value)
+
+
+def _mean_ttft_ms(row: dict[str, Any]) -> float:
+    return _float(row, "metrics.mean_ttft_ms", 1e30)
+
+
+def _mean_tpot_ms(row: dict[str, Any]) -> float:
+    return _float(row, "metrics.mean_tpot_ms", 1e30)
+
+
+def _rank_key(row: dict[str, Any]) -> tuple[Any, ...]:
+    return (
+        _get(row, "status") == "ok",
+        _bool(row, "sla.passed"),
+        _float(row, "metrics.request_throughput"),
+        _float(row, "metrics.output_token_throughput"),
+        -_mean_ttft_ms(row),
+        -_mean_tpot_ms(row),
+        -_float(row, "hardware.gpu_count", 1e30),
+    )
+
+
+def _is_winner_candidate(row: dict[str, Any]) -> bool:
+    return _get(row, "status") == "ok" and _bool(row, "sla.passed")
+
+
+def _fmt(value: Any, digits: int = 2) -> str:
+    if value is None:
+        return ""
+    if isinstance(value, float):
+        return f"{value:.{digits}f}"
+    return str(value)
+
+
+def _cell(value: Any, digits: int = 2) -> str:
+    text = _fmt(value, digits)
+    return text.replace("\n", "<br>").replace("|", "\\|")
+
+
+def _scenario(row: dict[str, Any]) -> str:
+    for path in (
+        "workload.scenario",
+        "workload.scenario_name",
+        "workload.dataset_scenario",
+        "workload.dataset_name",
+        "workload.kind",
+        "scenario",
+    ):
+        value = _get(row, path)
+        if value:
+            return str(value)
+    return "default"
+
+
+def _server_command(row: dict[str, Any]) -> str:
+    return str(_get(row, "server_command") or _get(row, "launch_command") or "")
+
+
+def _artifact_summary(row: dict[str, Any]) -> str:
+    artifacts = _get(row, "artifacts", {})
+    if not isinstance(artifacts, dict):
+        return ""
+    parts = []
+    for key in ("raw_result", "server_log", "benchmark_log", "summary"):
+        value = artifacts.get(key)
+        if value:
+            parts.append(f"{key}: {value}")
+    return "<br>".join(parts)
+
+
+def load_rows(path: Path) -> list[dict[str, Any]]:
+    rows: list[dict[str, Any]] = []
+    with path.open(encoding="utf-8") as f:
+        for line_no, line in enumerate(f, 1):
+            stripped = line.strip()
+            if not stripped:
+                continue
+            try:
+                row = json.loads(stripped)
+            except json.JSONDecodeError as exc:
+                raise SystemExit(f"{path}:{line_no}: invalid JSON: {exc}") from exc
+            if not isinstance(row, dict):
+                raise SystemExit(f"{path}:{line_no}: expected a JSON object")
+            rows.append(row)
+    return rows
+
+
+def best_by_framework_and_scenario(rows: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    best: dict[tuple[str, str], dict[str, Any]] = {}
+    for row in rows:
+        if not _is_winner_candidate(row):
+            continue
+        key = (str(_get(row, "framework", "unknown")), _scenario(row))
+        if key not in best or _rank_key(row) > _rank_key(best[key]):
+            best[key] = row
+    return sorted(
+        best.values(), key=lambda row: (_scenario(row), _rank_key(row)), reverse=True
+    )
+
+
+def write_csv(path: Path, rows: list[dict[str, Any]]) -> None:
+    fields = [
+        "framework",
+        "scenario",
+        "candidate_id",
+        "status",
+        "sla_passed",
+        "request_throughput",
+        "output_token_throughput",
+        "mean_ttft_ms",
+        "mean_tpot_ms",
+        "p99_ttft_ms",
+        "p99_tpot_ms",
+        "gpu_count",
+        "server_command",
+        "failure_reason",
+    ]
+    with path.open("w", encoding="utf-8", newline="") as f:
+        writer = csv.DictWriter(f, fieldnames=fields)
+        writer.writeheader()
+        for row in rows:
+            writer.writerow(
+                {
+                    "framework": _get(row, "framework", ""),
+                    "scenario": _scenario(row),
+                    "candidate_id": _get(row, "candidate_id", ""),
+                    "status": _get(row, "status", ""),
+                    "sla_passed": _bool(row, "sla.passed"),
+                    "request_throughput": _get(row, "metrics.request_throughput", ""),
+                    "output_token_throughput": _get(
+                        row, "metrics.output_token_throughput", ""
+                    ),
+                    "mean_ttft_ms": _get(row, "metrics.mean_ttft_ms", ""),
+                    "mean_tpot_ms": _get(row, "metrics.mean_tpot_ms", ""),
+                    "p99_ttft_ms": _get(row, "metrics.p99_ttft_ms", ""),
+                    "p99_tpot_ms": _get(row, "metrics.p99_tpot_ms", ""),
+                    "gpu_count": _get(row, "hardware.gpu_count", ""),
+                    "server_command": _server_command(row),
+                    "failure_reason": _get(row, "failure_reason", ""),
+                }
+            )
+
+
+def _append_best_commands_by_framework(
+    lines: list[str], scenario_winners: list[dict[str, Any]]
+) -> None:
+    frameworks = sorted(
+        {str(_get(row, "framework", "unknown")) for row in scenario_winners}
+    )
+    lines.extend(["## Best Commands By Framework", ""])
+    for framework in frameworks:
+        lines.extend(
+            [
+                f"### `{framework}`",
+                "",
+                "| Scenario | Candidate | Status | SLA | Req/s | Output tok/s | Total tok/s | Mean TTFT ms | Mean TPOT ms | Success rate | GPUs | Server command | Artifacts |",
+                "| --- | --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- | --- |",
+            ]
+        )
+        rows = [row for row in scenario_winners if _get(row, "framework") == framework]
+        for row in sorted(rows, key=_scenario):
+            lines.append(
+                "| {scenario} | {candidate} | {status} | {sla} | {rps} | {otps} | {ttps} | {ttft} | {tpot} | {success} | {gpus} | {command} | {artifacts} |".format(
+                    scenario=_cell(_scenario(row)),
+                    candidate=_cell(_get(row, "candidate_id", "")),
+                    status=_cell(_get(row, "status", "")),
+                    sla=_cell(_bool(row, "sla.passed")),
+                    rps=_cell(_get(row, "metrics.request_throughput")),
+                    otps=_cell(_get(row, "metrics.output_token_throughput")),
+                    ttps=_cell(_get(row, "metrics.total_token_throughput")),
+                    ttft=_cell(_get(row, "metrics.mean_ttft_ms")),
+                    tpot=_cell(_get(row, "metrics.mean_tpot_ms")),
+                    success=_cell(_get(row, "metrics.success_rate")),
+                    gpus=_cell(_get(row, "hardware.gpu_count")),
+                    command=_cell(_server_command(row)),
+                    artifacts=_cell(_artifact_summary(row)),
+                )
+            )
+        lines.append("")
+
+
+def _append_cross_framework_table(
+    lines: list[str], scenario_winners: list[dict[str, Any]]
+) -> None:
+    lines.extend(
+        [
+            "## Cross-Framework Best Comparison",
+            "",
+            "| Scenario | Rank | Framework | Candidate | SLA | Req/s | Output tok/s | Mean TTFT ms | Mean TPOT ms | GPUs | Server command |",
+            "| --- | ---: | --- | --- | --- | ---: | ---: | ---: | ---: | ---: | --- |",
+        ]
+    )
+    scenario_names = sorted({_scenario(row) for row in scenario_winners})
+    for scenario_name in scenario_names:
+        rows = [row for row in scenario_winners if _scenario(row) == scenario_name]
+        for rank, row in enumerate(sorted(rows, key=_rank_key, reverse=True), 1):
+            lines.append(
+                "| {scenario} | {rank} | {framework} | {candidate} | {sla} | {rps} | {otps} | {ttft} | {tpot} | {gpus} | {command} |".format(
+                    scenario=_cell(scenario_name),
+                    rank=rank,
+                    framework=_cell(_get(row, "framework", "")),
+                    candidate=_cell(_get(row, "candidate_id", "")),
+                    sla=_cell(_bool(row, "sla.passed")),
+                    rps=_cell(_get(row, "metrics.request_throughput")),
+                    otps=_cell(_get(row, "metrics.output_token_throughput")),
+                    ttft=_cell(_get(row, "metrics.mean_ttft_ms")),
+                    tpot=_cell(_get(row, "metrics.mean_tpot_ms")),
+                    gpus=_cell(_get(row, "hardware.gpu_count")),
+                    command=_cell(_server_command(row)),
+                )
+            )
+    lines.append("")
+
+
+def render_markdown(rows: list[dict[str, Any]]) -> str:
+    scenario_winners = best_by_framework_and_scenario(rows)
+
+    lines = ["# Benchmark Summary", ""]
+    if not rows:
+        lines.append("No rows found.")
+        return "\n".join(lines) + "\n"
+
+    _append_best_commands_by_framework(lines, scenario_winners)
+    _append_cross_framework_table(lines, scenario_winners)
+
+    failed = [
+        row
+        for row in rows
+        if _get(row, "status") != "ok" or not _bool(row, "sla.passed")
+    ]
+    if failed:
+        lines.extend(
+            [
+                "",
+                "## Failed Or SLA-Failing Candidates",
+                "",
+                "This table records tried configs that were not selected. They either failed, were skipped by policy, or completed without passing the SLA.",
+                "",
+                "| Framework | Candidate | Status | SLA | Reason |",
+                "| --- | --- | --- | --- | --- |",
+            ]
+        )
+        for row in failed:
+            lines.append(
+                "| {framework} | {candidate} | {status} | {sla} | {reason} |".format(
+                    framework=_cell(_get(row, "framework", "")),
+                    candidate=_cell(_get(row, "candidate_id", "")),
+                    status=_cell(_get(row, "status", "")),
+                    sla=_cell(_bool(row, "sla.passed")),
+                    reason=_cell(_get(row, "failure_reason", "")),
+                )
+            )
+    return "\n".join(lines) + "\n"
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--input", required=True, type=Path, help="Normalized JSONL")
+    parser.add_argument("--output", required=True, type=Path, help="Markdown summary")
+    parser.add_argument("--csv", type=Path, help="Optional CSV table")
+    args = parser.parse_args()
+
+    rows = load_rows(args.input)
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(render_markdown(rows), encoding="utf-8")
+    if args.csv:
+        args.csv.parent.mkdir(parents=True, exist_ok=True)
+        write_csv(args.csv, sorted(rows, key=_rank_key, reverse=True))
+
+
+if __name__ == "__main__":
+    main()

.claude/skills/llm-serving-auto-benchmark/scripts/validate_cookbook_configs.py

@@ -0,0 +1,434 @@
+#!/usr/bin/env python3
+"""Validate cross-framework cookbook benchmark configs.
+
+The validator is intentionally shallow: it proves that every config can be
+loaded, translated into bounded candidate commands, and checked against the
+known server flag surface. It does not launch model servers.
+"""
+
+from __future__ import annotations
+
+import argparse
+import itertools
+import re
+import shlex
+from pathlib import Path
+from typing import Any
+
+import yaml
+
+FRAMEWORKS = ("sglang", "vllm", "tensorrt_llm")
+ALLOWED_SOURCE_KINDS = {"llm_serving_cookbook"}
+
+SEQUENCE_LIMIT_KEY = {
+    "sglang": "context_length",
+    "vllm": "max_model_len",
+    "tensorrt_llm": "max_seq_len",
+}
+
+ALLOWED_SLA_KEYS = {
+    "max_p99_ttft_ms",
+    "max_p99_tpot_ms",
+    "min_success_rate",
+    "max_p99_e2e_ms",
+}
+
+DEPRECATED_SLA_KEYS = {
+    "max_ttft_ms": "max_p99_ttft_ms",
+    "max_tpot_ms": "max_p99_tpot_ms",
+    "max_e2e_ms": "max_p99_e2e_ms",
+}
+
+STATIC_SERVER_FLAGS = {
+    "sglang": {
+        "attention_backend",
+        "chunked_prefill_size",
+        "context_length",
+        "decode_attention_backend",
+        "dllm_algorithm",
+        "dtype",
+        "enable_multimodal",
+        "enable_symm_mem",
+        "ep_size",
+        "host",
+        "kv_cache_dtype",
+        "max_running_requests",
+        "mem_fraction_static",
+        "model_loader_extra_config",
+        "model_path",
+        "moe_runner_backend",
+        "nnodes",
+        "port",
+        "pp_size",
+        "prefill_attention_backend",
+        "reasoning_parser",
+        "schedule_policy",
+        "tool_call_parser",
+        "tp_size",
+        "trust_remote_code",
+    },
+    "vllm": {
+        "block_size",
+        "dtype",
+        "enable_chunked_prefill",
+        "enable_prefix_caching",
+        "gpu_memory_utilization",
+        "host",
+        "kv_cache_dtype",
+        "long_prefill_token_threshold",
+        "max_long_partial_prefills",
+        "max_model_len",
+        "max_num_batched_tokens",
+        "max_num_partial_prefills",
+        "max_num_seqs",
+        "pipeline_parallel_size",
+        "port",
+        "tensor_parallel_size",
+        "trust_remote_code",
+    },
+    "tensorrt_llm": {
+        "backend",
+        "ep_size",
+        "extra_llm_api_options",
+        "host",
+        "kv_cache_free_gpu_memory_fraction",
+        "max_batch_size",
+        "max_num_tokens",
+        "max_seq_len",
+        "port",
+        "pp_size",
+        "tp_size",
+        "trust_remote_code",
+    },
+}
+
+HELP_FILE_HINTS = {
+    "sglang": ("sglang", "launch"),
+    "vllm": ("vllm", "serve"),
+    "tensorrt_llm": ("trtllm", "serve"),
+}
+
+
+def flag_name(framework: str, key: str) -> str:
+    if framework in {"sglang", "vllm"}:
+        return "--" + key.replace("_", "-")
+    return "--" + key
+
+
+def load_yaml(path: Path) -> dict[str, Any]:
+    with path.open(encoding="utf-8") as f:
+        data = yaml.safe_load(f)
+    if not isinstance(data, dict):
+        raise ValueError(f"{path}: expected a YAML mapping")
+    return data
+
+
+def _as_list(value: Any) -> list[Any]:
+    if isinstance(value, list):
+        return value
+    return [value]
+
+
+def _enabled(config: dict[str, Any], framework: str) -> bool:
+    return bool(config.get("frameworks", {}).get(framework, {}).get("enabled", False))
+
+
+def _max_required_sequence(dataset: dict[str, Any]) -> int:
+    input_len = dataset.get("input_len")
+    output_len = dataset.get("output_len")
+    if not isinstance(input_len, list) or not isinstance(output_len, list):
+        raise ValueError("dataset.input_len and dataset.output_len must be lists")
+    if len(input_len) != len(output_len):
+        raise ValueError("dataset.input_len and dataset.output_len must be aligned")
+    if not input_len:
+        raise ValueError("dataset.input_len and dataset.output_len must not be empty")
+    return max(int(i) + int(o) for i, o in zip(input_len, output_len, strict=True))
+
+
+def _candidate_dicts(
+    base_flags: dict[str, Any],
+    search_space: dict[str, Any],
+    limit: int,
+) -> list[dict[str, Any]]:
+    candidates = [dict(base_flags)]
+    keys = list(search_space)
+    values = [_as_list(search_space[key]) for key in keys]
+    for combo in itertools.product(*values):
+        candidate = dict(base_flags)
+        candidate.update(dict(zip(keys, combo, strict=True)))
+        if candidate not in candidates:
+            candidates.append(candidate)
+        if len(candidates) >= limit:
+            break
+    return candidates
+
+
+def _command_tokens(
+    framework: str,
+    config: dict[str, Any],
+    flags: dict[str, Any],
+) -> list[str]:
+    server = config["frameworks"][framework]
+    command = shlex.split(server["server_command"])
+    model = config["model"]["name"]
+
+    if framework in {"vllm", "tensorrt_llm"}:
+        command.append(model)
+
+    for key, value in flags.items():
+        if value is None or value is False:
+            continue
+        command.append(flag_name(framework, key))
+        if value is not True:
+            command.append(str(value))
+
+    return command
+
+
+def render_command(
+    framework: str, config: dict[str, Any], flags: dict[str, Any]
+) -> str:
+    return shlex.join(_command_tokens(framework, config, flags))
+
+
+def _extract_help_flags(text: str) -> set[str]:
+    return {
+        item.lstrip("-") for item in re.findall(r"--[A-Za-z0-9][A-Za-z0-9_-]*", text)
+    }
+
+
+def load_help_flags(help_dir: Path) -> dict[str, set[str]]:
+    help_flags: dict[str, set[str]] = {}
+    for framework, hints in HELP_FILE_HINTS.items():
+        matches = []
+        for path in help_dir.rglob("*.txt"):
+            name = path.name.lower()
+            if all(hint in name for hint in hints):
+                matches.append(path)
+        if matches:
+            text = "\n".join(
+                path.read_text(encoding="utf-8", errors="replace") for path in matches
+            )
+            help_flags[framework] = _extract_help_flags(text)
+    return help_flags
+
+
+def _known_flag(
+    framework: str,
+    key: str,
+    help_flags: dict[str, set[str]] | None,
+) -> bool:
+    static_keys = STATIC_SERVER_FLAGS[framework]
+    if key not in static_keys:
+        return False
+    if not help_flags or framework not in help_flags:
+        return True
+
+    concrete = flag_name(framework, key).lstrip("-")
+    aliases = {concrete, concrete.replace("-", "_"), concrete.replace("_", "-")}
+    return bool(aliases & help_flags[framework])
+
+
+def _validate_framework(
+    config: dict[str, Any],
+    framework: str,
+    help_flags: dict[str, set[str]] | None,
+    max_candidates: int,
+) -> list[str]:
+    errors: list[str] = []
+    server = config["frameworks"].get(framework)
+    if not isinstance(server, dict):
+        return [f"missing frameworks.{framework}"]
+    if not server.get("enabled", False):
+        return []
+
+    base_flags = server.get("base_server_flags")
+    search_space = server.get("search_space")
+    if not isinstance(base_flags, dict):
+        errors.append(f"{framework}: base_server_flags must be a mapping")
+        base_flags = {}
+    if not isinstance(search_space, dict):
+        errors.append(f"{framework}: search_space must be a mapping")
+        search_space = {}
+    server_command_is_valid = isinstance(server.get("server_command"), str)
+    if not server_command_is_valid:
+        errors.append(f"{framework}: server_command must be a string")
+
+    for key in set(base_flags) | set(search_space):
+        if not _known_flag(framework, key, help_flags):
+            errors.append(f"{framework}: unknown or unsupported server flag {key!r}")
+
+    if framework == "tensorrt_llm":
+        if server.get("backend_policy") != "fixed_pytorch":
+            errors.append("tensorrt_llm: backend_policy must be fixed_pytorch")
+        if base_flags.get("backend") != "pytorch":
+            errors.append("tensorrt_llm: base backend must be pytorch")
+        if "backend" in search_space:
+            errors.append("tensorrt_llm: backend must not appear in search_space")
+
+    candidates = _candidate_dicts(base_flags, search_space, max_candidates)
+    if not candidates:
+        errors.append(f"{framework}: no candidates generated")
+    can_render = server_command_is_valid and isinstance(
+        config.get("model", {}).get("name"), str
+    )
+    if can_render:
+        for candidate in candidates:
+            command = render_command(framework, config, candidate)
+            if not command:
+                errors.append(f"{framework}: rendered an empty command")
+
+    return errors
+
+
+def validate_config(
+    path: Path,
+    help_flags: dict[str, set[str]] | None = None,
+) -> list[str]:
+    errors: list[str] = []
+    try:
+        config = load_yaml(path)
+    except Exception as exc:  # noqa: BLE001
+        return [str(exc)]
+
+    if config.get("schema_version") != 1:
+        errors.append("schema_version must be 1")
+    if not isinstance(config.get("model", {}).get("name"), str):
+        errors.append("model.name must be set")
+    if config.get("source", {}).get("kind") not in ALLOWED_SOURCE_KINDS:
+        errors.append(f"source.kind must be one of {sorted(ALLOWED_SOURCE_KINDS)}")
+
+    try:
+        required_sequence = _max_required_sequence(config["dataset"])
+    except Exception as exc:  # noqa: BLE001
+        errors.append(str(exc))
+        required_sequence = 0
+
+    search = config.get("search")
+    if not isinstance(search, dict):
+        errors.append("search must be a mapping")
+        max_candidates = 1
+    else:
+        try:
+            max_candidates = int(search.get("max_candidates_per_framework", 0))
+        except (TypeError, ValueError):
+            errors.append("search.max_candidates_per_framework must be an integer")
+            max_candidates = 1
+        if max_candidates < 1:
+            errors.append("search.max_candidates_per_framework must be positive")
+            max_candidates = 1
+
+    frameworks = config.get("frameworks")
+    if not isinstance(frameworks, dict):
+        return errors + ["frameworks must be a mapping"]
+
+    for framework in FRAMEWORKS:
+        errors.extend(
+            _validate_framework(config, framework, help_flags, max_candidates)
+        )
+
+    for framework in FRAMEWORKS:
+        if not _enabled(config, framework):
+            continue
+        key = SEQUENCE_LIMIT_KEY[framework]
+        fw = frameworks[framework]
+        base_flags = fw.get("base_server_flags", {}) or {}
+        search_space = fw.get("search_space", {}) or {}
+        if not isinstance(base_flags, dict) or not isinstance(search_space, dict):
+            continue
+
+        try:
+            if framework == "sglang":
+                base_value = int(base_flags.get(key, required_sequence))
+            else:
+                base_value = int(base_flags.get(key, 0))
+        except (TypeError, ValueError):
+            errors.append(f"{framework}: base {key} is not an integer")
+            continue
+        if base_value < required_sequence:
+            errors.append(
+                f"{framework}: base {key} ({base_value}) is smaller than the largest dataset scenario ({required_sequence})"
+            )
+
+        if key in search_space:
+            for value in _as_list(search_space[key]):
+                try:
+                    if int(value) < required_sequence:
+                        errors.append(
+                            f"{framework}: search_space {key} candidate {value} is smaller than the largest dataset scenario ({required_sequence})"
+                        )
+                except (TypeError, ValueError):
+                    errors.append(
+                        f"{framework}: search_space {key} candidate {value!r} is not an integer"
+                    )
+
+    sla_block = (
+        config.get("benchmark", {}).get("sla")
+        if isinstance(config.get("benchmark"), dict)
+        else None
+    )
+    if sla_block is None:
+        sla_block = config.get("sla")
+    if isinstance(sla_block, dict):
+        for key in sla_block:
+            if key in DEPRECATED_SLA_KEYS:
+                errors.append(
+                    f"sla: {key!r} is deprecated; use {DEPRECATED_SLA_KEYS[key]!r} (see references/result-schema.md)"
+                )
+            elif key not in ALLOWED_SLA_KEYS:
+                errors.append(
+                    f"sla: unknown key {key!r}; allowed keys are {sorted(ALLOWED_SLA_KEYS)}"
+                )
+
+    return errors
+
+
+def iter_config_files(paths: list[Path]) -> list[Path]:
+    files: list[Path] = []
+    for path in paths:
+        if path.is_dir():
+            files.extend(sorted(path.rglob("*.yaml")))
+            files.extend(sorted(path.rglob("*.yml")))
+        else:
+            files.append(path)
+    return sorted(dict.fromkeys(files))
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("paths", nargs="+", type=Path)
+    parser.add_argument("--help-dir", type=Path)
+    parser.add_argument("--print-commands", action="store_true")
+    args = parser.parse_args()
+
+    help_flags = load_help_flags(args.help_dir) if args.help_dir else None
+    failed = False
+    for path in iter_config_files(args.paths):
+        errors = validate_config(path, help_flags)
+        if errors:
+            failed = True
+            for error in errors:
+                print(f"{path}: {error}")
+            continue
+
+        if args.print_commands:
+            config = load_yaml(path)
+            limit = int(config["search"].get("max_candidates_per_framework", 1))
+            for framework in FRAMEWORKS:
+                if not _enabled(config, framework):
+                    continue
+                server = config["frameworks"][framework]
+                candidates = _candidate_dicts(
+                    server["base_server_flags"],
+                    server["search_space"],
+                    limit,
+                )
+                print(f"# {path.name} {framework}")
+                print(render_command(framework, config, candidates[0]))
+
+    if failed:
+        raise SystemExit(1)
+
+
+if __name__ == "__main__":
+    main()

.claude/skills/llm-torch-profiler-analysis/SKILL.md

@@ -55,18 +55,39 @@ add one short note after the tables with exactly one of:
 | Existing trace triage | yes | yes | yes |
 | Single-trace live capture | yes | yes, if torch profiler is enabled on server | requires profiler control endpoints |
 | Two-trace mapping+formal triage | yes | yes | yes |
-| Stage-aware live capture | yes | no | no |
+| Stage-separated live workload | yes | yes | yes, with a writable shared trace dir or per-stage host runner |
+| `--profile-by-stage` capture | yes | no | no |
 | `--profile-prefix` control | yes | usually ignored on HTTP profiler route | usually ignored on HTTP profiler route |
 
 For TensorRT-LLM, live capture only works when the server exposes `/start_profile` and
 `/stop_profile`, and when the deployment already provides a shared trace path plus the
 required env vars.
 
-## Validation Notes
+## Real H100 Validation
 
-This unified workflow has been validated with a `4x H100` matrix across SGLang,
+The current reference run is the `4x H100` matrix captured on `2026-04-23` on
-vLLM, and TensorRT-LLM. Use these model shapes as representative coverage when
+`h100_sglang` under:
-refreshing or extending the skill:
+
+- `/data/bbuf/validate/unified_llm_profiler_skill/runs/20260423_h100_large_model_matrix_v3`
+
+Rendered markdown bundle:
+
+- `/data/bbuf/validate/unified_llm_profiler_skill/runs/20260423_h100_large_model_matrix_v3/h100_large_model_matrix_v3_bundle.md`
+
+Validated model directories:
+
+- `mixtral_8x7b_instruct`
+- `qwen2_5_32b_instruct`
+- `qwen3_32b`
+
+Each model directory contains:
+
+- `analysis_sglang.txt`
+- `analysis_vllm.txt`
+- `analysis_trtllm.txt`
+- framework-specific trace roots and probe artifacts
+
+Validated matrix:
 
 | Model | SGLang | vLLM | TensorRT-LLM | Result |
 | --- | --- | --- | --- | --- |
@@ -74,17 +95,49 @@ refreshing or extending the skill:
 | `Qwen/Qwen2.5-32B-Instruct` | `4x H100` | `4x H100` | `4x H100` | three tables rendered correctly on all three frameworks; benchmark probes returned direct, non-empty text |
 | `Qwen/Qwen3-32B` | `4x H100` | `4x H100` | `4x H100` | three tables rendered correctly on all three frameworks; vLLM and TensorRT-LLM chat probes often emitted `<think>` prefixes |
 
+Use this run as the main H100 reference.
+The older `2026-04-22` single-card Qwen3 matrix is still useful for bring-up, but it is
+not the default reference anymore.
+
+Stage-separated workload validation captured on `2026-05-01` on `h100_sglang`:
+
+- `/data/bbuf/validate/unified_llm_profiler_skill/runs/20260501_stage_split_validation`
+- `/data/bbuf/validate/unified_llm_profiler_skill/runs/20260501_stage_split_validation_large`
+
+Validated models:
+
+| Model | GPU | Workloads | Result |
+| --- | --- | --- | --- |
+| `Qwen/Qwen2.5-0.5B-Instruct` | `1x H100` | prefill `4090->1`, decode `1->2048` | generated separate `prefill/*.trace.json.gz` and `decode/*.trace.json.gz`; kernel, overlap, and fuse tables rendered with separate `extend/prefill` and `decode` sections |
+| `Qwen/Qwen2.5-1.5B-Instruct` | `1x H100` | prefill `4090->1`, decode `1->2048` | generated separate `prefill/*.trace.json.gz` and `decode/*.trace.json.gz`; kernel, overlap, and fuse tables rendered with separate `extend/prefill` and `decode` sections |
+| `Qwen/Qwen2.5-7B-Instruct` | `1x H100` | prefill `4090->1`, decode `1->2048` | generated separate traces; prefill kernel table captured 28-layer GEMM/FA3/RMSNorm work, decode captured 5-step graph launches, and fuse rows were split by stage |
+| `Qwen/Qwen2.5-14B-Instruct` | `1x H100` | prefill `4090->1`, decode `1->2048` | generated separate traces; prefill kernel table captured 48-layer GEMM/FA3/RMSNorm work, decode captured 5-step graph launches, and fuse rows were split by stage |
+| `Qwen/Qwen3-8B` | `2x H100`, TP=2 | prefill `4090->1`, decode `1->2048`, warmup 10/capture 5 | generated separate prefill/decode traces and all three tables; unique probe prompts avoided prefix-cache pollution in the prefill table |
+| `mistralai/Mistral-7B-Instruct-v0.3` | `2x H100`, TP=2 | prefill `4090->1`, decode `1->2048`, warmup 10/capture 5 | generated separate prefill/decode traces and all three tables; server logs showed no repeated-prompt prefix-cache shortcut during the active prefill window |
+
+This validation also covers the compatibility fix for older SGLang profiler
+state machines: workload-separated live capture labels stages by output
+directory and avoids nesting SGLang's internal `profile_by_stage` state machine
+inside each workload. The helper
+adds one internal scheduler guard step because SGLang increments `forward_ct`
+before checking whether the profiler should stop; without that guard, a
+`num_steps=1` prefill capture can stop just before the actual prefill forward.
+The 2026-05-01 two-card validation artifacts for the additional models are:
+
+- `/data/bbuf/validate/core_skill_validation_20260501/qwen3_8b/profiler`
+- `/data/bbuf/validate/core_skill_validation_20260501/mistral_7b_instruct_v03/profiler`
+
 To render a validated run into one markdown document:
 
 ```bash
 python3 scripts/render_triage_markdown_bundle.py \
-  --analysis-root /path/to/analysis_root \
+  --analysis-root /data/bbuf/validate/unified_llm_profiler_skill/runs/20260423_h100_large_model_matrix_v3 \
-  --output /path/to/analysis_bundle.md
+  --output /data/bbuf/validate/unified_llm_profiler_skill/runs/20260423_h100_large_model_matrix_v3/h100_large_model_matrix_v3_bundle.md
 ```
 
 The bundle groups by model and keeps the three tables for each framework.
 
-Validation notes:
+H100 notes:
 
 - all three frameworks now render kernel, overlap, and fuse tables with separate `extend/prefill` and `decode` sections when the trace contains a clean stage split
 - SGLang live capture is validated and calls the server profiler API directly instead of shelling out to `sglang.profiler`
@@ -92,7 +145,8 @@ Validation notes:
 - SGLang kernel-site reconstruction keeps sampling disabled in the mapping path so the optimized parser does not perturb SGLang table output; equality rechecks matched for `Mixtral-8x7B-Instruct-v0.1`, `Qwen3-32B`, and `nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8`
 - vLLM live capture requires `--output-dir` to match the server `torch_profiler_dir`; the validated H100 flow uses `--profiler-config {"profiler":"torch","torch_profiler_dir":"..."}` and then drives `/start_profile` and `/stop_profile`
 - TensorRT-LLM validation stays on `--backend pytorch`; the H100 flow writes the trace with `TLLM_TORCH_PROFILE_TRACE` and then analyzes the saved trace
-- the 2026-04-22 TensorRT-LLM 1.0.0 `py_executor.py` profiler setup still needed a `with_stack=True` override for table-quality Python locations; re-check this on TensorRT-LLM 1.2.1 or any 1.3.x release-candidate image before assuming the override is still required
+- the 2026-04-22 TensorRT-LLM 1.0.0 `py_executor.py` profiler setup still needed a `with_stack=True` override for table-quality Python locations, and the matrix runner generated that override under `/data/bbuf/validate/unified_llm_profiler_skill/overrides/trtllm`; re-check this on TensorRT-LLM 1.2.1 or any 1.3.x release-candidate image before assuming the override is still required
+- on this host, keep all trace roots under `/data/...`, not `/home/...`
 
 ## When To Use It
 
@@ -119,6 +173,43 @@ Handle it as a backend-selection issue, not as native-kernel profiler evidence.
 
 ## Main Flows
 
+## Stage-Separated Live Capture Contract
+
+Live capture must not use one mixed prompt as the default.
+By default, `analyze_llm_torch_profile.py --url ...` captures two labeled
+workloads and then renders the same three tables with separate stage sections:
+
+- prefill: synthetic input length `4090`, output length `1`
+- decode: synthetic input length `1`, output length `2048`
+
+Every live profiler path warms up `10` steps before arming the profiler and then
+captures `5` active steps by default. Keep this warmup/active split aligned
+across SGLang, vLLM, and TensorRT-LLM before comparing kernel tables.
+
+Use these options to override the contract when the benchmark workload is known:
+
+```bash
+--profile-workload both \
+--warmup-steps 10 --num-steps 5 \
+--prefill-input-len 4090 --prefill-output-len 1 \
+--decode-input-len 1 --decode-output-len 2048
+```
+
+Allowed `--profile-workload` values:
+
+- `both`: default; capture prefill and decode separately
+- `prefill`: capture only the long-input / one-token workload
+- `decode`: capture only the one-input / long-output workload
+- `legacy`: keep the old `--probe-prompt` / `--probe-max-new-tokens` behavior
+
+For `sglang-sota-performance`, do not use the defaults if the slow SGLang
+benchmark scenario has a known input/output distribution.
+Set the profiler lengths from that slow scenario instead: prefill uses the slow
+input length with output `1`, and decode uses input `1` with the slow output
+length. For a mixed dataset, profile the slowest representative bucket such as
+the p50 or p95 input/output pair used in the benchmark report, and record the
+bucket in the artifact notes.
+
 ### 1. Single-trace triage from an existing profile dir or trace
 
 ```bash
@@ -136,14 +227,24 @@ mapping/formal pair is needed.
 python3 scripts/analyze_llm_torch_profile.py \
   --framework sglang \
   --url http://127.0.0.1:30000 \
-  --output-dir /tmp/llm-profiler/sglang_profile_live \
+  --output-dir /data/bbuf/validate/unified_llm_profiler_skill/runs/example/sglang_profile_live \
   --num-steps 5 \
-  --profile-by-stage
+  --warmup-steps 10 \
+  --profile-by-stage \
+  --profile-workload both
 ```
 
 The script sends `POST /start_profile` to the SGLang server directly.
-The script writes `server_args.json`, sends the probe requests after profiling is armed,
+Keep `--output-dir` under `/data/...` so later analysis and docs can see the trace.
-and waits longer for trace flush than the earlier implementation.
+The script writes `server_args.json`, warms up with the same workload shape,
+sends the active probe requests after profiling is armed, captures separate
+`prefill/` and `decode/` profile roots by default, and waits longer for trace
+flush than the earlier implementation.
+For the default workload-separated capture, the directory name labels the stage
+and the SGLang internal `profile_by_stage` mode is not used inside each
+workload. This avoids mixing a one-token prefill probe with a separate decode
+profile. The helper still adds one internal guard step because older SGLang
+profilers check the target counter before running the next forward.
 
 ### 3. Single-trace live capture from vLLM
 
@@ -151,7 +252,7 @@ Launch vLLM with torch profiler enabled, for example:
 
 ```bash
 vllm serve meta-llama/Llama-3.1-8B-Instruct \
-  --profiler-config '{"profiler":"torch","torch_profiler_dir":"/tmp/llm-profiler/vllm_profile"}'
+  --profiler-config '{"profiler":"torch","torch_profiler_dir":"/data/bbuf/validate/unified_llm_profiler_skill/runs/example/vllm_profile"}'
 ```
 
 Then run:
@@ -160,14 +261,20 @@ Then run:
 python3 scripts/analyze_llm_torch_profile.py \
   --framework vllm \
   --url http://127.0.0.1:8000 \
-  --output-dir /tmp/llm-profiler/vllm_profile \
+  --output-dir /data/bbuf/validate/unified_llm_profiler_skill/runs/example/vllm_profile \
   --num-steps 5 \
-  --no-profile-by-stage
+  --warmup-steps 10 \
+  --no-profile-by-stage \
+  --profile-workload both
 ```
 
 For vLLM, `--output-dir` must point to the same `torch_profiler_dir` the server uses.
 The current vLLM profiler config already defaults `torch_profiler_with_stack=true`,
 so the runner only needs to set `torch_profiler_dir`.
+On `h100_sglang`, external vLLM containers should mount both:
+
+- `/data/.cache/huggingface:/root/.cache/huggingface`
+- `/data/bbuf/validate/unified_llm_profiler_skill:/data/bbuf/validate/unified_llm_profiler_skill`
 
 ### 4. Single-trace live capture from TensorRT-LLM
 
@@ -187,11 +294,15 @@ python3 scripts/analyze_llm_torch_profile.py \
   --url http://127.0.0.1:8000 \
   --output-dir /shared/path \
   --num-steps 5 \
-  --no-profile-by-stage
+  --no-profile-by-stage \
+  --profile-workload both
 ```
 
 If the deployment does not expose the profiler control endpoints, fall back to analyzing
 an existing trace instead of trying live capture.
+If the TensorRT-LLM trace output is configured as one fixed file path, use
+`scripts/run_trtllm_pytorch_profile_host.sh --stage prefill` and `--stage decode`
+instead of direct `--profile-workload both`, so each stage gets its own trace file.
 
 On the current TensorRT-LLM mainline path, `py_executor.py` creates the torch profiler
 with `record_shapes=True` and `with_modules=True` but not `with_stack=True`.
@@ -200,19 +311,21 @@ For table-quality validation, use the override generator:
 ```bash
 python3 scripts/make_trtllm_py_executor_override.py \
   --source /path/to/original/py_executor.py \
-  --output /tmp/llm-profiler/py_executor_with_stack.py
+  --output /data/bbuf/validate/unified_llm_profiler_skill/overrides/trtllm/py_executor_with_stack.py
 ```
 
-The validated TensorRT-LLM flow is:
+The matrix runner does this automatically on H100 before TensorRT-LLM capture starts.
 
-1. launch `trtllm-serve` with `TLLM_TORCH_PROFILE_TRACE=/shared/path/trace.json`
+This is the validated TensorRT-LLM flow on `h100_sglang`:
+
+1. launch `trtllm-serve` with `TLLM_TORCH_PROFILE_TRACE=/data/.../trace.json`
 2. run a few benchmark requests
-3. analyze the emitted trace with `--input /shared/path/trace.json`
+3. analyze the emitted trace with `--input /data/.../trace.json`
 
 ### 5. Two-trace triage from existing profile dirs or traces
 
 ```bash
-python3 scripts/analyze_llm_torch_profile.py triage \
+python3 scripts/analyze_llm_torch_profile.py \
   --mapping-input /path/to/graph_off_profile_dir \
   --formal-input /path/to/graph_on_profile_dir
 ```
@@ -222,7 +335,7 @@ Use this when you need stronger overlap attribution and kernel-to-source mapping
 ### 6. Two-trace triage from running servers
 
 ```bash
-python3 scripts/analyze_llm_torch_profile.py triage \
+python3 scripts/analyze_llm_torch_profile.py \
   --framework sglang \
   --mapping-url http://127.0.0.1:31025 \
   --formal-url http://127.0.0.1:31026 \
@@ -241,7 +354,13 @@ For `vllm` or `TensorRT-LLM`, use the same shape but pass:
 
 `--profile-by-stage` is only meaningful on the SGLang live-capture path.
 
-- On ordinary non-PD SGLang serving, it is still useful because prefill and decode usually have very different bottlenecks.
+- With `--profile-workload both` / `prefill` / `decode`, workload directories
+  are the stage labels; the live-capture helper disables SGLang's internal
+  stage profiler per workload, warms up first, and captures the requested
+  active step count for the selected workload.
+- On legacy or hand-captured SGLang serving, internal `profile_by_stage` is
+  still useful because prefill and decode usually have very different
+  bottlenecks.
 - On the current profile-v2 path inside SGLang, stage-based profiling is effectively the normal path.
 - PD-disaggregated serving adds one extra rule: prefill workers and decode workers must be profiled separately. That is stricter than ordinary `profile_by_stage`.
 - For `vllm` and `TensorRT-LLM`, disable it with `--no-profile-by-stage`.
@@ -279,8 +398,10 @@ It exists to recover `kernel -> cpu_op -> python scope`.
 1. If the user only wants a diagnosis, one trace is enough.
 2. Prefer one-rank traces over merged traces whenever the profiler emitted both.
 3. For a live server, let the script drive the profiler only when the framework-specific prerequisites are already met.
-4. Prefer SGLang `--profile-by-stage` unless the user explicitly wants an all-stage mixed trace.
+4. Prefer `--profile-workload both`; use `legacy` only when reproducing an old trace contract.
-5. Create or clean the target trace directory before live capture so the profiler can write artifacts without permission surprises.
+5. Prefer workload-separated SGLang capture; use internal `--profile-by-stage`
+   mainly for `legacy` or manually collected traces.
+6. When on `h100_sglang`, create or clean the target trace directory through `docker exec sglang_bbuf ...` so the path is definitely writable under `/data`.
 
 ### Two-trace workflow
 
@@ -312,6 +433,8 @@ Load these only when needed:
   - overlap labels, dependency-risk interpretation, and limits
 - [references/fuse-overlap-catalog.md](references/fuse-overlap-catalog.md)
   - mixed source-backed catalog of existing fuse and overlap patterns, including mainline rows plus PR-backed / in-flight rows
+- [references/vllm-torch-compile-fusions.md](references/vllm-torch-compile-fusions.md)
+  - current vLLM torch.compile fusion passes and the source patterns they target
 - [references/overlap-catalog.md](references/overlap-catalog.md)
   - overlap-only lookup table across LLM, VLM, diffusion, disaggregation, HiSparse, and speculative scheduling
 

.claude/skills/llm-torch-profiler-analysis/references/fuse-overlap-catalog.md

@@ -28,21 +28,14 @@ overlap opportunity as novel.
 
 The catalog is grouped by reusable optimization family, not by one specific model.
 
-Refresh note `2026-04-22`: rescanned current `sglang`, `flashinfer`,
+Refresh note `2026-05-01`: rescanned current `sglang` and vLLM mainline, then
-`TensorRT-LLM`, and `vllm` mainline plus rechecked referenced PR state via the
+rechecked recent merged and open optimization PRs through the GitHub CLI/API.
-GitHub API on `2026-04-22`. Stable current-code families such as Qwen-style
+The vLLM torch.compile pass inventory is now split out in
-shared-expert top-k append, TensorRT-LLM Triton fused add+RMSNorm+FP8 quant,
+[`vllm-torch-compile-fusions.md`](vllm-torch-compile-fusions.md). Stable
-and vLLM `merge_attn_states` attention-output quant are folded into the
+current-code families remain folded into the mainline rows below. New
-mainline rows below. Closed-unmerged SGLang
+status-sensitive rows were added for DeepSeek-V4, GLM5 NSA / PDL, NVFP4 MoE,
-[#22410](https://github.com/sgl-project/sglang/pull/22410) and FlashInfer
+torch.compile decode, vLLM DSV4, vLLM ROCm WMMA, and vLLM GPU/CPU sync-removal
-[#2840](https://github.com/flashinfer-ai/flashinfer/pull/2840) were removed
+work. Recheck PR state before treating an in-flight row as shipped.
-from the PR-backed sections. Keep FlashInfer
-[#3058](https://github.com/flashinfer-ai/flashinfer/pull/3058) /
-[#3079](https://github.com/flashinfer-ai/flashinfer/pull/3079) in mind because
-that branch was reverted, and keep vLLM
-[#40057](https://github.com/vllm-project/vllm/pull/40057) in mind when using
-B200 FP4 MoE test coverage as a signal: it disables some B200 FP4 MoE layer
-tests rather than proving the kernel family is absent.
 
 ## 1. LLM / SRT fused-kernel families
 
@@ -147,12 +140,21 @@ Stable entries should be folded into the mainline family rows above.
 | PR `#22005` fused add + RMSNorm + per-token FP8 quant | `fused_add_rmsnorm_per_token_quant`<br>`per_token_quant_fp8` | `PR #22005`<br>`python/sglang/jit_kernel/csrc/elementwise/fused_add_rmsnorm_per_token_quant.cuh`<br>`python/sglang/jit_kernel/fused_add_rmsnorm_per_token_quant.py` | CUDA JIT kernel keeps normed values in registers and emits BF16 + FP8 outputs plus per-token scales | If FP8 online-quant traces show add+norm followed by per-token quant, treat this as an in-flight upstream CUDA fuse family. |
 | PR `#20667` Qwen3.5 fused QK norm + RoPE + KV cache write | `fused_qk_norm_rope_cache_pts_quant_shuffle`<br>`fused_qk_norm_mrope_3d_cache_pts_quant_shuffle`<br>`rotary_dim` | `PR #20667`<br>`python/sglang/srt/models/qwen3_5.py`<br>`python/sglang/srt/models/utils.py` | ROCm / AITER path fuses Q / K RMSNorm, partial or 3D RoPE, and direct KV cache write for Qwen3.5 attention | Treat split QK-norm + RoPE + cache-store on Qwen3.5 as a concrete in-flight upstream family, not a novel idea. |
 | PR `#22392` CUTLASS FP8 GEMM replacing nvjet | `cutlass_scaled_mm`<br>`fp8_scaled_mm`<br>`nvjet`<br>`cudaMemsetAsync` | `PR #22392`<br>`sgl-kernel/python/sgl_kernel/gemm.py`<br>`python/sglang/srt/layers/quantization/fp8_utils.py` | Runtime replacement swaps nvjet FP8 GEMMs for CUTLASS kernels, removing per-launch memset bubbles and extra output-copy kernels | Treat nvjet GEMM + memset bubble ladders as an in-flight SGLang linear-kernel family before calling them novel. |
+| PR `#18612` NVFP4 CUTLASS MoE fused SiLU+Mul+quant | `silu_and_mul_scaled_nvfp4`<br>`nvfp4 expert quant`<br>`cutlass moe` | `PR #18612`<br>`python/sglang/srt/layers/moe/cutlass_w4a8_moe.py`<br>`python/sglang/jit_kernel/nvfp4.py` | Fuses MoE activation epilogue and NVFP4 expert quantization before the CUTLASS MoE second GEMM | Treat split SiLU+Mul then NVFP4 expert quant in CUTLASS MoE traces as an in-flight upstream SGLang family. |
+| PR `#22918` FlashInfer per-token NVFP4 MoE | `per_token_nvfp4`<br>`trtllm_fp4_block_scale_moe`<br>`FlashInfer MoE` | `PR #22918`<br>`python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py` | Adds FlashInfer-backed per-token NVFP4 MoE execution so expert quant/dequant work can move into the fused MoE backend | Treat standalone per-token NVFP4 MoE support kernels as a candidate missing backend-selection path, not an automatically novel kernel idea. |
+| PR `#22851` NSA top-k backend and FlashInfer / PyTorch top-k split | `nsa topk`<br>`flashinfer_topk`<br>`pytorch_topk`<br>`fast_topk_transform` | `PR #22851`<br>`python/sglang/srt/layers/attention/nsa_backend.py` | Makes NSA top-k backend selection explicit and aligns fused top-k transform with FlashInfer / PyTorch fallbacks | When NSA top-k dominates decode, first classify it as backend selection or fused-transform eligibility work. |
+| PR `#24125` GLM5 NSA decode CatArrayBatchedCopy removal | `CatArrayBatchedCopy`<br>`GLM-5`<br>`NSA`<br>`TileLang decode` | `PR #24125`<br>`python/sglang/srt/layers/attention/nsa_backend.py` | Skips redundant cat/copy work in the GLM5 NSA TileLang decode path | Treat cat/copy bursts in GLM5 NSA decode as a concrete in-flight cleanup opportunity. |
+| PR `#24007` MoE LoRA virtual experts for csgmv backend | `csgmv`<br>`virtual experts`<br>`MoE LoRA`<br>`fused_moe_lora` | `PR #24007`<br>`python/sglang/srt/layers/lora_backend.py`<br>`python/sglang/srt/layers/moe` | Routes MoE LoRA adapter work through virtual experts so csgmv-style kernels can batch it instead of launching fragmented adapter work | Treat MoE-LoRA tiny-kernel ladders as an in-flight batching/fusion family. |
+| PR `#24150` torch.compile local decode support | `enable_torch_compile`<br>`local compile`<br>`decode compile`<br>`torchinductor` | `PR #24150`<br>`python/sglang/srt` | Extends SGLang torch.compile coverage to local decode regions, so Inductor-generated fusion may replace hand-authored tiny kernels | When decode traces show compiler-generated kernels or missing named fused kernels, check this in-flight compile path before calling the shape unsupported. |
 
 ## 7. PR-backed / in-flight kernel-overlap families
 
 | Pattern | Trace keywords | Primary code | Existing path | Skill should conclude |
 | --- | --- | --- | --- | --- |
 | PR `#21877` fused down-GEMM + combine superseding SBO | `enable_fused_grouped_gemm_combine`<br>`combine`<br>`down_gemm` | `PR #21877`<br>`python/sglang/srt/server_args.py`<br>`python/sglang/srt/layers/moe/token_dispatcher/deepep.py` | Fused combine eliminates the standalone combine window, so SBO is intentionally disabled when this path is on | If the trace discussion is about combine overlap, first classify it as this upstream fused-overlap family. |
+| PR `#23965` PDL for DSV32 / GLM5 kernels | `enable_pdl`<br>`TRTLLM_ENABLE_PDL`<br>`cudaGridDependencySynchronize`<br>`DSV32`<br>`GLM5` | `PR #23965`<br>`python/sglang/srt/layers`<br>`sgl-kernel` | Enables programmatic dependent launch on selected DeepSeek / GLM kernels so dependent decode kernels can overlap launch-to-start gaps | Treat tight same-stream decode windows around DSV32 / GLM5 as an in-flight PDL overlap family. |
+| PR `#21878` TTFT / TPOT torch.compile optimization | `enable_torch_compile`<br>`decode graph`<br>`piecewise cudagraph` | `PR #21878`<br>`python/sglang/srt` | Uses compiler and graph capture changes to shave TTFT / TPOT rather than adding one handwritten kernel | If the trace shows many small compiler-visible decode ops, compare against this compile-overlap / graph-capture family first. |
+| PR `#24168` batched GPU-to-CPU sync for logprobs / embeddings | `logprobs`<br>`embeddings`<br>`GPU->CPU sync`<br>`batch sync` | `PR #24168`<br>`python/sglang/srt` | Batches per-request synchronization work that can otherwise serialize decode progress around logprob or embedding outputs | Treat per-request CPU sync stalls in logprob / embedding traces as a concrete in-flight SGLang scheduler/data-movement family. |
 
 ## 8. FlashInfer mainline fused-kernel families
 
@@ -272,6 +274,15 @@ contain the same implementation.
 | PR `#37646` ROCm AITER fused allreduce + RMSNorm | `rocm_aiter_fused_allreduce_rmsnorm`<br>`custom_fused_ar_rms`<br>`RocmAiterAllReduceFusionPass` | `PR #37646`<br>`vllm/_aiter_ops.py`<br>`vllm/compilation/passes/pass_manager.py` | ROCm-specific compile-time path swaps the generic all-reduce fusion pass for an AITER fused allreduce-plus-RMSNorm kernel family | Treat ROCm TP all-reduce + RMSNorm ladders as an in-flight upstream fused-collective family first. |
 | PR `#36413` FlashInfer RMSNorm + FP4 quant fusion | `fuse_norm_quant`<br>`flashinfer`<br>`NVFP4`<br>`rmsnorm + fp4 quant` | `PR #36413`<br>`vllm/compilation/passes/fusion/rms_quant_fusion.py`<br>`vllm/docs/design/fusions.md` | FlashInfer-backed norm-plus-FP4 quant fusion extends the existing RMSNorm+quant family to NVFP4 flows | Treat split RMSNorm + FP4 quant ladders as an upstream in-flight family, not a fresh idea. |
 | PR `#39301` GLM5 router GEMM with PDL overlap | `TRTLLM_ENABLE_PDL`<br>`router_gemm`<br>`GLM5`<br>`FI AR RMS fusion` | `PR #39301`<br>`vllm/model_executor/layers/fused_moe/router/gate_linear.py`<br>`vllm/csrc/moe/dsv3_router_gemm_utils.h` | Extends the specialized router GEMM family to GLM5 hidden size and uses PDL to overlap the router launch with the preceding fused allreduce-plus-RMS block | Treat this as an in-flight upstream router-kernel plus launch-overlap family before calling it novel. |
+| PR `#41455` ROCm WMMA paged prefill and split-K decode | `wmma`<br>`paged prefill`<br>`split-K decode`<br>`ROCm attention` | `PR #41455`<br>`vllm/v1/attention`<br>`vllm/_aiter_ops.py` | Adds ROCm WMMA attention kernels for paged prefill and split-K decode shapes | Treat split attention support kernels on AMD as an in-flight vLLM attention-kernel family before calling them novel. |
+| PR `#41263` DeepSeek-V4 fused norm / router low-latency path | `DSV4`<br>`fuse norm router`<br>`low latency`<br>`router` | `PR #41263`<br>`vllm/model_executor/models/deepseek_v2.py`<br>`vllm/model_executor/layers/fused_moe/router` | Targets DeepSeek-V4 decode latency by fusing norm / router-adjacent work and low-latency model paths | Treat DSV4 norm-router ladders as a concrete in-flight upstream family. |
+| PR `#41428` DSV4 fused indexer Q quant kernel | `DSV4`<br>`fused Indexer Q quant`<br>`indexer q`<br>`fp4` | `PR #41428`<br>`vllm/model_executor/models/deepseek_v2.py`<br>`vllm/csrc` | Improves the fused DeepSeek-V4 indexer Q quant kernel instead of materializing Q then quantizing separately | Treat DSV4 indexer-Q quant ladders as an in-flight upstream fused quant family. |
+| PR `#41255` DeepSeek-V4 Tile kernels / `head_compute_mix_kernel` | `head_compute_mix_kernel`<br>`Tile kernel`<br>`DSV4`<br>`MLA` | `PR #41255`<br>`vllm/model_executor/models/deepseek_v2.py`<br>`vllm/csrc` | Adds DeepSeek-V4 Tile kernels that mix head compute work in one specialized kernel | Treat DSV4 MLA head-compute ladders as a known in-flight specialized-kernel family. |
+| PR `#41441` DSV4 all-reduce plus `mhc_post` fusion | `DSV4`<br>`AR+mhc_post`<br>`allreduce`<br>`mhc_post` | `PR #41441`<br>`vllm/model_executor/models/deepseek_v2.py`<br>`vllm/compilation/passes/fusion` | Fuses or overlaps DSV4 all-reduce with post-MLA head-compute work | Treat all-reduce followed by `mhc_post` in DSV4 traces as an in-flight vLLM overlap/fusion family. |
+| PR `#41446` AMD GatedDeltaNet FLA prefill kernels | `GatedDeltaNet`<br>`FLA prefill`<br>`AMD`<br>`Qwen3-Next` | `PR #41446`<br>`vllm/model_executor/models/qwen3_next.py`<br>`vllm/v1/attention` | Optimizes GatedDeltaNet / FLA prefill kernels on AMD linear-attention models | Treat split GDN prefill kernels on ROCm as an in-flight upstream family. |
+| PR `#39748` dual-stream GDN input projection | `dual-stream`<br>`input projection`<br>`GatedDeltaNet`<br>`Qwen3.5` | `PR #39748`<br>`vllm/model_executor/models/qwen3_next.py` | Overlaps sibling input-projection branches for Qwen3 / Qwen3.5 GDN-style blocks | Treat serial GDN input projections as a known in-flight overlap opportunity. |
+| PRs `#41433` / `#41434` / `#41429` / `#40561` GPU/CPU sync removal | `GPU->CPU sync`<br>`cpu sync`<br>`item()`<br>`non_blocking` | `PR #41433`<br>`PR #41434`<br>`PR #41429`<br>`PR #40561` | Removes or gates accidental GPU-to-CPU synchronization points and adds sync-detection coverage | Treat CPU gaps next to small GPU kernels as an upstream vLLM sync-removal family before proposing a kernel-only fix. |
+| PR `#36823` vLLM IR `fused_add_rms_norm` overload | `vllm_ir`<br>`fused_add_rms_norm`<br>`maybe_inplace` | `PR #36823`<br>`vllm/compilation/passes/ir`<br>`vllm/compilation/passes/fusion/rms_quant_fusion.py` | Extends vLLM IR lowering so fused-add-RMSNorm variants remain visible to later compile-time fusions | Treat missing norm/quant compile fusion as potentially an IR-lowering visibility issue. |
 
 ## 17. Important toggles and caveats
 
@@ -306,8 +317,11 @@ contain the same implementation.
 | `PassConfig.fuse_norm_quant` | `vllm/config/compilation.py` | Enables vLLM's RMSNorm(+residual add) -> FP8 / FP4 quant compile-time fusion family. |
 | `PassConfig.fuse_act_quant` | `vllm/config/compilation.py` | Enables vLLM's `SiLU+Mul -> quant` fusion family, plus ROCm AITER variants where applicable. |
 | `PassConfig.fuse_attn_quant` | `vllm/config/compilation.py` | Enables attention-epilogue quant fusion; requires the right backend / graph visibility, so split kernels may still be expected. |
+| `PassConfig.fuse_mla_dual_rms_norm` | `vllm/config/compilation.py` | Enables the AITER-backed MLA paired-Q/KV RMSNorm fusion family on ROCm. |
 | `PassConfig.enable_qk_norm_rope_fusion` | `vllm/config/compilation.py` | Enables the compile-time QK RMSNorm + RoPE family on CUDA-like backends. |
 | `PassConfig.fuse_rope_kvcache` | `vllm/config/compilation.py` | Enables ROCm / AITER RoPE + KV-cache update fusion and is range-limited by token count. |
+| `PassConfig.fuse_minimax_qk_norm` | `vllm/config/compilation.py` | Enables the MiniMax decode Q/K allreduce-plus-RMSNorm compile-time fusion family. |
+| `PassConfig.fuse_act_padding` | `vllm/config/compilation.py` | Enables the ROCm AITER add-RMSNorm-plus-pad fusion family when AITER is available. |
 | `PassConfig.enable_sp` | `vllm/config/compilation.py` | Rewrites all-reduce into sequence-parallel staging; this is often a prerequisite for the overlap family, not just a pure fuse toggle. |
 | `PassConfig.fuse_gemm_comms` | `vllm/config/compilation.py` | Enables AsyncTP GEMM + collective overlap and auto-enables `enable_sp` when valid. |
 | `TRTLLM_ENABLE_PDL` | `vllm/csrc/dsv3_fused_a_gemm.cu`<br>`vllm/csrc/moe/dsv3_router_gemm_utils.h` | Enables programmatic dependent launch for the DSV3 specialized CUDA kernels, which can change launch grouping and trace shape for router / QKV-A paths. |

.claude/skills/llm-torch-profiler-analysis/references/vllm-torch-compile-fusions.md

@@ -0,0 +1,66 @@
+# vLLM Torch Compile Fusion Patterns
+
+Refresh: `2026-05-01`.
+Source tree: vLLM `origin/main` at `7075df79b`.
+
+Use this file when the fuse-pattern table reports split kernels in a trace and
+you need to decide whether the shape is already covered by vLLM's
+`torch.compile` pattern matcher. Treat every row here as an upstream precedent
+before calling a similar SGLang opportunity novel.
+
+## Pass Registration
+
+vLLM registers these passes from
+`vllm/compilation/passes/pass_manager.py` through `PassConfig`.
+
+| Toggle | Pass | Target shape |
+| --- | --- | --- |
+| `enable_sp` | `SequenceParallelismPass` | all-reduce around residual/norm blocks becomes reduce-scatter, local work, and all-gather |
+| `fuse_gemm_comms` | `AsyncTPPass` | GEMM plus reduce-scatter / all-gather overlap through symmetric-memory collectives |
+| `fuse_allreduce_rms` | `AllReduceFusionPass` | all-reduce followed by RMSNorm, optional residual add, optional FP8 / NVFP4 quant |
+| `fuse_minimax_qk_norm` | `MiniMaxQKNormPass` | MiniMax Q/K all-reduce plus RMSNorm decode path |
+| `fuse_norm_quant` | `RMSNormQuantFusionPass` | RMSNorm or fused-add-RMSNorm followed by FP8 / FP4 quant |
+| `fuse_norm_quant` + AITER | `RocmAiterRMSNormQuantFusionPass` | ROCm AITER RMSNorm / fused-add-RMSNorm followed by AITER or vLLM quant |
+| `fuse_act_quant` | `ActivationQuantFusionPass` | SiLU-and-mul followed by FP8 / NVFP4 / block quant |
+| `fuse_act_quant` + AITER | `RocmAiterSiluMulFp8GroupQuantFusionPass` | AITER SiLU-and-mul followed by FP8 group quant |
+| `fuse_act_padding` + AITER | `RocmAiterTritonAddRMSNormPadFusionPass` | AITER fused-add-RMSNorm followed by padding into the next layout |
+| `fuse_mla_dual_rms_norm` + AITER | `MLADualRMSNormFusionPass` | MLA paired Q and KV RMSNorms become `fused_mla_dual_rms_norm` |
+| `fuse_rope_kvcache` | `RopeKVCacheFusionPass` | RoPE plus paged KV-cache update, after split cleanup passes |
+| `fuse_attn_quant` | `AttnQuantFusionPass` | attention output followed by FP8 / NVFP4 quant |
+| `fuse_attn_quant` | `MLAAttnQuantFusionPass` | MLA attention output followed by FP8 / NVFP4 / FP8 group quant |
+| `enable_qk_norm_rope_fusion` | `QKNormRoPEFusionPass` | Q/K RMSNorm plus RoPE on packed QKV tensors |
+
+## Pattern Inventory
+
+| Source file | Pattern classes | Trace clue | Replacement |
+| --- | --- | --- | --- |
+| `fusion/allreduce_rms_fusion.py` | `AllReduceRMSNormPattern`, `AllReduceFusedAddRMSNormPattern`, `AllReduceFusedRMSNormStaticQuantFP8Pattern`, `AllReduceFusedAddRMSNormStaticQuantFP8Pattern`, `AllReduceFusedRMSNormStaticQuantNVFP4Pattern`, `AllReduceFusedAddRMSNormStaticQuantNVFP4Pattern` | TP all-reduce directly before RMSNorm, residual-add RMSNorm, or quant | `flashinfer_trtllm_fused_allreduce_norm` with FlashInfer allreduce fusion pattern codes |
+| `fusion/rms_quant_fusion.py` | `RMSNormStaticQuantPattern`, `FusedAddRMSNormStaticQuantPattern`, `RMSNormDynamicQuantPattern`, `FusedAddRMSNormDynamicQuantPattern`, `RMSNormGroupQuantPattern`, `FusedAddRMSNormGroupQuantPattern` | RMSNorm or fused-add-RMSNorm followed by static FP8, dynamic per-token FP8, FP8 group quant, or NVFP4 quant | `_C.rms_norm_*_quant`, `_C.fused_add_rms_norm_*_quant`, or per-block quant custom op |
+| `fusion/rocm_aiter_fusion.py` | `AiterRMSNormDynamicQuantPattern`, `AiterFusedAddRMSNormDynamicQuantPattern`, `AiterRMSFp8GroupQuantPattern`, `AiterFusedAddRMSFp8GroupQuantPattern` | AITER RMSNorm/fused-add-RMSNorm followed by AITER or vLLM FP8 quant | AITER fused RMSNorm-quant custom ops |
+| `fusion/act_quant_fusion.py` | `SiluMulFp8StaticQuantPattern`, `SiluMulNvfp4QuantPattern`, `SiluMulBlockQuantPattern` | SiLU-and-mul activation output immediately quantized | fused activation-plus-quant custom op |
+| `fusion/rocm_aiter_fusion.py` | `AiterSiluMulFp8GroupQuantPattern` | AITER SiLU-and-mul followed by FP8 group quant | AITER `act_mul_fused_fp8_group_quant` |
+| `fusion/rocm_aiter_fusion.py` | `AddAiterRMSNormPadPattern` | AITER fused-add-RMSNorm output padded before the next op | AITER add-RMSNorm-pad op |
+| `fusion/rocm_aiter_fusion.py` | `MLADualRMSNormPattern` | MLA Q branch and KV branch each run RMSNorm | `torch.ops.vllm.fused_mla_dual_rms_norm` backed by AITER fused QK RMSNorm |
+| `fusion/qk_norm_rope_fusion.py` | `QkNormRopePattern` | Q/K RMSNorm, split/getitem reshapes, then RoPE | `_C.fused_qk_norm_rope` |
+| `fusion/rope_kvcache_fusion.py` | `RopeReshapeKVCachePattern` | RoPE output followed by reshape/cache update | `vllm.fused_rope_and_unified_kv_cache_update` |
+| `fusion/attn_quant_fusion.py` | `AttnFp8StaticQuantPattern`, `AttnNvfp4QuantPattern` | attention output followed by FP8 static quant or NVFP4 quant | backend attention op with fused output quant when supported |
+| `fusion/mla_attn_quant_fusion.py` | `MLAAttnFp8StaticQuantPattern`, `MLAAttnNvfp4QuantPattern`, `MLAAttnFp8GroupQuantPattern` | MLA attention output followed by static FP8, NVFP4, or FP8 group quant | MLA attention op with fused output quant when supported |
+| `fusion/minimax_qk_norm_fusion.py` | `MiniMaxQKNormPattern` | MiniMax `forward_qk`: Q/K variance all-reduce divided by TP world size, then RMS apply | `vllm.minimax_qk_norm_fused` / Lamport fused kernel |
+| `fusion/sequence_parallelism.py` | `FirstAllReduceRMSNormPattern`, `MiddleAllReduceRMSNormPattern`, `FirstAllReduceRMSNormStaticFP8Pattern`, `MiddleAllReduceRMSNormStaticFP8Pattern` | all-reduce plus norm block in a full-graph TP model | sequence-parallel reduce-scatter, local norm, all-gather staging |
+| `fusion/collective_fusion.py` | `GEMMReduceScatterPattern`, `AllGatherGEMMPattern`, `ScaledMMReduceScatterPattern`, `AllGatherScaledMMPattern`, `CutlassScaledMMReduceScatterPattern`, `AllGatherCutlassScaledMMPattern`, `FlashInferBMMFP8ReduceScatterPattern`, `FlashInferAllGatherBMMFP8Pattern` | matmul / scaled-mm / FlashInfer BMM adjacent to TP collectives | symmetric-memory fused matmul+reduce-scatter or all-gather+matmul |
+
+## Triage Rules
+
+- If the trace shows split norm/add/quant, compare first against
+  `RMSNormQuantFusionPass`, AITER variants, and `AllReduceFusionPass`.
+- If the trace shows attention output followed by quant kernels, compare against
+  `AttnQuantFusionPass` or `MLAAttnQuantFusionPass`, not only handwritten
+  attention kernels.
+- If the trace shows Q/K norm followed by RoPE or cache update, compare both
+  `QKNormRoPEFusionPass` and `RopeKVCacheFusionPass`; they are separate passes.
+- If the trace is a TP decode trace with visible collectives, check whether
+  `enable_sp` and `fuse_gemm_comms` would transform the same region into
+  sequence-parallel or AsyncTP overlap.
+- A missing vLLM compile fusion may be intentional when the graph range, backend
+  support check, dtype, token count, or AITER / FlashInfer availability does not
+  satisfy the pass-specific guard.

.claude/skills/llm-torch-profiler-analysis/scripts/analyze_llm_torch_profile.py

@@ -11,6 +11,12 @@ from typing import Dict, List, Optional, Sequence, Tuple
 import triage_kernel_helpers as kernel_helpers
 import triage_overlap_helpers as overlap_helpers
 from profile_common import (
+    DEFAULT_DECODE_INPUT_LEN,
+    DEFAULT_DECODE_OUTPUT_LEN,
+    DEFAULT_PREFILL_INPUT_LEN,
+    DEFAULT_PREFILL_OUTPUT_LEN,
+    DEFAULT_WARMUP_STEPS,
+    PROFILE_WORKLOAD_CHOICES,
     discover_trace_targets,
     framework_display_name,
     load_server_args,
@@ -57,8 +63,8 @@ def build_triage_parser() -> argparse.ArgumentParser:
         default=None,
         help=(
             "Running server URL for single-trace triage. SGLang supports direct "
-            "capture through its profiler HTTP API. vLLM and TensorRT-LLM require "
+            "capture via sglang.profiler. vLLM and TensorRT-LLM require a server-side "
-            "a server-side torch-profiler output path exposed via --output-dir."
+            "torch-profiler output path exposed via --output-dir."
         ),
     )
     parser.add_argument(
@@ -132,7 +138,13 @@ def build_triage_parser() -> argparse.ArgumentParser:
         "--num-steps",
         type=int,
         default=5,
-        help="Profiler steps when generating traces from URLs.",
+        help="Active profiler steps when generating traces from URLs.",
+    )
+    parser.add_argument(
+        "--warmup-steps",
+        type=int,
+        default=DEFAULT_WARMUP_STEPS,
+        help="Warmup steps to run before arming the profiler for URL capture.",
     )
     parser.add_argument(
         "--profile-by-stage", action=argparse.BooleanOptionalAction, default=True
@@ -151,11 +163,45 @@ def build_triage_parser() -> argparse.ArgumentParser:
     )
     parser.add_argument("--probe-max-new-tokens", type=int, default=None)
     parser.add_argument("--probe-delay", type=float, default=0.5)
+    parser.add_argument(
+        "--profile-workload",
+        choices=PROFILE_WORKLOAD_CHOICES,
+        default="both",
+        help=(
+            "Live-capture workload shape. Default 'both' captures separate "
+            "prefill and decode profiles instead of one mixed request. Use "
+            "'legacy' to keep the old --probe-prompt behavior."
+        ),
+    )
+    parser.add_argument(
+        "--prefill-input-len",
+        type=int,
+        default=DEFAULT_PREFILL_INPUT_LEN,
+        help="Synthetic input length for the prefill profile workload.",
+    )
+    parser.add_argument(
+        "--prefill-output-len",
+        type=int,
+        default=DEFAULT_PREFILL_OUTPUT_LEN,
+        help="Output length for the prefill profile workload.",
+    )
+    parser.add_argument(
+        "--decode-input-len",
+        type=int,
+        default=DEFAULT_DECODE_INPUT_LEN,
+        help="Synthetic input length for the decode profile workload.",
+    )
+    parser.add_argument(
+        "--decode-output-len",
+        type=int,
+        default=DEFAULT_DECODE_OUTPUT_LEN,
+        help="Output length for the decode profile workload.",
+    )
     parser.add_argument(
         "--start-step",
         type=int,
         default=None,
-        help="SGLang-only profiler start step when generating traces from URLs.",
+        help="Pass through to sglang.profiler when generating traces from URLs.",
     )
     parser.add_argument(
         "--pid-substring",
@@ -239,9 +285,15 @@ def resolve_profile_targets(
             probe_prompt=args.probe_prompt,
             probe_max_new_tokens=args.probe_max_new_tokens,
             probe_delay=args.probe_delay,
+            warmup_steps=args.warmup_steps,
             start_step=args.start_step,
             framework=framework,
             framework_hint_path=output_dir,
+            profile_workload=args.profile_workload,
+            prefill_input_len=args.prefill_input_len,
+            prefill_output_len=args.prefill_output_len,
+            decode_input_len=args.decode_input_len,
+            decode_output_len=args.decode_output_len,
         )
         traces, server_args = discover_trace_targets(target_dir, all_traces=False)
         resolved_framework = resolve_framework(

.claude/skills/llm-torch-profiler-analysis/scripts/probe_llm_server.py

@@ -15,7 +15,7 @@ from urllib import request
 from profile_common import extract_openai_chat_text
 
 DEFAULT_PROMPTS = [
-    "用一句中文介绍上海。",
+    "Introduce Shanghai in one short sentence.",
     "What is 2+2? Answer briefly.",
     "Write one short haiku about GPUs.",
 ]

.claude/skills/llm-torch-profiler-analysis/scripts/profile_common.py

@@ -5,10 +5,12 @@ from __future__ import annotations
 import gzip
 import json
 import re
+import shutil
 import sys
 import tempfile
 import time
 from collections import Counter, defaultdict
+from dataclasses import dataclass
 from functools import lru_cache
 from pathlib import Path
 from typing import Callable, Dict, Iterable, List, Optional, Sequence, Tuple
@@ -42,6 +44,21 @@ TRACE_METADATA_NAMES = {
 }
 NON_KERNEL_TRACE_CATEGORIES = ("python_function", "cpu_op", "trace")
 PYTHON_SCOPE_NAME_PREFIXES = ("python/", "nn.module:")
+PROFILE_WORKLOAD_CHOICES = ("legacy", "prefill", "decode", "both")
+DEFAULT_PREFILL_INPUT_LEN = 4090
+DEFAULT_PREFILL_OUTPUT_LEN = 1
+DEFAULT_DECODE_INPUT_LEN = 1
+DEFAULT_DECODE_OUTPUT_LEN = 2048
+DEFAULT_WARMUP_STEPS = 10
+
+
+@dataclass(frozen=True)
+class ProbePlan:
+    prompt: str
+    capture_max_new_tokens: int
+    capture_requests: int
+    warmup_max_new_tokens: int
+    warmup_requests: int
 
 
 @lru_cache(maxsize=65536)
@@ -416,10 +433,16 @@ def resolve_framework(
 
 
 def parse_stage(path: Path) -> str:
-    name = path.name.lower()
+    parts = [part.lower() for part in path.parts[-6:]]
-    if "-extend" in name or "-prefill" in name:
+    name = " ".join(parts)
+    segment_path = "/" + "/".join(parts) + "/"
+    if any(marker in name for marker in ("-extend", "-prefill", "_extend", "_prefill")):
         return "extend"
-    if "-decode" in name:
+    if any(f"/{segment}/" in segment_path for segment in ("extend", "prefill")):
+        return "extend"
+    if any(marker in name for marker in ("-decode", "_decode")):
+        return "decode"
+    if "/decode/" in segment_path:
         return "decode"
     return "all"
 
@@ -514,8 +537,19 @@ def discover_trace_targets(
     if path.is_file():
         return [path], load_server_args(path)
 
-    trace_dir = newest_trace_dir(path)
+    direct_traces = discover_trace_files(path, recursive=False)
-    traces = discover_trace_files(trace_dir, recursive=False)
+    recursive_traces = discover_trace_files(path, recursive=True)
+    recursive_stages = {parse_stage(trace) for trace in recursive_traces}
+    if (
+        not direct_traces
+        and recursive_traces
+        and any(stage != "all" for stage in recursive_stages)
+    ):
+        traces = recursive_traces
+        trace_dir = path
+    else:
+        trace_dir = newest_trace_dir(path)
+        traces = discover_trace_files(trace_dir, recursive=False)
     if not traces:
         raise FileNotFoundError(f"No trace files found under {trace_dir}")
 
@@ -606,6 +640,109 @@ def send_probe_request(
         )
 
 
+def unique_probe_prompt(prompt: str, probe_index: int) -> str:
+    marker = f"profile_probe_{max(0, int(probe_index))}"
+    parts = prompt.split(maxsplit=1)
+    suffix = parts[1] if len(parts) == 2 else prompt
+    return f"{marker} {suffix}".strip()
+
+
+def send_probe_requests(
+    *,
+    url: str,
+    prompt: str,
+    max_new_tokens: int,
+    request_count: int,
+    framework: str,
+    model: Optional[str] = None,
+    sampling_seed_offset: int = 0,
+) -> None:
+    request_count = max(0, int(request_count))
+    seed_offset = max(0, int(sampling_seed_offset))
+    for request_idx in range(request_count):
+        probe_index = seed_offset + request_idx
+        send_probe_request(
+            url=url,
+            prompt=unique_probe_prompt(prompt, probe_index),
+            max_new_tokens=max_new_tokens,
+            sampling_seed=probe_index,
+            framework=framework,
+            model=model,
+        )
+
+
+def synthetic_prompt(input_len: int) -> str:
+    token_count = max(1, int(input_len))
+    return " ".join(["profile"] * token_count)
+
+
+def workload_probe(
+    stage: str,
+    *,
+    prefill_input_len: int,
+    prefill_output_len: int,
+    decode_input_len: int,
+    decode_output_len: int,
+) -> Tuple[str, int]:
+    if stage == "prefill":
+        return synthetic_prompt(prefill_input_len), max(1, int(prefill_output_len))
+    if stage == "decode":
+        return synthetic_prompt(decode_input_len), max(1, int(decode_output_len))
+    raise ValueError(f"unknown profile workload stage: {stage}")
+
+
+def build_probe_plan(
+    stage: str,
+    *,
+    prompt: str,
+    max_new_tokens: int,
+    num_steps: int,
+    probe_requests: int,
+    warmup_steps: int,
+) -> ProbePlan:
+    active_steps = max(1, int(num_steps))
+    requested_probes = max(1, int(probe_requests))
+    warmup_steps = max(0, int(warmup_steps))
+    max_new_tokens = max(1, int(max_new_tokens))
+
+    if stage == "prefill":
+        return ProbePlan(
+            prompt=prompt,
+            capture_max_new_tokens=max_new_tokens,
+            capture_requests=max(requested_probes, active_steps),
+            warmup_max_new_tokens=max_new_tokens,
+            warmup_requests=warmup_steps,
+        )
+    if stage == "decode":
+        return ProbePlan(
+            prompt=prompt,
+            capture_max_new_tokens=max_new_tokens,
+            capture_requests=requested_probes,
+            warmup_max_new_tokens=max(1, warmup_steps),
+            warmup_requests=1 if warmup_steps else 0,
+        )
+    return ProbePlan(
+        prompt=prompt,
+        capture_max_new_tokens=max_new_tokens,
+        capture_requests=requested_probes,
+        warmup_max_new_tokens=max_new_tokens,
+        warmup_requests=warmup_steps,
+    )
+
+
+def expand_profile_workload(profile_workload: str) -> List[str]:
+    workload = normalize_text(profile_workload).lower()
+    if workload not in PROFILE_WORKLOAD_CHOICES:
+        raise ValueError(
+            f"--profile-workload must be one of {', '.join(PROFILE_WORKLOAD_CHOICES)}"
+        )
+    if workload == "both":
+        return ["prefill", "decode"]
+    if workload == "legacy":
+        return ["legacy"]
+    return [workload]
+
+
 def discover_openai_model(url: str) -> str:
     payload = try_get_json(url.rstrip("/") + "/v1/models", timeout=60.0)
     if not isinstance(payload, dict):
@@ -690,35 +827,50 @@ def run_remote_profiler(
     url: str,
     output_dir: Optional[str],
     framework: str,
-    probe_requests: int,
+    probe_plan: ProbePlan,
-    probe_prompt: str,
-    probe_max_new_tokens: Optional[int],
     probe_delay: float,
-    num_steps: int,
+    stage: Optional[str] = None,
 ) -> Path:
     framework = canonicalize_framework(framework)
     output_path = ensure_remote_profiler_output_path(output_dir, framework)
+    if stage and output_path.is_file():
+        raise ValueError(
+            "--profile-workload both requires a directory output path for "
+            f"{framework_display_name(framework)} so each stage trace can be labeled."
+        )
+    before_traces = (
+        set(discover_trace_files(output_path, recursive=True))
+        if output_path.exists()
+        else set()
+    )
+    model = discover_openai_model(url) if framework in {"vllm", "trtllm"} else None
+    if probe_plan.warmup_requests > 0:
+        send_probe_requests(
+            url=url,
+            prompt=probe_plan.prompt,
+            max_new_tokens=probe_plan.warmup_max_new_tokens,
+            request_count=probe_plan.warmup_requests,
+            framework=framework,
+            model=model,
+        )
+
     start_remote_profiler(url, framework)
     stop_error: Optional[BaseException] = None
     try:
-        if probe_requests > 0:
+        if probe_plan.capture_requests > 0:
-            # Some profiler endpoints need a brief setup window after
+            # `sglang.profiler` performs its own startup work before it reaches
             # POST /start_profile. A very short delay can send probes too early
             # and miss the profiling window entirely.
             time.sleep(max(5.0, probe_delay))
-            effective_max_new_tokens = probe_max_new_tokens or max(64, num_steps * 8)
+            send_probe_requests(
-            model = (
+                url=url,
-                discover_openai_model(url) if framework in {"vllm", "trtllm"} else None
+                prompt=probe_plan.prompt,
+                max_new_tokens=probe_plan.capture_max_new_tokens,
+                request_count=probe_plan.capture_requests,
+                framework=framework,
+                model=model,
+                sampling_seed_offset=probe_plan.warmup_requests,
             )
-            for request_idx in range(probe_requests):
-                send_probe_request(
-                    url=url,
-                    prompt=probe_prompt,
-                    max_new_tokens=effective_max_new_tokens,
-                    sampling_seed=request_idx,
-                    framework=framework,
-                    model=model,
-                )
     finally:
         try:
             stop_remote_profiler(url, framework)
@@ -726,7 +878,22 @@ def run_remote_profiler(
             stop_error = exc
     if stop_error is not None:
         raise stop_error
-    return wait_for_profiler_artifact(output_path)
+    artifact = wait_for_profiler_artifact(output_path)
+    if stage and output_path.is_dir():
+        after_traces = set(discover_trace_files(output_path, recursive=True))
+        new_traces = sorted(after_traces - before_traces, key=lambda item: item.name)
+        if new_traces:
+            stage_dir = output_path / stage
+            stage_dir.mkdir(parents=True, exist_ok=True)
+            for trace in new_traces:
+                if stage_dir in trace.parents:
+                    continue
+                target = stage_dir / trace.name
+                if target.exists():
+                    target = stage_dir / f"{time.time_ns()}-{trace.name}"
+                shutil.move(str(trace), str(target))
+            return stage_dir
+    return artifact
 
 
 def run_sglang_profiler(
@@ -736,9 +903,7 @@ def run_sglang_profiler(
     profile_by_stage: bool,
     merge_profiles: bool,
     profile_prefix: Optional[str],
-    probe_requests: int,
+    probe_plan: ProbePlan,
-    probe_prompt: str,
-    probe_max_new_tokens: Optional[int],
     probe_delay: float,
     start_step: Optional[int] = None,
 ) -> Path:
@@ -765,6 +930,15 @@ def run_sglang_profiler(
     if start_step is not None:
         payload["start_step"] = str(start_step)
 
+    if probe_plan.warmup_requests > 0:
+        send_probe_requests(
+            url=url,
+            prompt=probe_plan.prompt,
+            max_new_tokens=probe_plan.warmup_max_new_tokens,
+            request_count=probe_plan.warmup_requests,
+            framework="sglang",
+        )
+
     req = request.Request(
         url.rstrip("/") + "/start_profile",
         data=json.dumps(payload).encode("utf-8"),
@@ -773,17 +947,20 @@ def run_sglang_profiler(
     with request.urlopen(req, timeout=300.0):
         pass
 
-    if probe_requests > 0:
+    if probe_plan.capture_requests > 0:
         time.sleep(max(0.0, probe_delay))
-        effective_max_new_tokens = probe_max_new_tokens or max(64, num_steps * 8)
+        send_probe_requests(
-        for request_idx in range(probe_requests):
+            url=url,
-            send_probe_request(
+            prompt=probe_plan.prompt,
-                url=url,
+            max_new_tokens=probe_plan.capture_max_new_tokens,
-                prompt=probe_prompt,
+            request_count=probe_plan.capture_requests,
-                max_new_tokens=effective_max_new_tokens,
+            framework="sglang",
-                sampling_seed=request_idx,
+            sampling_seed_offset=probe_plan.warmup_requests,
-                framework="sglang",
+        )
-            )
+        try:
+            stop_remote_profiler(url, "sglang")
+        except RuntimeError:
+            pass
 
     return wait_for_profiler_artifact(output_path, timeout_s=180.0)
 
@@ -799,9 +976,15 @@ def run_profiler(
     probe_prompt: str,
     probe_max_new_tokens: Optional[int],
     probe_delay: float,
+    warmup_steps: int = DEFAULT_WARMUP_STEPS,
     start_step: Optional[int] = None,
     framework: str = "auto",
     framework_hint_path: Optional[str] = None,
+    profile_workload: str = "both",
+    prefill_input_len: int = DEFAULT_PREFILL_INPUT_LEN,
+    prefill_output_len: int = DEFAULT_PREFILL_OUTPUT_LEN,
+    decode_input_len: int = DEFAULT_DECODE_INPUT_LEN,
+    decode_output_len: int = DEFAULT_DECODE_OUTPUT_LEN,
 ) -> Path:
     resolved_framework = resolve_framework(
         framework,
@@ -813,6 +996,58 @@ def run_profiler(
         ),
     )
     if resolved_framework == "sglang":
+        stages = expand_profile_workload(profile_workload)
+        if stages != ["legacy"]:
+            output_root = (
+                Path(output_dir).expanduser().resolve()
+                if output_dir
+                else Path(tempfile.mkdtemp(prefix="sglang-torch-profile-"))
+            )
+            output_root.mkdir(parents=True, exist_ok=True)
+            for stage in stages:
+                prompt, max_new_tokens = workload_probe(
+                    stage,
+                    prefill_input_len=prefill_input_len,
+                    prefill_output_len=prefill_output_len,
+                    decode_input_len=decode_input_len,
+                    decode_output_len=decode_output_len,
+                )
+                probe_plan = build_probe_plan(
+                    stage,
+                    prompt=prompt,
+                    max_new_tokens=max_new_tokens,
+                    num_steps=num_steps,
+                    probe_requests=probe_requests,
+                    warmup_steps=warmup_steps,
+                )
+                # SGLang increments `forward_ct` before checking whether the
+                # profiler reached its target. Ask for one extra step so the
+                # requested stage forward is captured instead of stopping just
+                # before it runs.
+                stage_num_steps = max(1, int(num_steps)) + 1
+                run_sglang_profiler(
+                    url=url,
+                    output_dir=str(output_root / stage),
+                    num_steps=stage_num_steps,
+                    profile_by_stage=False,
+                    merge_profiles=merge_profiles,
+                    profile_prefix=(
+                        f"{profile_prefix}-{stage}" if profile_prefix else stage
+                    ),
+                    probe_plan=probe_plan,
+                    probe_delay=probe_delay,
+                    start_step=start_step,
+                )
+            return output_root
+        legacy_max_new_tokens = probe_max_new_tokens or max(64, num_steps * 8)
+        legacy_plan = build_probe_plan(
+            "legacy",
+            prompt=probe_prompt,
+            max_new_tokens=legacy_max_new_tokens,
+            num_steps=num_steps,
+            probe_requests=probe_requests,
+            warmup_steps=warmup_steps,
+        )
         return run_sglang_profiler(
             url=url,
             output_dir=output_dir,
@@ -820,9 +1055,7 @@ def run_profiler(
             profile_by_stage=profile_by_stage,
             merge_profiles=merge_profiles,
             profile_prefix=profile_prefix,
-            probe_requests=probe_requests,
+            probe_plan=legacy_plan,
-            probe_prompt=probe_prompt,
-            probe_max_new_tokens=probe_max_new_tokens,
             probe_delay=probe_delay,
             start_step=start_step,
         )
@@ -844,16 +1077,48 @@ def run_profiler(
             "--profile-prefix on the HTTP profiler control path.",
             file=sys.stderr,
         )
-    return run_remote_profiler(
+    stages = expand_profile_workload(profile_workload)
-        url=url,
+    if stages == ["legacy"]:
-        output_dir=output_dir,
+        legacy_max_new_tokens = probe_max_new_tokens or max(64, num_steps * 8)
-        framework=resolved_framework,
+        return run_remote_profiler(
-        probe_requests=probe_requests,
+            url=url,
-        probe_prompt=probe_prompt,
+            output_dir=output_dir,
-        probe_max_new_tokens=probe_max_new_tokens,
+            framework=resolved_framework,
-        probe_delay=probe_delay,
+            probe_plan=build_probe_plan(
-        num_steps=num_steps,
+                "legacy",
-    )
+                prompt=probe_prompt,
+                max_new_tokens=legacy_max_new_tokens,
+                num_steps=num_steps,
+                probe_requests=probe_requests,
+                warmup_steps=warmup_steps,
+            ),
+            probe_delay=probe_delay,
+        )
+    output_root = ensure_remote_profiler_output_path(output_dir, resolved_framework)
+    for stage in stages:
+        prompt, max_new_tokens = workload_probe(
+            stage,
+            prefill_input_len=prefill_input_len,
+            prefill_output_len=prefill_output_len,
+            decode_input_len=decode_input_len,
+            decode_output_len=decode_output_len,
+        )
+        run_remote_profiler(
+            url=url,
+            output_dir=str(output_root),
+            framework=resolved_framework,
+            probe_plan=build_probe_plan(
+                stage,
+                prompt=prompt,
+                max_new_tokens=max_new_tokens,
+                num_steps=num_steps,
+                probe_requests=probe_requests,
+                warmup_steps=warmup_steps,
+            ),
+            probe_delay=probe_delay,
+            stage=stage,
+        )
+    return output_root
 
 
 def select_heaviest_pid(

.claude/skills/llm-torch-profiler-analysis/scripts/run_llm_single_model_matrix_host.sh

@@ -0,0 +1,274 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+usage() {
+  cat <<'EOF'
+Usage:
+  run_llm_single_model_matrix_host.sh \
+    --model-id gpt_oss_20b \
+    --model openai/gpt-oss-20b \
+    --root /data/bbuf/validate/unified_llm_profiler_skill/runs/20260423_h100_large_model_matrix \
+    --gpus 2,3,4,5 \
+    --sglang-port 30098 \
+    --vllm-formal-port 31098 \
+    --vllm-mapping-port 31099 \
+    --trt-formal-prefill-port 32098 \
+    --trt-formal-decode-port 32099 \
+    --trt-mapping-prefill-port 32198 \
+    --trt-mapping-decode-port 32199
+
+This script is intended to run on the H100 host. It:
+1. captures SGLang live profiling and writes `analysis_sglang.txt`
+2. captures vLLM formal + eager mapping traces and writes `analysis_vllm.txt`
+3. captures TensorRT-LLM formal + graph-off mapping traces and writes `analysis_trtllm.txt`
+4. stores one benchmark JSON per framework under the model run directory
+
+Default profiler workloads are stage-separated:
+  prefill: input 4090, output 1
+  decode:  input 1, output 2048
+
+Environment:
+  Export `HF_TOKEN` and `HUGGINGFACE_HUB_TOKEN` before running.
+EOF
+}
+
+MODEL_ID=""
+MODEL=""
+ROOT=""
+GPUS=""
+TP_SIZE=""
+SGLANG_PORT=""
+VLLM_FORMAL_PORT=""
+VLLM_MAPPING_PORT=""
+TRT_FORMAL_PREFILL_PORT=""
+TRT_FORMAL_DECODE_PORT=""
+TRT_MAPPING_PREFILL_PORT=""
+TRT_MAPPING_DECODE_PORT=""
+SGLANG_MEM_FRACTION="0.85"
+MAX_MODEL_LEN="4096"
+KV_FRACTION="0.85"
+SGLANG_SERVER_EXTRA=""
+PROFILE_WORKLOAD="both"
+PREFILL_INPUT_LEN=4090
+PREFILL_OUTPUT_LEN=1
+DECODE_INPUT_LEN=1
+DECODE_OUTPUT_LEN=2048
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+TRT_IMAGE="nvcr.io/nvidia/tensorrt-llm/release:latest"
+TRT_OVERRIDE_ROOT="/data/bbuf/validate/unified_llm_profiler_skill/overrides/trtllm"
+TRT_OVERRIDE_SOURCE="$TRT_OVERRIDE_ROOT/py_executor.original.py"
+TRT_OVERRIDE_PATH="$TRT_OVERRIDE_ROOT/py_executor_with_stack.py"
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --model-id) MODEL_ID="$2"; shift 2 ;;
+    --model) MODEL="$2"; shift 2 ;;
+    --root) ROOT="$2"; shift 2 ;;
+    --gpus) GPUS="$2"; shift 2 ;;
+    --tp-size) TP_SIZE="$2"; shift 2 ;;
+    --sglang-port) SGLANG_PORT="$2"; shift 2 ;;
+    --vllm-formal-port) VLLM_FORMAL_PORT="$2"; shift 2 ;;
+    --vllm-mapping-port) VLLM_MAPPING_PORT="$2"; shift 2 ;;
+    --trt-formal-prefill-port) TRT_FORMAL_PREFILL_PORT="$2"; shift 2 ;;
+    --trt-formal-decode-port) TRT_FORMAL_DECODE_PORT="$2"; shift 2 ;;
+    --trt-mapping-prefill-port) TRT_MAPPING_PREFILL_PORT="$2"; shift 2 ;;
+    --trt-mapping-decode-port) TRT_MAPPING_DECODE_PORT="$2"; shift 2 ;;
+    --sglang-mem-fraction) SGLANG_MEM_FRACTION="$2"; shift 2 ;;
+    --sglang-server-extra) SGLANG_SERVER_EXTRA="$2"; shift 2 ;;
+    --max-model-len) MAX_MODEL_LEN="$2"; shift 2 ;;
+    --kv-fraction) KV_FRACTION="$2"; shift 2 ;;
+    --profile-workload) PROFILE_WORKLOAD="$2"; shift 2 ;;
+    --prefill-input-len) PREFILL_INPUT_LEN="$2"; shift 2 ;;
+    --prefill-output-len) PREFILL_OUTPUT_LEN="$2"; shift 2 ;;
+    --decode-input-len) DECODE_INPUT_LEN="$2"; shift 2 ;;
+    --decode-output-len) DECODE_OUTPUT_LEN="$2"; shift 2 ;;
+    --help|-h) usage; exit 0 ;;
+    *)
+      echo "Unknown argument: $1" >&2
+      usage >&2
+      exit 2
+      ;;
+  esac
+done
+
+if [[ -z "${HF_TOKEN:-}" && -z "${HUGGINGFACE_HUB_TOKEN:-}" ]]; then
+  echo "Set HF_TOKEN or HUGGINGFACE_HUB_TOKEN before running." >&2
+  exit 2
+fi
+if [[ -z "${HF_TOKEN:-}" ]]; then
+  HF_TOKEN="$HUGGINGFACE_HUB_TOKEN"
+fi
+if [[ -z "${HUGGINGFACE_HUB_TOKEN:-}" ]]; then
+  HUGGINGFACE_HUB_TOKEN="$HF_TOKEN"
+fi
+
+for value in \
+  MODEL_ID MODEL ROOT GPUS \
+  SGLANG_PORT VLLM_FORMAL_PORT VLLM_MAPPING_PORT \
+  TRT_FORMAL_PREFILL_PORT TRT_FORMAL_DECODE_PORT \
+  TRT_MAPPING_PREFILL_PORT TRT_MAPPING_DECODE_PORT; do
+  if [[ -z "${!value}" ]]; then
+    echo "Missing required argument: $value" >&2
+    usage >&2
+    exit 2
+  fi
+done
+
+IFS=',' read -r -a GPU_LIST <<< "$GPUS"
+GPU_COUNT="${#GPU_LIST[@]}"
+if [[ "$GPU_COUNT" -lt 1 ]]; then
+  echo "Could not parse --gpus: $GPUS" >&2
+  exit 2
+fi
+if [[ -z "$TP_SIZE" ]]; then
+  TP_SIZE="$GPU_COUNT"
+fi
+if (( TP_SIZE < 1 || TP_SIZE > GPU_COUNT )); then
+  echo "--tp-size must be between 1 and the visible GPU count ($GPU_COUNT)." >&2
+  exit 2
+fi
+
+MODEL_ROOT="$ROOT/$MODEL_ID"
+SGLANG_ANALYSIS="$MODEL_ROOT/analysis_sglang.txt"
+VLLM_FORMAL_DIR="$MODEL_ROOT/vllm_formal"
+VLLM_MAPPING_DIR="$MODEL_ROOT/vllm_mapping"
+VLLM_ANALYSIS="$MODEL_ROOT/analysis_vllm.txt"
+TRT_FORMAL_DIR="$MODEL_ROOT/trtllm_formal"
+TRT_MAPPING_DIR="$MODEL_ROOT/trtllm_mapping"
+TRT_ANALYSIS="$MODEL_ROOT/analysis_trtllm.txt"
+
+docker exec sglang_bbuf bash -lc "mkdir -p '$MODEL_ROOT'"
+
+if [[ ! -s "$TRT_OVERRIDE_SOURCE" ]]; then
+  echo "[bootstrap] TensorRT-LLM py_executor source snapshot"
+  docker exec sglang_bbuf bash -lc "mkdir -p '$TRT_OVERRIDE_ROOT'"
+  docker run --rm --entrypoint cat "$TRT_IMAGE" \
+    /usr/local/lib/python3.12/dist-packages/tensorrt_llm/_torch/pyexecutor/py_executor.py \
+    | docker exec -i sglang_bbuf bash -lc "cat > '$TRT_OVERRIDE_SOURCE'"
+fi
+echo "[bootstrap] TensorRT-LLM py_executor override with with_stack=True and rank0-only trace export"
+docker exec sglang_bbuf bash -lc "cd '$SCRIPT_DIR' && python3 make_trtllm_py_executor_override.py --source '$TRT_OVERRIDE_SOURCE' --output '$TRT_OVERRIDE_PATH'"
+
+sglang_args=(
+  --model "$MODEL"
+  --run-dir "$MODEL_ROOT"
+  --port "$SGLANG_PORT"
+  --gpus "$GPUS"
+  --tp-size "$TP_SIZE"
+  --mem-fraction "$SGLANG_MEM_FRACTION"
+  --profile-workload "$PROFILE_WORKLOAD"
+  --prefill-input-len "$PREFILL_INPUT_LEN"
+  --prefill-output-len "$PREFILL_OUTPUT_LEN"
+  --decode-input-len "$DECODE_INPUT_LEN"
+  --decode-output-len "$DECODE_OUTPUT_LEN"
+  --trust-remote-code
+)
+if [[ -n "$SGLANG_SERVER_EXTRA" ]]; then
+  sglang_args+=(--server-extra "$SGLANG_SERVER_EXTRA")
+fi
+
+echo "[1/6] SGLang server + live triage"
+HF_TOKEN="$HF_TOKEN" HUGGINGFACE_HUB_TOKEN="$HUGGINGFACE_HUB_TOKEN" \
+  "$SCRIPT_DIR/run_sglang_torch_profile_host.sh" \
+  "${sglang_args[@]}"
+
+echo "[2/6] vLLM formal"
+HF_TOKEN="$HF_TOKEN" HUGGINGFACE_HUB_TOKEN="$HUGGINGFACE_HUB_TOKEN" \
+  "$SCRIPT_DIR/run_vllm_torch_profile_host.sh" \
+  --model "$MODEL" \
+  --run-dir "$VLLM_FORMAL_DIR" \
+  --port "$VLLM_FORMAL_PORT" \
+  --gpus "$GPUS" \
+  --tensor-parallel-size "$TP_SIZE" \
+  --max-model-len "$MAX_MODEL_LEN" \
+  --profile-workload "$PROFILE_WORKLOAD" \
+  --prefill-input-len "$PREFILL_INPUT_LEN" \
+  --prefill-output-len "$PREFILL_OUTPUT_LEN" \
+  --decode-input-len "$DECODE_INPUT_LEN" \
+  --decode-output-len "$DECODE_OUTPUT_LEN" \
+  --trust-remote-code
+
+echo "[3/6] vLLM mapping"
+HF_TOKEN="$HF_TOKEN" HUGGINGFACE_HUB_TOKEN="$HUGGINGFACE_HUB_TOKEN" \
+  "$SCRIPT_DIR/run_vllm_torch_profile_host.sh" \
+  --model "$MODEL" \
+  --run-dir "$VLLM_MAPPING_DIR" \
+  --port "$VLLM_MAPPING_PORT" \
+  --gpus "$GPUS" \
+  --tensor-parallel-size "$TP_SIZE" \
+  --profiler-active-iterations 2 \
+  --max-model-len "$MAX_MODEL_LEN" \
+  --profile-workload "$PROFILE_WORKLOAD" \
+  --prefill-input-len "$PREFILL_INPUT_LEN" \
+  --prefill-output-len "$PREFILL_OUTPUT_LEN" \
+  --decode-input-len "$DECODE_INPUT_LEN" \
+  --decode-output-len "$DECODE_OUTPUT_LEN" \
+  --trust-remote-code \
+  --enforce-eager
+
+echo "[4/6] vLLM mapping-formal analysis"
+docker exec sglang_bbuf bash -lc "cd '$SCRIPT_DIR' && python3 analyze_llm_torch_profile.py --framework vllm --mapping-input '$VLLM_MAPPING_DIR' --formal-input '$VLLM_FORMAL_DIR' > '$VLLM_ANALYSIS'"
+
+echo "[5/6] TensorRT-LLM formal + mapping captures"
+HF_TOKEN="$HF_TOKEN" HUGGINGFACE_HUB_TOKEN="$HUGGINGFACE_HUB_TOKEN" \
+  "$SCRIPT_DIR/run_trtllm_pytorch_profile_host.sh" \
+  --model "$MODEL" \
+  --run-dir "$TRT_FORMAL_DIR" \
+  --stage prefill \
+  --port "$TRT_FORMAL_PREFILL_PORT" \
+  --gpus "$GPUS" \
+  --tp-size "$TP_SIZE" \
+  --kv-fraction "$KV_FRACTION" \
+  --input-len "$PREFILL_INPUT_LEN" \
+  --output-len "$PREFILL_OUTPUT_LEN" \
+  --override-py-executor "$TRT_OVERRIDE_PATH" \
+  --trust-remote-code
+HF_TOKEN="$HF_TOKEN" HUGGINGFACE_HUB_TOKEN="$HUGGINGFACE_HUB_TOKEN" \
+  "$SCRIPT_DIR/run_trtllm_pytorch_profile_host.sh" \
+  --model "$MODEL" \
+  --run-dir "$TRT_FORMAL_DIR" \
+  --stage decode \
+  --port "$TRT_FORMAL_DECODE_PORT" \
+  --gpus "$GPUS" \
+  --tp-size "$TP_SIZE" \
+  --kv-fraction "$KV_FRACTION" \
+  --input-len "$DECODE_INPUT_LEN" \
+  --output-len "$DECODE_OUTPUT_LEN" \
+  --override-py-executor "$TRT_OVERRIDE_PATH" \
+  --trust-remote-code
+HF_TOKEN="$HF_TOKEN" HUGGINGFACE_HUB_TOKEN="$HUGGINGFACE_HUB_TOKEN" \
+  "$SCRIPT_DIR/run_trtllm_pytorch_profile_host.sh" \
+  --model "$MODEL" \
+  --run-dir "$TRT_MAPPING_DIR" \
+  --stage prefill \
+  --port "$TRT_MAPPING_PREFILL_PORT" \
+  --gpus "$GPUS" \
+  --tp-size "$TP_SIZE" \
+  --kv-fraction "$KV_FRACTION" \
+  --input-len "$PREFILL_INPUT_LEN" \
+  --output-len "$PREFILL_OUTPUT_LEN" \
+  --override-py-executor "$TRT_OVERRIDE_PATH" \
+  --disable-cudagraph \
+  --trust-remote-code
+HF_TOKEN="$HF_TOKEN" HUGGINGFACE_HUB_TOKEN="$HUGGINGFACE_HUB_TOKEN" \
+  "$SCRIPT_DIR/run_trtllm_pytorch_profile_host.sh" \
+  --model "$MODEL" \
+  --run-dir "$TRT_MAPPING_DIR" \
+  --stage decode \
+  --port "$TRT_MAPPING_DECODE_PORT" \
+  --gpus "$GPUS" \
+  --tp-size "$TP_SIZE" \
+  --kv-fraction "$KV_FRACTION" \
+  --input-len "$DECODE_INPUT_LEN" \
+  --output-len "$DECODE_OUTPUT_LEN" \
+  --override-py-executor "$TRT_OVERRIDE_PATH" \
+  --disable-cudagraph \
+  --trust-remote-code
+
+echo "[6/6] TensorRT-LLM mapping-formal analysis"
+docker exec sglang_bbuf bash -lc "cd '$SCRIPT_DIR' && python3 analyze_llm_torch_profile.py --framework trtllm --mapping-input '$TRT_MAPPING_DIR' --formal-input '$TRT_FORMAL_DIR' > '$TRT_ANALYSIS'"
+
+echo "MODEL_ROOT=$MODEL_ROOT"
+echo "ANALYSIS_SGLANG=$SGLANG_ANALYSIS"
+echo "ANALYSIS_VLLM=$VLLM_ANALYSIS"
+echo "ANALYSIS_TRTLLM=$TRT_ANALYSIS"

.claude/skills/llm-torch-profiler-analysis/scripts/run_sglang_torch_profile_host.sh

@@ -0,0 +1,241 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+usage() {
+  cat <<'EOF'
+Usage:
+  run_sglang_torch_profile_host.sh \
+    --model Qwen/Qwen3-8B \
+    --run-dir /data/bbuf/validate/unified_llm_profiler_skill/runs/example_sglang \
+    --port 30088 \
+    --gpus 0
+
+  run_sglang_torch_profile_host.sh \
+    --model openai/gpt-oss-20b \
+    --run-dir /data/bbuf/validate/unified_llm_profiler_skill/runs/example_sglang_4gpu \
+    --port 30088 \
+    --gpus 2,3,4,5 \
+    --tp-size 4
+
+Options:
+  --model TEXT                  Model id or local path for SGLang.
+  --run-dir PATH               Shared /data directory for logs and traces.
+  --port INT                   Server port.
+  --gpus TEXT                  CUDA_VISIBLE_DEVICES value, for example 0 or 2,3,4,5.
+  --gpu TEXT                   Alias for --gpus.
+  --tp-size INT                Tensor parallel size. Defaults to the visible GPU count.
+  --trust-remote-code          Pass --trust-remote-code.
+  --mem-fraction FLOAT         SGLang static memory fraction.
+  --request-max-tokens INT     Generation length for the probe request.
+  --prompt TEXT                Probe prompt.
+  --warmup-steps INT           Warmup steps before profiling. Defaults to 10.
+  --profile-workload TEXT      legacy|prefill|decode|both. Defaults to both.
+  --prefill-input-len INT      Synthetic prefill prompt length. Defaults to 4090.
+  --prefill-output-len INT     Synthetic prefill output length. Defaults to 1.
+  --decode-input-len INT       Synthetic decode prompt length. Defaults to 1.
+  --decode-output-len INT      Synthetic decode output length. Defaults to 2048.
+  --repo-dir PATH              SGLang repo path inside `sglang_bbuf`.
+  --server-extra TEXT          Extra args appended to launch_server.
+  --help                       Show this message.
+
+Notes:
+  - Run this on the H100 host. It uses `docker exec sglang_bbuf`.
+  - The server is launched first, then the profiler capture runs with
+    stage-separated prefill/decode workloads and `--profile-by-stage`.
+  - A small benchmark summary is written after profiling.
+EOF
+}
+
+MODEL=""
+RUN_DIR=""
+PORT=""
+GPUS=""
+TP_SIZE=""
+TRUST_REMOTE_CODE=0
+MEM_FRACTION=0.85
+REQUEST_MAX_TOKENS=12
+PROMPT="Explain the difference between CUDA graph mode and eager mode in two sentences."
+WARMUP_STEPS=10
+PROFILE_WORKLOAD="both"
+PREFILL_INPUT_LEN=4090
+PREFILL_OUTPUT_LEN=1
+DECODE_INPUT_LEN=1
+DECODE_OUTPUT_LEN=2048
+SGLANG_REPO_DIR="${SGLANG_REPO_DIR:-/data/bbuf/repos/sglang}"
+SERVER_EXTRA=""
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --model)
+      MODEL="$2"
+      shift 2
+      ;;
+    --run-dir)
+      RUN_DIR="$2"
+      shift 2
+      ;;
+    --port)
+      PORT="$2"
+      shift 2
+      ;;
+    --gpu)
+      GPUS="$2"
+      shift 2
+      ;;
+    --gpus)
+      GPUS="$2"
+      shift 2
+      ;;
+    --tp-size)
+      TP_SIZE="$2"
+      shift 2
+      ;;
+    --trust-remote-code)
+      TRUST_REMOTE_CODE=1
+      shift
+      ;;
+    --mem-fraction)
+      MEM_FRACTION="$2"
+      shift 2
+      ;;
+    --request-max-tokens)
+      REQUEST_MAX_TOKENS="$2"
+      shift 2
+      ;;
+    --prompt)
+      PROMPT="$2"
+      shift 2
+      ;;
+    --warmup-steps)
+      WARMUP_STEPS="$2"
+      shift 2
+      ;;
+    --profile-workload)
+      PROFILE_WORKLOAD="$2"
+      shift 2
+      ;;
+    --prefill-input-len)
+      PREFILL_INPUT_LEN="$2"
+      shift 2
+      ;;
+    --prefill-output-len)
+      PREFILL_OUTPUT_LEN="$2"
+      shift 2
+      ;;
+    --decode-input-len)
+      DECODE_INPUT_LEN="$2"
+      shift 2
+      ;;
+    --decode-output-len)
+      DECODE_OUTPUT_LEN="$2"
+      shift 2
+      ;;
+    --repo-dir)
+      SGLANG_REPO_DIR="$2"
+      shift 2
+      ;;
+    --server-extra)
+      SERVER_EXTRA="$2"
+      shift 2
+      ;;
+    --help|-h)
+      usage
+      exit 0
+      ;;
+    *)
+      echo "Unknown argument: $1" >&2
+      usage >&2
+      exit 2
+      ;;
+  esac
+done
+
+if [[ -z "$MODEL" || -z "$RUN_DIR" || -z "$PORT" || -z "$GPUS" ]]; then
+  usage >&2
+  exit 2
+fi
+
+IFS=',' read -r -a GPU_LIST <<< "$GPUS"
+GPU_COUNT="${#GPU_LIST[@]}"
+if [[ "$GPU_COUNT" -lt 1 ]]; then
+  echo "Could not parse --gpus: $GPUS" >&2
+  exit 2
+fi
+if [[ -z "$TP_SIZE" ]]; then
+  TP_SIZE="$GPU_COUNT"
+fi
+if (( TP_SIZE < 1 || TP_SIZE > GPU_COUNT )); then
+  echo "--tp-size must be between 1 and the visible GPU count ($GPU_COUNT)." >&2
+  exit 2
+fi
+
+LOG_PATH="$RUN_DIR/sglang_server.log"
+ANALYSIS_PATH="$RUN_DIR/analysis_sglang.txt"
+PROFILE_ROOT="$RUN_DIR/sglang_profile_live"
+BENCHMARK_PATH="$RUN_DIR/benchmark_sglang.json"
+PID_PATH="$RUN_DIR/sglang_server.pid"
+LAUNCH_PATTERN="[s]glang.launch_server.*--port $PORT"
+SERVER_ARGS="python3 -m sglang.launch_server --model-path \"$MODEL\" --port \"$PORT\" --tp-size \"$TP_SIZE\" --mem-fraction-static \"$MEM_FRACTION\""
+
+if [[ "$TRUST_REMOTE_CODE" -eq 1 ]]; then
+  SERVER_ARGS="$SERVER_ARGS --trust-remote-code"
+fi
+if [[ -n "$SERVER_EXTRA" ]]; then
+  SERVER_ARGS="$SERVER_ARGS $SERVER_EXTRA"
+fi
+
+docker exec sglang_bbuf bash -lc "mkdir -p '$RUN_DIR' '$PROFILE_ROOT'"
+docker exec sglang_bbuf bash -lc "pkill -f '$LAUNCH_PATTERN' >/dev/null 2>&1 || true"
+docker exec sglang_bbuf bash -lc "mkdir -p '$RUN_DIR' '$PROFILE_ROOT' && cd '$SGLANG_REPO_DIR' && rm -f '$PID_PATH' && (CUDA_VISIBLE_DEVICES=$GPUS PYTHONPATH=python nohup $SERVER_ARGS > '$LOG_PATH' 2>&1 < /dev/null & echo \$! > '$PID_PATH')"
+
+cleanup() {
+  docker exec sglang_bbuf bash -lc "pkill -f '$LAUNCH_PATTERN' >/dev/null 2>&1 || true" >/dev/null 2>&1 || true
+}
+trap cleanup EXIT
+
+ready=0
+for _ in $(seq 1 180); do
+  if curl -sf "http://127.0.0.1:${PORT}/v1/models" >/dev/null; then
+    ready=1
+    break
+  fi
+  sleep 2
+done
+if [[ "$ready" -ne 1 ]]; then
+  echo "SGLang server did not become ready on port ${PORT}. Recent logs:" >&2
+  ssh_log=$(docker exec sglang_bbuf bash -lc "tail -n 120 '$LOG_PATH'" 2>/dev/null || true)
+  printf '%s\n' "$ssh_log" >&2
+  exit 1
+fi
+
+python3 - <<PY
+import json
+import urllib.request
+
+payload = {
+    "text": ${PROMPT@Q},
+    "sampling_params": {
+        "temperature": 0.0,
+        "max_new_tokens": int(${REQUEST_MAX_TOKENS@Q}),
+    },
+    "stream": False,
+}
+req = urllib.request.Request(
+    "http://127.0.0.1:${PORT}/generate",
+    data=json.dumps(payload).encode(),
+    headers={"Content-Type": "application/json"},
+)
+with urllib.request.urlopen(req, timeout=600) as resp:
+    body = json.loads(resp.read().decode())
+text = body.get("text", "")
+print(text[:400])
+PY
+
+docker exec sglang_bbuf bash -lc "cd '$SCRIPT_DIR' && python3 analyze_llm_torch_profile.py --framework sglang --url http://127.0.0.1:${PORT} --output-dir '$PROFILE_ROOT' --num-steps 5 --warmup-steps '$WARMUP_STEPS' --probe-requests 1 --profile-by-stage --profile-workload '$PROFILE_WORKLOAD' --prefill-input-len '$PREFILL_INPUT_LEN' --prefill-output-len '$PREFILL_OUTPUT_LEN' --decode-input-len '$DECODE_INPUT_LEN' --decode-output-len '$DECODE_OUTPUT_LEN' > '$ANALYSIS_PATH'"
+python3 "$SCRIPT_DIR/probe_llm_server.py" \
+  --framework sglang \
+  --url "http://127.0.0.1:${PORT}" \
+  | docker exec -i sglang_bbuf bash -lc "cat > '$BENCHMARK_PATH'" >/dev/null
+docker exec sglang_bbuf bash -lc "sed -n '1,240p' '$ANALYSIS_PATH'"
+echo "BENCHMARK_PATH=$BENCHMARK_PATH"

.claude/skills/llm-torch-profiler-analysis/scripts/run_trtllm_pytorch_profile_host.sh

@@ -0,0 +1,407 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+usage() {
+  cat <<'EOF'
+Usage:
+  run_trtllm_pytorch_profile_host.sh \
+    --model Qwen/Qwen3-8B \
+    --run-dir /data/bbuf/validate/unified_llm_profiler_skill/runs/example \
+    --stage prefill \
+    --port 32188 \
+    --gpus 0
+
+  run_trtllm_pytorch_profile_host.sh \
+    --model openai/gpt-oss-20b \
+    --run-dir /data/bbuf/validate/unified_llm_profiler_skill/runs/example_4gpu \
+    --stage prefill \
+    --port 32188 \
+    --gpus 2,3,4,5 \
+    --tp-size 4
+
+Options:
+  --model TEXT                     Hugging Face model id.
+  --run-dir PATH                  Shared /data run directory for logs and traces.
+  --stage prefill|decode          Capture window. Prefill profiles 4090->1 by
+                                  default; decode profiles 1->2048 by default.
+  --port INT                      Host port for trtllm-serve.
+  --gpus TEXT                     CUDA_VISIBLE_DEVICES value, for example 0 or 2,3,4,5.
+  --gpu TEXT                      Alias for --gpus.
+  --tp-size INT                   Tensor parallel size. Defaults to the visible GPU count.
+  --image TEXT                    Container image.
+  --shared-root PATH              Shared validation root mounted into the container.
+  --hf-cache PATH                 Host Hugging Face cache path.
+  --override-py-executor PATH     Optional py_executor.py override path.
+  --disable-cudagraph             Generate/use a YAML override with cuda_graph_config: null.
+  --input-len INT                 Synthetic prompt length for this stage.
+                                  Defaults: prefill 4090, decode 1.
+  --request-max-tokens INT        Generation length for this stage.
+                                  Defaults: prefill 1, decode 2048.
+  --output-len INT                Alias for --request-max-tokens.
+  --prompt TEXT                   Probe prompt. Defaults to a synthetic prompt
+                                  sized by --input-len.
+  --warmup-steps INT              Warmup steps before the profiler window. Defaults to 10.
+  --active-steps INT              Active profiler steps to capture. Defaults to 5.
+  --max-seq-len INT               Serve max sequence length.
+  --kv-fraction FLOAT             KV cache free GPU memory fraction.
+  --container-name TEXT           Override container name.
+  --trust-remote-code             Pass --trust_remote_code to trtllm-serve.
+  --help                          Show this message.
+
+Environment:
+  HF_TOKEN or HUGGINGFACE_HUB_TOKEN must be set.
+
+Notes:
+  - Run this on the H100 host, not inside `sglang_bbuf`.
+  - It always pins TensorRT-LLM to `--backend pytorch`.
+  - The default image tag is floating; record the resolved TensorRT-LLM version
+    in the run manifest and pass --image for reproducible validation.
+  - Profiling uses `TLLM_PROFILE_START_STOP` and `TLLM_TORCH_PROFILE_TRACE`.
+  - For Python-location recovery, prefer a `py_executor.py` override with `with_stack=True`.
+  - A small benchmark summary is written after the trace is emitted.
+EOF
+}
+
+IMAGE="nvcr.io/nvidia/tensorrt-llm/release:latest"
+SHARED_ROOT="/data/bbuf/validate/unified_llm_profiler_skill"
+HF_CACHE="/data/.cache/huggingface"
+OVERRIDE_PY_EXECUTOR=""
+DISABLE_CUDAGRAPH=0
+REQUEST_MAX_TOKENS=""
+INPUT_LEN=""
+PROMPT=""
+WARMUP_STEPS=10
+ACTIVE_STEPS=5
+MAX_SEQ_LEN=4096
+KV_FRACTION=0.85
+CONTAINER_NAME=""
+TRUST_REMOTE_CODE=0
+TP_SIZE=""
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+MODEL=""
+RUN_DIR=""
+STAGE=""
+PORT=""
+GPUS=""
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --model)
+      MODEL="$2"
+      shift 2
+      ;;
+    --run-dir)
+      RUN_DIR="$2"
+      shift 2
+      ;;
+    --stage)
+      STAGE="$2"
+      shift 2
+      ;;
+    --port)
+      PORT="$2"
+      shift 2
+      ;;
+    --gpu)
+      GPUS="$2"
+      shift 2
+      ;;
+    --gpus)
+      GPUS="$2"
+      shift 2
+      ;;
+    --tp-size)
+      TP_SIZE="$2"
+      shift 2
+      ;;
+    --image)
+      IMAGE="$2"
+      shift 2
+      ;;
+    --shared-root)
+      SHARED_ROOT="$2"
+      shift 2
+      ;;
+    --hf-cache)
+      HF_CACHE="$2"
+      shift 2
+      ;;
+    --override-py-executor)
+      OVERRIDE_PY_EXECUTOR="$2"
+      shift 2
+      ;;
+    --disable-cudagraph)
+      DISABLE_CUDAGRAPH=1
+      shift
+      ;;
+    --input-len)
+      INPUT_LEN="$2"
+      shift 2
+      ;;
+    --request-max-tokens)
+      REQUEST_MAX_TOKENS="$2"
+      shift 2
+      ;;
+    --output-len)
+      REQUEST_MAX_TOKENS="$2"
+      shift 2
+      ;;
+    --prompt)
+      PROMPT="$2"
+      shift 2
+      ;;
+    --warmup-steps)
+      WARMUP_STEPS="$2"
+      shift 2
+      ;;
+    --active-steps)
+      ACTIVE_STEPS="$2"
+      shift 2
+      ;;
+    --max-seq-len)
+      MAX_SEQ_LEN="$2"
+      shift 2
+      ;;
+    --kv-fraction)
+      KV_FRACTION="$2"
+      shift 2
+      ;;
+    --container-name)
+      CONTAINER_NAME="$2"
+      shift 2
+      ;;
+    --trust-remote-code)
+      TRUST_REMOTE_CODE=1
+      shift
+      ;;
+    --help|-h)
+      usage
+      exit 0
+      ;;
+    *)
+      echo "Unknown argument: $1" >&2
+      usage >&2
+      exit 2
+      ;;
+  esac
+done
+
+if [[ -z "${HF_TOKEN:-}" && -z "${HUGGINGFACE_HUB_TOKEN:-}" ]]; then
+  echo "Set HF_TOKEN or HUGGINGFACE_HUB_TOKEN before running." >&2
+  exit 2
+fi
+if [[ -z "${HF_TOKEN:-}" ]]; then
+  HF_TOKEN="$HUGGINGFACE_HUB_TOKEN"
+fi
+if [[ -z "${HUGGINGFACE_HUB_TOKEN:-}" ]]; then
+  HUGGINGFACE_HUB_TOKEN="$HF_TOKEN"
+fi
+
+if [[ -z "$MODEL" || -z "$RUN_DIR" || -z "$STAGE" || -z "$PORT" || -z "$GPUS" ]]; then
+  usage >&2
+  exit 2
+fi
+
+IFS=',' read -r -a GPU_LIST <<< "$GPUS"
+GPU_COUNT="${#GPU_LIST[@]}"
+if [[ "$GPU_COUNT" -lt 1 ]]; then
+  echo "Could not parse --gpus: $GPUS" >&2
+  exit 2
+fi
+if [[ -z "$TP_SIZE" ]]; then
+  TP_SIZE="$GPU_COUNT"
+fi
+if (( TP_SIZE < 1 || TP_SIZE > GPU_COUNT )); then
+  echo "--tp-size must be between 1 and the visible GPU count ($GPU_COUNT)." >&2
+  exit 2
+fi
+
+case "$STAGE" in
+  prefill)
+    TRACE_PATH="$RUN_DIR/trace-prefill.json"
+    LOG_PATH="$RUN_DIR/server-prefill.log"
+    BENCHMARK_PATH="$RUN_DIR/benchmark-prefill.json"
+    if [[ -z "$INPUT_LEN" ]]; then
+      INPUT_LEN=4090
+    fi
+    if [[ -z "$REQUEST_MAX_TOKENS" ]]; then
+      REQUEST_MAX_TOKENS=1
+    fi
+    ;;
+  decode)
+    TRACE_PATH="$RUN_DIR/trace-decode.json"
+    LOG_PATH="$RUN_DIR/server-decode.log"
+    BENCHMARK_PATH="$RUN_DIR/benchmark-decode.json"
+    if [[ -z "$INPUT_LEN" ]]; then
+      INPUT_LEN=1
+    fi
+    if [[ -z "$REQUEST_MAX_TOKENS" ]]; then
+      REQUEST_MAX_TOKENS=2048
+    fi
+    ;;
+  *)
+    echo "--stage must be prefill or decode." >&2
+    exit 2
+    ;;
+esac
+
+if (( WARMUP_STEPS < 0 || ACTIVE_STEPS < 1 )); then
+  echo "--warmup-steps must be >= 0 and --active-steps must be >= 1." >&2
+  exit 2
+fi
+
+case "$STAGE" in
+  prefill)
+    profile_start=$((WARMUP_STEPS + 1))
+    ;;
+  decode)
+    profile_start=$((WARMUP_STEPS + 2))
+    ;;
+esac
+profile_stop=$((profile_start + ACTIVE_STEPS - 1))
+PROFILE_START_STOP="${profile_start}-${profile_stop}"
+
+if [[ -z "$CONTAINER_NAME" ]]; then
+  model_slug="${MODEL##*/}"
+  model_slug="${model_slug//\//-}"
+  model_slug="${model_slug//./-}"
+  model_slug="${model_slug//_/-}"
+  model_slug="${model_slug// /-}"
+  gpu_slug="${GPUS//,/-}"
+  CONTAINER_NAME="trtllm-${model_slug}-${STAGE}-g${gpu_slug}-p${PORT}"
+fi
+
+EXTRA_LLM_OPTIONS=""
+if [[ "$DISABLE_CUDAGRAPH" -eq 1 ]]; then
+  EXTRA_CFG_PATH="$SHARED_ROOT/tmp/trt_no_cudagraph.yaml"
+  docker exec sglang_bbuf bash -lc "mkdir -p '$(dirname "$EXTRA_CFG_PATH")' && printf 'cuda_graph_config: null\n' > '$EXTRA_CFG_PATH'"
+  EXTRA_LLM_OPTIONS="--extra_llm_api_options $EXTRA_CFG_PATH"
+fi
+
+docker exec sglang_bbuf bash -lc "mkdir -p '$RUN_DIR'"
+docker rm -f "$CONTAINER_NAME" >/dev/null 2>&1 || true
+
+docker_args=(
+  run -d --rm
+  --name "$CONTAINER_NAME"
+  --gpus all
+  --ipc=host
+  --network host
+  --entrypoint bash
+  -e "CUDA_VISIBLE_DEVICES=$GPUS"
+  -e "HF_TOKEN=$HF_TOKEN"
+  -e "HUGGINGFACE_HUB_TOKEN=$HUGGINGFACE_HUB_TOKEN"
+  -e "TLLM_PROFILE_START_STOP=$PROFILE_START_STOP"
+  -e "TLLM_LLMAPI_ENABLE_NVTX=1"
+  -e "TLLM_TORCH_PROFILE_TRACE=$TRACE_PATH"
+  -e "RUN_DIR=$RUN_DIR"
+  -e "LOG_PATH=$LOG_PATH"
+  -e "MODEL_ID=$MODEL"
+  -e "SERVE_PORT=$PORT"
+  -v "$HF_CACHE:/root/.cache/huggingface"
+  -v "$SHARED_ROOT:$SHARED_ROOT"
+)
+
+if [[ -n "$OVERRIDE_PY_EXECUTOR" ]]; then
+  docker_args+=(
+    -v "$OVERRIDE_PY_EXECUTOR:/usr/local/lib/python3.12/dist-packages/tensorrt_llm/_torch/pyexecutor/py_executor.py:ro"
+  )
+fi
+
+trust_remote_code_arg=""
+if [[ "$TRUST_REMOTE_CODE" -eq 1 ]]; then
+  trust_remote_code_arg="--trust_remote_code"
+fi
+
+container_cmd=$(
+  cat <<EOF
+mkdir -p "$RUN_DIR" && trtllm-serve serve "$MODEL" \
+  --backend pytorch \
+  --tp_size "$TP_SIZE" \
+  --gpus_per_node "$GPU_COUNT" \
+  --host 0.0.0.0 \
+  --port "$PORT" \
+  --max_seq_len "$MAX_SEQ_LEN" \
+  --kv_cache_free_gpu_memory_fraction "$KV_FRACTION" \
+  $trust_remote_code_arg \
+  $EXTRA_LLM_OPTIONS \
+  > "$LOG_PATH" 2>&1
+EOF
+)
+
+docker_args+=("$IMAGE" -lc "$container_cmd")
+docker "${docker_args[@]}" >/dev/null
+
+cleanup() {
+  docker rm -f "$CONTAINER_NAME" >/dev/null 2>&1 || true
+}
+trap cleanup EXIT
+
+ready=0
+for _ in $(seq 1 180); do
+  if curl -sf "http://127.0.0.1:${PORT}/v1/models" >/dev/null; then
+    ready=1
+    break
+  fi
+  sleep 2
+done
+if [[ "$ready" -ne 1 ]]; then
+  echo "Server did not become ready on port ${PORT}. Recent logs:" >&2
+  docker logs "$CONTAINER_NAME" 2>&1 | tail -n 120 >&2 || true
+  exit 1
+fi
+
+python3 - <<PY
+import json
+import sys
+import urllib.request
+
+sys.path.insert(0, ${SCRIPT_DIR@Q})
+from profile_common import extract_openai_chat_text, synthetic_prompt
+
+prompt = ${PROMPT@Q} or synthetic_prompt(int(${INPUT_LEN@Q}))
+stage = ${STAGE@Q}
+warmup_steps = int(${WARMUP_STEPS@Q})
+active_steps = int(${ACTIVE_STEPS@Q})
+request_count = warmup_steps + active_steps if stage == "prefill" else 1
+
+payload = {
+    "model": ${MODEL@Q},
+    "messages": [{"role": "user", "content": prompt}],
+    "temperature": 0,
+    "max_tokens": int(${REQUEST_MAX_TOKENS@Q}),
+}
+for request_idx in range(request_count):
+    req = urllib.request.Request(
+        "http://127.0.0.1:${PORT}/v1/chat/completions",
+        data=json.dumps(payload).encode(),
+        headers={"Content-Type": "application/json"},
+    )
+    with urllib.request.urlopen(req, timeout=600) as resp:
+        body = json.loads(resp.read().decode())
+text, source = extract_openai_chat_text(body)
+print(text[:400] if text else f"[empty completion; source={source}]")
+PY
+
+for _ in $(seq 1 120); do
+  if [[ -s "$TRACE_PATH" ]]; then
+    break
+  fi
+  sleep 2
+done
+
+if [[ ! -s "$TRACE_PATH" ]]; then
+  echo "Trace was not written: $TRACE_PATH" >&2
+  exit 1
+fi
+
+python3 "$SCRIPT_DIR/probe_llm_server.py" \
+  --framework trtllm \
+  --url "http://127.0.0.1:${PORT}" \
+  --model "$MODEL" \
+  | docker exec -i sglang_bbuf bash -lc "cat > '$BENCHMARK_PATH'" >/dev/null
+
+echo "TRACE_PATH=$TRACE_PATH"
+echo "LOG_PATH=$LOG_PATH"
+echo "BENCHMARK_PATH=$BENCHMARK_PATH"

.claude/skills/llm-torch-profiler-analysis/scripts/run_vllm_torch_profile_host.sh

@@ -0,0 +1,343 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+usage() {
+  cat <<'EOF'
+Usage:
+  run_vllm_torch_profile_host.sh \
+    --model Qwen/Qwen3-8B \
+    --run-dir /data/bbuf/validate/unified_llm_profiler_skill/runs/example_vllm_formal \
+    --port 31088 \
+    --gpus 1
+
+  run_vllm_torch_profile_host.sh \
+    --model openai/gpt-oss-20b \
+    --run-dir /data/bbuf/validate/unified_llm_profiler_skill/runs/example_vllm_4gpu \
+    --port 31088 \
+    --gpus 2,3,4,5 \
+    --tensor-parallel-size 4
+
+Options:
+  --model TEXT                   Hugging Face model id.
+  --run-dir PATH                Shared /data directory for logs and traces.
+  --port INT                    Host port for vllm serve.
+  --gpus TEXT                   CUDA_VISIBLE_DEVICES value, for example 1 or 2,3,4,5.
+  --gpu TEXT                    Alias for --gpus.
+  --image TEXT                  Container image.
+  --hf-cache PATH               Host Hugging Face cache path.
+  --gpu-memory-util FLOAT       vLLM --gpu-memory-utilization.
+  --max-model-len INT           vLLM --max-model-len.
+    --tensor-parallel-size INT    vLLM --tensor-parallel-size. Defaults to the visible GPU count.
+    --profiler-active-iterations INT
+                                 Torch-profiler active iterations.
+    --enforce-eager               Launch vLLM with --enforce-eager for mapping traces.
+  --trust-remote-code           Pass --trust-remote-code.
+  --request-max-tokens INT      Generation length for the probe request.
+  --prompt TEXT                 Probe prompt.
+  --warmup-steps INT            Warmup steps before profiling. Defaults to 10.
+  --profile-workload TEXT       legacy|prefill|decode|both. Defaults to both.
+  --prefill-input-len INT       Synthetic prefill prompt length. Defaults to 4090.
+  --prefill-output-len INT      Synthetic prefill output length. Defaults to 1.
+  --decode-input-len INT        Synthetic decode prompt length. Defaults to 1.
+  --decode-output-len INT       Synthetic decode output length. Defaults to 2048.
+  --container-name TEXT         Override container name.
+  --help                        Show this message.
+
+Environment:
+  HF_TOKEN or HUGGINGFACE_HUB_TOKEN must be set.
+
+Notes:
+  - Run this on the H100 host, not inside `sglang_bbuf`.
+  - This uses the vLLM torch-profiler flow: `--profiler-config`, then POST
+    `/start_profile` and `/stop_profile`.
+  - Default capture is two labeled profiles: prefill 4090->1 and decode 1->2048.
+  - Current vLLM profiler config already defaults `torch_profiler_with_stack=true`.
+  - A small benchmark summary is written after profiling.
+EOF
+}
+
+IMAGE="vllm/vllm-openai:latest"
+HF_CACHE="/data/.cache/huggingface"
+GPU_MEMORY_UTIL=0.90
+MAX_MODEL_LEN=4096
+TP_SIZE=""
+ENFORCE_EAGER=0
+TRUST_REMOTE_CODE=0
+REQUEST_MAX_TOKENS=12
+PROFILER_ACTIVE_ITERATIONS=5
+PROMPT="Explain the difference between CUDA graph mode and eager mode in two sentences."
+WARMUP_STEPS=10
+PROFILE_WORKLOAD="both"
+PREFILL_INPUT_LEN=4090
+PREFILL_OUTPUT_LEN=1
+DECODE_INPUT_LEN=1
+DECODE_OUTPUT_LEN=2048
+CONTAINER_NAME=""
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+MODEL=""
+RUN_DIR=""
+PORT=""
+GPUS=""
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --model)
+      MODEL="$2"
+      shift 2
+      ;;
+    --run-dir)
+      RUN_DIR="$2"
+      shift 2
+      ;;
+    --port)
+      PORT="$2"
+      shift 2
+      ;;
+    --gpu)
+      GPUS="$2"
+      shift 2
+      ;;
+    --gpus)
+      GPUS="$2"
+      shift 2
+      ;;
+    --image)
+      IMAGE="$2"
+      shift 2
+      ;;
+    --hf-cache)
+      HF_CACHE="$2"
+      shift 2
+      ;;
+    --gpu-memory-util)
+      GPU_MEMORY_UTIL="$2"
+      shift 2
+      ;;
+    --max-model-len)
+      MAX_MODEL_LEN="$2"
+      shift 2
+      ;;
+    --tensor-parallel-size)
+      TP_SIZE="$2"
+      shift 2
+      ;;
+    --profiler-active-iterations)
+      PROFILER_ACTIVE_ITERATIONS="$2"
+      shift 2
+      ;;
+    --enforce-eager)
+      ENFORCE_EAGER=1
+      shift
+      ;;
+    --trust-remote-code)
+      TRUST_REMOTE_CODE=1
+      shift
+      ;;
+    --request-max-tokens)
+      REQUEST_MAX_TOKENS="$2"
+      shift 2
+      ;;
+    --prompt)
+      PROMPT="$2"
+      shift 2
+      ;;
+    --warmup-steps)
+      WARMUP_STEPS="$2"
+      shift 2
+      ;;
+    --profile-workload)
+      PROFILE_WORKLOAD="$2"
+      shift 2
+      ;;
+    --prefill-input-len)
+      PREFILL_INPUT_LEN="$2"
+      shift 2
+      ;;
+    --prefill-output-len)
+      PREFILL_OUTPUT_LEN="$2"
+      shift 2
+      ;;
+    --decode-input-len)
+      DECODE_INPUT_LEN="$2"
+      shift 2
+      ;;
+    --decode-output-len)
+      DECODE_OUTPUT_LEN="$2"
+      shift 2
+      ;;
+    --container-name)
+      CONTAINER_NAME="$2"
+      shift 2
+      ;;
+    --help|-h)
+      usage
+      exit 0
+      ;;
+    *)
+      echo "Unknown argument: $1" >&2
+      usage >&2
+      exit 2
+      ;;
+  esac
+done
+
+if [[ -z "${HF_TOKEN:-}" && -z "${HUGGINGFACE_HUB_TOKEN:-}" ]]; then
+  echo "Set HF_TOKEN or HUGGINGFACE_HUB_TOKEN before running." >&2
+  exit 2
+fi
+if [[ -z "${HF_TOKEN:-}" ]]; then
+  HF_TOKEN="$HUGGINGFACE_HUB_TOKEN"
+fi
+if [[ -z "${HUGGINGFACE_HUB_TOKEN:-}" ]]; then
+  HUGGINGFACE_HUB_TOKEN="$HF_TOKEN"
+fi
+
+if [[ -z "$MODEL" || -z "$RUN_DIR" || -z "$PORT" || -z "$GPUS" ]]; then
+  usage >&2
+  exit 2
+fi
+
+IFS=',' read -r -a GPU_LIST <<< "$GPUS"
+GPU_COUNT="${#GPU_LIST[@]}"
+if [[ "$GPU_COUNT" -lt 1 ]]; then
+  echo "Could not parse --gpus: $GPUS" >&2
+  exit 2
+fi
+if [[ -z "$TP_SIZE" ]]; then
+  TP_SIZE="$GPU_COUNT"
+fi
+if (( TP_SIZE < 1 || TP_SIZE > GPU_COUNT )); then
+  echo "--tensor-parallel-size must be between 1 and the visible GPU count ($GPU_COUNT)." >&2
+  exit 2
+fi
+if (( PROFILER_ACTIVE_ITERATIONS < 1 )); then
+  echo "--profiler-active-iterations must be >= 1." >&2
+  exit 2
+fi
+
+PROFILE_DIR="$RUN_DIR/vllm_profile"
+LOG_PATH="$RUN_DIR/server.log"
+ANALYSIS_PATH="$RUN_DIR/analysis_vllm_live.txt"
+BENCHMARK_PATH="$RUN_DIR/benchmark_vllm.json"
+
+if [[ -z "$CONTAINER_NAME" ]]; then
+  model_slug="${MODEL##*/}"
+  model_slug="${model_slug//\//-}"
+  model_slug="${model_slug//./-}"
+  model_slug="${model_slug//_/-}"
+  gpu_slug="${GPUS//,/-}"
+  CONTAINER_NAME="vllm-${model_slug}-g${gpu_slug}-p${PORT}"
+  if [[ "$ENFORCE_EAGER" -eq 1 ]]; then
+    CONTAINER_NAME="${CONTAINER_NAME}-eager"
+  fi
+fi
+
+docker exec sglang_bbuf bash -lc "mkdir -p '$PROFILE_DIR'"
+docker rm -f "$CONTAINER_NAME" >/dev/null 2>&1 || true
+
+profiler_config=$(python3 - <<PY
+import json
+print(json.dumps({
+    "profiler": "torch",
+    "torch_profiler_dir": ${PROFILE_DIR@Q},
+    "active_iterations": int(${PROFILER_ACTIVE_ITERATIONS@Q}),
+}))
+PY
+)
+
+docker_args=(
+  run -d --rm
+  --name "$CONTAINER_NAME"
+  --gpus all
+  --ipc=host
+  --network host
+  -e "CUDA_VISIBLE_DEVICES=$GPUS"
+  -e "HF_TOKEN=$HF_TOKEN"
+  -e "HUGGINGFACE_HUB_TOKEN=$HUGGINGFACE_HUB_TOKEN"
+  -e "VLLM_RPC_TIMEOUT=1800000"
+  -v "$HF_CACHE:/root/.cache/huggingface"
+  -v "$RUN_DIR:$RUN_DIR"
+)
+
+docker_cmd=(
+  "$IMAGE"
+  "$MODEL"
+  --host 0.0.0.0
+  --port "$PORT"
+  --tensor-parallel-size "$TP_SIZE"
+  --max-model-len "$MAX_MODEL_LEN"
+  --gpu-memory-utilization "$GPU_MEMORY_UTIL"
+  --profiler-config "$profiler_config"
+)
+
+if [[ "$ENFORCE_EAGER" -eq 1 ]]; then
+  docker_cmd+=(--enforce-eager)
+fi
+if [[ "$TRUST_REMOTE_CODE" -eq 1 ]]; then
+  docker_cmd+=(--trust-remote-code)
+fi
+
+docker "${docker_args[@]}" "${docker_cmd[@]}" >/dev/null
+cleanup() {
+  docker rm -f "$CONTAINER_NAME" >/dev/null 2>&1 || true
+}
+trap cleanup EXIT
+
+ready=0
+for _ in $(seq 1 180); do
+  if curl -sf "http://127.0.0.1:${PORT}/v1/models" >/dev/null; then
+    ready=1
+    break
+  fi
+  sleep 2
+done
+if [[ "$ready" -ne 1 ]]; then
+  echo "Server did not become ready on port ${PORT}. Recent logs:" >&2
+  docker logs "$CONTAINER_NAME" 2>&1 | tail -n 120 >&2 || true
+  exit 1
+fi
+
+python3 "$SCRIPT_DIR/analyze_llm_torch_profile.py" \
+  --framework vllm \
+  --url "http://127.0.0.1:${PORT}" \
+  --output-dir "$PROFILE_DIR" \
+  --num-steps "$PROFILER_ACTIVE_ITERATIONS" \
+  --warmup-steps "$WARMUP_STEPS" \
+  --probe-requests 1 \
+  --no-profile-by-stage \
+  --profile-workload "$PROFILE_WORKLOAD" \
+  --probe-prompt "$PROMPT" \
+  --probe-max-new-tokens "$REQUEST_MAX_TOKENS" \
+  --prefill-input-len "$PREFILL_INPUT_LEN" \
+  --prefill-output-len "$PREFILL_OUTPUT_LEN" \
+  --decode-input-len "$DECODE_INPUT_LEN" \
+  --decode-output-len "$DECODE_OUTPUT_LEN" \
+  > "$ANALYSIS_PATH"
+
+profile_found=0
+for _ in $(seq 1 240); do
+  if find "$PROFILE_DIR" -type f \( -name '*.pt.trace.json' -o -name '*.pt.trace.json.gz' -o -name '*.trace.json' -o -name '*.trace.json.gz' \) | grep -q .; then
+    profile_found=1
+    break
+  fi
+  sleep 2
+done
+if [[ "$profile_found" -ne 1 ]]; then
+  echo "No vLLM profiler traces appeared under $PROFILE_DIR" >&2
+  docker logs "$CONTAINER_NAME" 2>&1 | tail -n 120 >&2 || true
+  exit 1
+fi
+
+python3 "$SCRIPT_DIR/probe_llm_server.py" \
+  --framework vllm \
+  --url "http://127.0.0.1:${PORT}" \
+  --model "$MODEL" \
+  | docker exec -i sglang_bbuf bash -lc "cat > '$BENCHMARK_PATH'" >/dev/null
+
+docker logs "$CONTAINER_NAME" 2>&1 | docker exec -i sglang_bbuf bash -lc "cat > '$LOG_PATH'" || true
+sed -n '1,240p' "$ANALYSIS_PATH"
+echo "PROFILE_DIR=$PROFILE_DIR"
+echo "LOG_PATH=$LOG_PATH"
+echo "ANALYSIS_PATH=$ANALYSIS_PATH"
+echo "BENCHMARK_PATH=$BENCHMARK_PATH"

.claude/skills/sglang-auto-benchmark/SKILL.md

@@ -1,496 +0,0 @@
----
-name: sglang-auto-benchmark
-description: Run SGLang auto benchmark searches with tiered server-flag sweeps, canonical dataset preparation, ShareGPT auto-download, custom-data conversion/validation, SLA or fixed-QPS benchmarking, CSV export, and optional second-stage speculative/EAGLE tuning. Use when the user wants an AI-operated benchmark workflow rather than a one-off bench_serving command.
----
-
-# SGLang Auto Benchmark
-
-This skill is for repeatable, AI-driven SGLang performance tuning.
-
-The preferred workflow is:
-- start from a mostly pure-TP baseline command,
-- move the rest of the performance knobs into `search_space`,
-- let auto benchmark search and compare candidates under the target SLA.
-
-The implementation lives in:
-- `python -m sglang.auto_benchmark`
-- canonical dataset loader in `python -m sglang.bench_serving --dataset-name autobench`
-- cookbook-derived LLM reference configs in `.claude/skills/sglang-auto-benchmark/references/cookbook-llm/`
-
-## Preconditions
-
-- SGLang can already launch and serve the target model in this environment.
-- The model path exists, or the model is otherwise launchable.
-- The goal is clear:
-  - benchmark a fixed QPS list, or
-  - search the maximum QPS that satisfies `max_ttft_ms` / `max_tpot_ms`.
-
-If those are not true yet, fix them before running a large search.
-
-Environment consistency check:
-- if the benchmark will run from a remote repo copy or ad-hoc synced workspace,
-  verify that the remote `python/sglang/bench_serving.py` matches the local
-  feature level needed by auto benchmark before launching a long run
-- at minimum, run a preflight such as `PYTHONPATH=<repo>/python python3 -m
-  sglang.bench_serving --help` and confirm that the dataset choices include
-  `autobench`
-- if `autobench` is missing remotely, do not start the benchmark; sync
-  `python/sglang/bench_serving.py` and any required dataset modules first
-
-## Remote Run Logging
-
-If the benchmark is executed on a remote machine, the progress bar output must be
-mirrored back to a local file for humans to watch.
-
-Scope note:
-- use the remote-log mirroring workflow only when the benchmark is running in a
-  different machine or a different remote container than the one the agent is
-  actively operating in
-- if the agent itself is already running inside the target container where auto
-  benchmark is executing, do not add a separate log-return loop just for parity;
-  inspect the live log files and result files directly in the current container
-- in other words, "remote container" needs mirrored local logs, while "current
-  container" should use direct local inspection
-
-Required behavior:
-- start the remote run with a persistent terminal/session log, for example with
-  `script -q -f <log> -c "<cmd>"`; on Linux containers that use util-linux
-  `script`, prefer the explicit `-c` form instead of BSD-style positional
-  command arguments
-- continuously sync a cleaned version of that remote session log back to a local
-  `progress.log`; this local `progress.log` should already have terminal control
-  sequences removed, because `script` + `tqdm` progress bars will otherwise leave
-  ANSI cursor-control bytes and carriage-return redraws that look like garbled text
-- if the benchmark itself is executed inside a remote container, the cleaned local
-  `progress.log` must be refreshed automatically at least once every 30
-  seconds while the run is active; do not rely on one-off manual polling
-- implement the sync loop as a dedicated local script file checked into neither
-  git nor the benchmark config; avoid fragile one-line `nohup zsh -lc '...'`
-  command strings with heavy nested quoting
-- prefer running the sync loop inside a long-lived local session such as a
-  dedicated `tmux` pane, `screen`, or the agent's own persistent PTY session;
-  detached child processes started from short-lived command runners can be
-  reaped unexpectedly, so plain `nohup ... &` is not the most stable default
-- immediately after starting the sync loop, verify that `progress.log` is
-  actually updating by checking its timestamp or size twice across a short wait;
-  if it is not changing, treat that as a broken sync setup and fix it before
-  telling the user that live log mirroring is working
-- tell the user the local log path up front
-- keep final result files synced back locally after the run ends
-- when scenario-level or top-level markdown summaries are produced, sync those
-  `summary.md` / `SUMMARY.md` files back locally as first-class result artifacts
-  rather than leaving them only on the remote machine
-
-This is important because long searches can run for hours, and people need a
-stable local file they can tail without logging into the remote box. The final
-local run folder should also be self-contained enough for someone to review the
-benchmark outcome without re-entering the remote environment.
-
-Recommended cleanup pipeline for the local mirrored log:
-
-```bash
-perl -pe 's/\e\[[0-9;?]*[ -\/]*[@-~]//g; s/\r/\n/g; s/\x08//g;' raw_progress.log \
-  > progress.log
-```
-
-Recommended remote-container sync pattern:
-
-```bash
-cat > sync_progress.sh <<'EOF'
-#!/bin/zsh
-set -euo pipefail
-while true; do
-  ssh <remote-host> "tail -n 200 <remote-progress-log>" > raw_progress.log
-  perl -pe 's/\e\[[0-9;?]*[ -\/]*[@-~]//g; s/\r/\n/g; s/\x08//g;' raw_progress.log \
-    > progress.log
-  sleep 15
-done
-EOF
-chmod +x sync_progress.sh
-```
-
-Run that script from a long-lived local session, for example:
-
-```bash
-tmux new-session -d -s autobench-sync './sync_progress.sh'
-```
-
-Use a persistent local background job, `tmux` pane, `screen`, or equivalent
-long-lived sync process so that humans can watch the cleaned local log in real
-time. Use `sleep 15` by default for long runs unless there is a specific need
-for tighter polling, and keep the cleaned local `progress.log` within the
-required 30-second refresh window while the run is active.
-
-At the end of the run, make sure the local artifact set includes any generated:
-- `results.jsonl`
-- `results.csv`
-- `summary.md`
-- `SUMMARY.md`
-- `scenario_summary.jsonl`
-- `scenario_summary.csv`
-
-Required health check after starting the sync script:
-
-```bash
-stat -f '%m %z' progress.log
-sleep 5
-stat -f '%m %z' progress.log
-```
-
-If the timestamp and size both stay unchanged while the remote benchmark is known
-to be producing new output, the sync loop is broken. Fix the script before
-continuing.
-
-Do not make the cleaned log optional. The default local progress artifact should
-be the cleaned `progress.log` that humans actually read.
-
-## Most Important Rule
-
-If the user wants the best command for a **real production or real workload scenario**, the benchmark must use **their real request distribution**.
-
-That means:
-- real prompt lengths,
-- real output lengths,
-- real multi-turn patterns,
-- real tool / reasoning / sampling settings,
-- real prefix-sharing behavior if it exists.
-
-`sharegpt`, `random`, and `generated-shared-prefix` are useful for sanity checks and broad tuning, but they are not a substitute for the user’s real traffic.
-
-The cookbook reference configs now default to `random` because it is portable and immediately runnable, but that should still be treated as a fallback benchmark shape rather than the final answer for a real deployment.
-
-## Supported Dataset Kinds
-
-The current implementation intentionally keeps the dataset surface small:
-
-- `sharegpt`
-  - Supports auto-download when no file path is provided.
-  - Will be prepared into canonical autobench JSONL on disk before benchmarking.
-- `custom`
-  - Supports two cases:
-    - old `bench_serving` custom conversation JSONL,
-    - already-converted canonical autobench JSONL.
-- `random`
-  - Uses SGLang’s existing synthetic/random benchmark path.
-  - This is the default dataset mode in the cookbook reference configs.
-  - `input_len` and `output_len` can be lists of equal length.
-  - Each aligned pair becomes one full benchmark scenario, not a cartesian product.
-  - Example:
-
-```yaml
-dataset:
-  kind: random
-  scenario_names: [chat, summarization]
-  input_len: [1000, 8000]
-  output_len: [1000, 1000]
-```
-
-  - The workflow will run one full search for `1000 -> 1000` and one full search for `8000 -> 1000`.
-- `generated-shared-prefix`
-  - Uses SGLang’s existing shared-prefix synthetic generator.
-
-Everything is normalized into one canonical autobench JSONL file before the benchmark loop starts.
-
-## Canonical Dataset Format
-
-Canonical format is JSONL, one request per line.
-
-Minimal rows:
-
-```json
-{"prompt": "Write a summary of this document.", "output_len": 256}
-{"prompt": [{"role": "user", "content": "Summarize this document."}], "output_len": 256}
-{"prompt": ["first turn", "follow-up turn"], "output_len": 128}
-```
-
-Optional fields:
-
-```json
-{
-  "prompt": [{"role": "user", "content": "Use the weather tool."}],
-  "output_len": 256,
-  "extra_request_body": {"temperature": 0.0, "top_p": 0.95},
-  "image_data": ["file:///tmp/example.png"],
-  "timestamp": 1710000000,
-  "routing_key": "group-a",
-  "metadata": {"source": "custom-upload"}
-}
-```
-
-Compatibility:
-- legacy `messages`
-- legacy `prompt_origin`
-- legacy `param_send`
-- legacy `system + content`
-
-## ShareGPT Auto-Prepare
-
-`sharegpt` does not need a full path.
-
-Example:
-
-```bash
-python3 -m sglang.auto_benchmark convert \
-  --kind sharegpt \
-  --tokenizer /path/to/tokenizer \
-  --num-prompts 200 \
-  --output /tmp/sharegpt.autobench.jsonl
-```
-
-This will:
-- auto-download ShareGPT through the existing SGLang cache path when needed,
-- convert it into canonical autobench JSONL,
-- save it to the requested output path.
-
-## Custom User Data Workflow
-
-When the user uploads custom data:
-
-1. Inspect a few raw rows first.
-2. Decide whether the file is:
-   - already canonical autobench JSONL,
-   - old `bench_serving` custom format,
-   - or an unsupported custom schema that must be transformed manually.
-3. If manual transformation is needed:
-   - map it into canonical JSONL,
-   - never hallucinate missing turns or answers,
-   - never keep the final assistant answer as part of the benchmark prompt if that answer is the target completion,
-   - preserve per-request generation settings in `extra_request_body`.
-4. Run:
-
-```bash
-python3 -m sglang.auto_benchmark validate \
-  --dataset-path /path/to/converted.autobench.jsonl \
-  --tokenizer /path/to/tokenizer
-```
-
-5. Manually inspect at least 3 converted rows and confirm:
-   - prompt shape is correct,
-   - final assistant answer was not accidentally left in the prompt,
-   - `output_len` is sensible,
-   - request extras were preserved.
-
-## Search Tiers
-
-`search.tier` controls search breadth.
-
-- Tier 1
-  - Fastest and smallest sweep.
-  - Best for quick checks, config validation, and confirming whether a model can run at all.
-  - Uses a very small subset of the search space and mainly does one-at-a-time changes on top of the baseline.
-  - Lowest search cost, but also the easiest to miss a better configuration.
-- Tier 2
-  - Recommended default.
-  - Good balance between coverage and runtime.
-  - Runs a small cartesian search on the first few high-priority keys, then expands the rest one at a time.
-  - Usually the right choice for everyday tuning when you want meaningful search without waiting too long.
-- Tier 3
-  - Largest search space.
-  - Runs the full cartesian product of the provided search space.
-  - Search time is the longest by far.
-  - Only use it when the search space is already tightly bounded and you intentionally want the most exhaustive sweep.
-  - This is the best chance of finding the strongest config, but it is also the easiest way to turn a benchmark into a multi-hour or multi-day run.
-
-`search.max_candidates` still applies at all tiers, including tier 3.
-When it is set together with tier 3, the workflow still enumerates the full cartesian order conceptually, but only keeps the first `max_candidates` unique candidates after deduplication.
-That makes it useful as a safety valve, but it also means tier 3 is no longer truly exhaustive unless you remove the cap or raise it high enough.
-
-If `search.max_candidates` is omitted, the workflow now defaults to `8`.
-Set it to `null` only when you intentionally want an unbounded sweep.
-
-The reference configs now default to tier 2 with `search.max_candidates: 8`.
-
-## Interrupt And Resume
-
-Long searches may need to be stopped and resumed later.
-
-Use:
-
-```yaml
-search:
-  tier: 2
-  resume: true
-```
-
-Behavior:
-- every completed trial is appended to `live_results.jsonl`
-- if the process receives `SIGINT` or `SIGTERM`, it will first save partial
-  `results.jsonl`, `results.csv`, and `summary.md`
-- on the next run with the same config and `search.resume: true`, completed
-  trials are reused and only unfinished trials are executed
-- resume works per scenario directory, so it is safest to keep the same
-  `benchmark.output_dir`
-
-Notes:
-- resume assumes the candidate order and dataset are unchanged
-- for maximum safety, reuse the same prepared dataset or keep the same dataset
-  seed/config
-- `SIGKILL` cannot be handled gracefully, so only the already-written
-  `live_results.jsonl` can be reused after a hard kill
-
-YAML key order matters. Put the most important search keys first.
-
-## What Is Tunable
-
-This workflow is not limited to attention backend tuning.
-
-`server.base_flags` and `server.search_space` are passed directly to `sglang.launch_server`, so in practice any valid server CLI flag can be set or searched.
-
-There is also a small convenience layer for parallel search:
-
-- `server.parallel.tp`
-- `server.parallel.pp_size`
-
-When `server.parallel` is used and `dp_size` is not set explicitly, the workflow auto-derives:
-
-`dp_size = visible_gpus / (tp_size * pp_size)`
-
-Visible GPU count is inferred from `server.env.CUDA_VISIBLE_DEVICES` by default, or from `server.parallel.gpu_count` if you set it explicitly.
-
-The most important performance-related groups are:
-
-- Kernel / backend
-  - `attention_backend`
-  - `prefill_attention_backend`
-  - `decode_attention_backend`
-  - `sampling_backend`
-  - `grammar_backend`
-- Batching / scheduling
-  - `max_running_requests`
-  - `max_queued_requests`
-  - `chunked_prefill_size`
-  - `prefill_max_requests`
-  - `max_prefill_tokens`
-  - `schedule_conservativeness`
-  - `num_continuous_decode_steps`
-  - `stream_interval`
-- Memory / cache
-  - `max_total_tokens`
-  - `page_size`
-  - `disable_radix_cache`
-- Parallel / distributed execution
-  - `tp_size`
-  - `pp_size`
-  - `dp_size`
-  - `ep_size`
-  - `load_balance_method`
-  - `enable_dp_attention`
-  - `enable_mixed_chunk`
-  - `disable_overlap_schedule`
-- Runtime / CUDA graph
-  - keep CUDA graph enabled by default for performance benchmarking
-  - `cuda_graph_max_bs`
-  - `disable_cuda_graph_padding`
-  - `enable_cudagraph_gc`
-- Optional speculative / EAGLE stage
-  - `speculative_num_steps`
-  - `speculative_eagle_topk`
-  - `speculative_num_draft_tokens`
-  - `speculative_attention_mode`
-  - `speculative_draft_attention_backend`
-  - `speculative_accept_threshold_single`
-  - `speculative_accept_threshold_acc`
-
-For cookbook-derived reference configs, keep `mem_fraction_static` and
-`schedule_policy` pinned to the cookbook baseline unless the user explicitly
-asks to search them. They are useful knobs, but they add a lot of search width
-for relatively low validation value in the default workflow.
-
-Do not put these into the default search space:
-- `mem_fraction_static`
-- `schedule_policy`
-- `enable_hierarchical_cache`
-- `hicache_ratio`
-- `hicache_size`
-- `enable_lmcache`
-
-Those features are not treated as standard auto-benchmark sweep knobs in this workflow.
-
-Budget guardrails for the default workflow:
-- use `dataset.num_prompts: 80` unless the user asks for a heavier study
-- prefer a coarse QPS search tolerance
-- keep `benchmark.qps.max_rounds <= 5`
-- keep `search.max_duration_hours <= 12`
-
-## Base Tuning Before EAGLE
-
-Never start by tuning EAGLE first.
-
-Use this order:
-
-1. Tune the non-speculative base server first.
-2. Find the best normal config for the target dataset and SLA.
-3. Only if the user explicitly asks for speculative/EAGLE tuning, and provides the required draft model or equivalent assets, run the second-stage speculative search.
-
-Do not put `disable_cuda_graph` into the default search space. For normal performance tuning, CUDA graph should stay enabled unless the user is debugging compatibility issues.
-
-When a candidate OOMs, keep it in the final result table as a failed row and add a hint such as:
-- increase GPU count, or
-- use GPUs with larger memory.
-
-## Running The Workflow
-
-Prepare a dataset explicitly:
-
-```bash
-python3 -m sglang.auto_benchmark convert \
-  --kind custom \
-  --path /path/to/data.jsonl \
-  --tokenizer /path/to/tokenizer \
-  --output /tmp/data.autobench.jsonl
-```
-
-Run from config:
-
-```bash
-python3 -m sglang.auto_benchmark run --config /path/to/config.yaml
-```
-
-Outputs:
-- prepared canonical dataset JSONL
-- per-run `results.jsonl`
-- summary `results.csv`
-- per-candidate server logs
-
-## Config Template
-
-Standalone example (uses ShareGPT as dataset, a good starting point for non-cookbook models):
-- `references/qwen3-32b.yaml`
-
-Cookbook-derived configs live in `references/cookbook-llm/`.
-They default to synthetic `random` traffic and are runnable out of the box.
-See `references/cookbook-llm/README.md` for the full list.
-
-Representative picks from that folder:
-- `references/cookbook-llm/llama-3.1-70b-instruct.yaml`
-- `references/cookbook-llm/llama-3.3-70b-instruct.yaml`
-- `references/cookbook-llm/llama-4-scout-17b-16e-instruct.yaml`
-- `references/cookbook-llm/llama-4-maverick-17b-128e-instruct-fp8.yaml`
-- `references/cookbook-llm/minimax-m2.5.yaml`
-- `references/cookbook-llm/minimax-m2.1.yaml`
-- `references/cookbook-llm/deepseek-v3.yaml`
-- `references/cookbook-llm/deepseek-v3.1.yaml`
-- `references/cookbook-llm/deepseek-v3.2.yaml`
-- `references/cookbook-llm/deepseek-r1-0528.yaml`
-- `references/cookbook-llm/qwen3-235b-a22b.yaml`
-- `references/cookbook-llm/qwen35-397b-a17b-fp8.yaml`
-- `references/cookbook-llm/mistral-small-4-119b-2603.yaml`
-- `references/cookbook-llm/kimi-k2-instruct.yaml`
-
-All reference configs use Hugging Face repo IDs by default.
-Replace `model_path` and `tokenizer` with local paths when the weights are already on disk.
-
-## What To Report Back
-
-After a run, summarize:
-- which tier was used,
-- which dataset kind was used,
-- whether the dataset was synthetic or real user traffic,
-- best base config,
-- best QPS that satisfied SLA,
-- whether speculative tuning was skipped or run,
-- paths to:
-  - prepared dataset JSONL
-  - `results.jsonl`
-  - `results.csv`
-  - key server logs

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/README.md

@@ -1,76 +0,0 @@
-# Cookbook LLM References
-
-These configs are derived from `sgl-cookbook` autoregressive text-model pages and normalized into the auto-benchmark config format.
-
-Rules used here:
-- Keep the baseline as close as possible to a pure-TP launch command.
-- Keep `mem_fraction_static` and `schedule_policy` at the cookbook baseline by
-  default; search higher-ROI knobs first.
-- Move the remaining common performance knobs into `search_space`.
-- Add `ep_size` search for relevant MoE pages.
-- Keep CUDA graph enabled by default.
-- Prefer cookbook H200 defaults first, then H100 defaults when H200 is not available; if neither exists, fall back to the cookbook's published baseline for that model and say so in the config comments.
-- Default to synthetic `random` data so every config is runnable out of the box.
-- Default to `dataset.num_prompts: 80` so the reference sweep stays cheap enough
-  for interactive validation.
-- Default to a coarse QPS search with `benchmark.qps.max_rounds <= 5`.
-- Default to `search.tier: 2` so the shipped configs stay reasonably practical to run.
-- Default to `search.max_candidates: 8` so the candidate sweep stays bounded by
-  default.
-- Default to `search.max_duration_hours: 12` because longer searches do not fit
-  the intended workflow budget.
-- Treat `dataset.input_len` and `dataset.output_len` as aligned scenario lists, not a cartesian product.
-- If a candidate OOMs, the result table should recommend increasing GPU count or using GPUs with larger memory.
-
-Default random scenarios in these configs:
-- `1000 -> 1000` for a chat-like shape
-- `8000 -> 1000` for a summarization-like shape
-
-Each scenario should run a full search independently, and each scenario should have its own best launch command and summary table.
-
-Excluded from this folder because they are OCR/VL-oriented rather than text-serving benchmark configs:
-- DeepSeekOCR / DeepSeekOCR2
-- GLMOCR
-- GLM45V / GLM46V
-- Qwen2.5-VL / Qwen3-VL
-- Step3-VL-10B
-
-Configs in this folder:
-- `deepseek-v3.2.yaml`
-- `deepseek-math-v2.yaml`
-- `deepseek-r1-0528.yaml`
-- `deepseek-v3.1.yaml`
-- `deepseek-v3.yaml`
-- `devstral-small-2-24b-instruct-2512.yaml`
-- `ernie-4.5-21b-a3b-pt.yaml`
-- `glm-4.5.yaml`
-- `glm-4.6.yaml`
-- `glm-4.7.yaml`
-- `glm-4.7-flash.yaml`
-- `glm-5-fp8.yaml`
-- `gpt-oss-120b.yaml`
-- `glyph.yaml`
-- `intern-s1.yaml`
-- `kimi-k2.5.yaml`
-- `kimi-k2-instruct.yaml`
-- `kimi-linear-48b-a3b-instruct.yaml`
-- `llada2-1-mini.yaml`
-- `ling-2.5-1t.yaml`
-- `llama-3.1-70b-instruct.yaml`
-- `llama-3.3-70b-instruct.yaml`
-- `llama-4-scout-17b-16e-instruct.yaml`
-- `llama-4-maverick-17b-128e-instruct-fp8.yaml`
-- `mimo-v2-flash.yaml`
-- `minimax-m2.5.yaml`
-- `minimax-m2.1.yaml`
-- `ministral-3-8b-instruct-2512.yaml`
-- `mistral-small-4-119b-2603.yaml`
-- `nemotron-3-nano-30b-a3b-bf16.yaml`
-- `nemotron-3-super-120b-a12b-bf16.yaml`
-- `qwen35-397b-a17b-fp8.yaml`
-- `qwen3-coder-480b-a35b-instruct.yaml`
-- `qwen3-coder-next.yaml`
-- `qwen3-235b-a22b.yaml`
-- `qwen3-next-80b-a3b-instruct.yaml`
-- `ring-2.5-1t.yaml`
-- `step-3.5-flash.yaml`

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/deepseek-math-v2.yaml

@@ -1,81 +0,0 @@
-# Cookbook auto benchmark config for deepseek math v2.
-# Cookbook does not expose an H100/H200 default for this model; this config follows the cookbook B200 baseline instead.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.8/deepseek-math-v2.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    model_path: deepseek-ai/DeepSeek-Math-V2
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - flashinfer
-    decode_attention_backend:
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-    - 8
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: deepseek-ai/DeepSeek-Math-V2
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/deepseek-math-v2
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/deepseek-r1-0528.yaml

@@ -1,84 +0,0 @@
-# Cookbook auto benchmark config for deepseek r1 0528.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/deepseek-r1.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    enable_symm_mem: true
-    model_path: deepseek-ai/DeepSeek-R1-0528
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-    - 8
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: deepseek-ai/DeepSeek-R1-0528
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/deepseek-r1-0528
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/deepseek-v3.1.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for deepseek v3.1.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/DeepSeekV31ConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    model_path: deepseek-ai/DeepSeek-V3.1
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-    - 8
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: deepseek-ai/DeepSeek-V3.1
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/deepseek-v3.1
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/deepseek-v3.2.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for deepseek v3.2.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/deepseek.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    model_path: deepseek-ai/DeepSeek-V3.2
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-    - 8
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: deepseek-ai/DeepSeek-V3.2
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/deepseek-v3.2
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/deepseek-v3.yaml

@@ -1,84 +0,0 @@
-# Cookbook auto benchmark config for deepseek v3.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/DeepSeekV3ConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    enable_symm_mem: true
-    model_path: deepseek-ai/DeepSeek-V3
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-    - 8
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: deepseek-ai/DeepSeek-V3
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/deepseek-v3
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/devstral-small-2-24b-instruct-2512.yaml

@@ -1,78 +0,0 @@
-# Cookbook auto benchmark config for devstral small 2 24b instruct 2512.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/Devstral2ConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: '0'
-  base_flags:
-    model_path: mistralai/Devstral-Small-2-24B-Instruct-2512
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: mistralai/Devstral-Small-2-24B-Instruct-2512
-  max_concurrency:
-  - null
-  - 16
-  - 32
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 16.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/devstral-small-2-24b-instruct-2512
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/ernie-4.5-21b-a3b-pt.yaml

@@ -1,76 +0,0 @@
-# Cookbook auto benchmark config for ernie 4.5 21b a3b pt.
-# Cookbook does not expose an H100/H200 default for this model; this config follows the cookbook MI300X baseline instead.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/Ernie45ConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: '0'
-  base_flags:
-    model_path: baidu/ERNIE-4.5-21B-A3B-PT
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-  command_prefix:
-  - python3
-  - -m
-  - sglang.launch_server
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: baidu/ERNIE-4.5-21B-A3B-PT
-  max_concurrency:
-  - null
-  - 16
-  - 32
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 16.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/ernie-4.5-21b-a3b-pt
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/glm-4.5.yaml

@@ -1,74 +0,0 @@
-# Cookbook auto benchmark config for glm 4.5.
-# Cookbook does not expose an H100/H200 default for this model; this config follows the cookbook MI300X baseline instead.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/GLM45ConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3
-  base_flags:
-    tp_size: 4
-    context_length: 8192
-    model_path: zai-org/GLM-4.5
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: zai-org/GLM-4.5
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 8.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/glm-4.5
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/glm-4.6.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for glm 4.6.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/glm46.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    model_path: zai-org/GLM-4.6
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-    - 8
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: zai-org/GLM-4.6
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/glm-4.6
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/glm-4.7-flash.yaml

@@ -1,78 +0,0 @@
-# Cookbook auto benchmark config for glm 4.7 flash.
-# Baseline follows the cookbook H100 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/GLM47FlashConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: '0'
-  base_flags:
-    model_path: zai-org/GLM-4.7-Flash
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: zai-org/GLM-4.7-Flash
-  max_concurrency:
-  - null
-  - 16
-  - 32
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 16.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/glm-4.7-flash
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/glm-4.7.yaml

@@ -1,78 +0,0 @@
-# Cookbook auto benchmark config for glm 4.7.
-# Cookbook does not expose an H100/H200 default for this model; this config follows the cookbook MI300X baseline instead.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/GLM47ConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3
-  base_flags:
-    tp_size: 4
-    context_length: 8192
-    model_path: zai-org/GLM-4.7
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-    ep_size:
-    - 1
-    - 2
-    - 4
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: zai-org/GLM-4.7
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 8.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/glm-4.7
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/glm-5-fp8.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for glm 5 fp8.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.8/glm5.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    model_path: zai-org/GLM-5-FP8
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-    - 8
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: zai-org/GLM-5-FP8
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/glm-5-fp8
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/glyph.yaml

@@ -1,81 +0,0 @@
-# Cookbook auto benchmark config for glyph.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/GlyphConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3
-  base_flags:
-    tp_size: 4
-    reasoning_parser: glm45
-    tool_call_parser: glm45
-    model_path: zai-org/Glyph
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: zai-org/Glyph
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 8.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/glyph
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/gpt-oss-120b.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for gpt oss 120b.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/gpt-oss.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    model_path: openai/gpt-oss-120b
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-    - 8
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: openai/gpt-oss-120b
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/gpt-oss-120b
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/intern-s1.yaml

@@ -1,84 +0,0 @@
-# Cookbook auto benchmark config for intern s1.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/intern-s1.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    trust_remote_code: true
-    model_path: internlm/Intern-S1
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-    - 8
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: internlm/Intern-S1
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/intern-s1
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/kimi-k2-instruct.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for kimi k2 instruct.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/kimi-k2.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    trust_remote_code: true
-    model_path: moonshotai/Kimi-K2-Instruct
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: moonshotai/Kimi-K2-Instruct
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/kimi-k2-instruct
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/kimi-k2.5.yaml

@@ -1,80 +0,0 @@
-# Cookbook auto benchmark config for kimi k2.5.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.8/kimi-k25.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    trust_remote_code: true
-    model_path: moonshotai/Kimi-K2.5
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: moonshotai/Kimi-K2.5
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/kimi-k2.5
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/kimi-linear-48b-a3b-instruct.yaml

@@ -1,79 +0,0 @@
-# Cookbook auto benchmark config for kimi linear 48b a3b instruct.
-# Cookbook does not expose an H100/H200 default for this model; this config follows the cookbook MI300X baseline instead.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/KimiK2linearConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    SGLANG_ROCM_FUSED_DECODE_MLA: '0'
-    CUDA_VISIBLE_DEVICES: 0,1,2,3
-  base_flags:
-    tp_size: 4
-    trust_remote_code: true
-    model_path: moonshotai/Kimi-Linear-48B-A3B-Instruct
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-  command_prefix:
-  - python3
-  - -m
-  - sglang.launch_server
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: moonshotai/Kimi-Linear-48B-A3B-Instruct
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 8.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/kimi-linear-48b-a3b-instruct
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/ling-2.5-1t.yaml

@@ -1,91 +0,0 @@
-# Cookbook auto benchmark config for ling 2.5 1t.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# This cookbook baseline is multi-node or external-service oriented. Keep launch: false and benchmark an already running deployment.
-# Source: src/components/autoregressive/Ling25ConfigGenerator/index.js
-server:
-  launch: false
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    pp_size: 2
-    nnodes: 2
-    trust_remote_code: true
-    tool_call_parser: qwen
-    model_path: inclusionAI/Ling-2.5-1T
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    pp_size:
-    - 1
-    - 2
-  command_prefix:
-  - python3
-  - -m
-  - sglang.launch_server
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: inclusionAI/Ling-2.5-1T
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 2.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/ling-2.5-1t
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/llada2-1-mini.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for llada2 1 mini.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/llada21.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: '0'
-  base_flags:
-    tp_size: 1
-    dllm_algorithm: JointThreshold
-    trust_remote_code: true
-    max_running_requests: 1
-    attention_backend: flashinfer
-    model_path: inclusionAI/LLaDA2.1-mini
-    mem_fraction_static: 0.77
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 1
-    - 2
-    - 4
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: inclusionAI/LLaDA2.1-mini
-  max_concurrency:
-  - 1
-  - 2
-  - 4
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/llada2-1-mini
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/llama-3.1-70b-instruct.yaml

@@ -1,79 +0,0 @@
-# Cookbook auto benchmark config for llama 3.1 70b instruct.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/llama31.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3
-  base_flags:
-    tp_size: 4
-    model_path: meta-llama/Llama-3.1-70B-Instruct
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: meta-llama/Llama-3.1-70B-Instruct
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 12.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/llama-3.1-70b-instruct
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/llama-3.3-70b-instruct.yaml

@@ -1,73 +0,0 @@
-# Cookbook auto benchmark config for llama 3.3 70b instruct.
-# Cookbook does not expose an H100/H200 default for this model; this config follows the cookbook MI300X baseline instead.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/Llama33ConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: '0'
-  base_flags:
-    tool_call_parser: llama3
-    model_path: meta-llama/Llama-3.3-70B-Instruct
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: meta-llama/Llama-3.3-70B-Instruct
-  max_concurrency:
-  - null
-  - 16
-  - 32
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 16.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/llama-3.3-70b-instruct
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/llama-4-maverick-17b-128e-instruct-fp8.yaml

@@ -1,79 +0,0 @@
-# Cookbook auto benchmark config for llama 4 maverick 17b 128e instruct fp8.
-# Cookbook does not expose an H100/H200 default for this model; this config follows the cookbook GENERIC-8GPU baseline instead.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: docs/autoregressive/Llama/Llama4.md
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    context_length: 1000000
-    trust_remote_code: true
-    enable_multimodal: true
-    model_path: meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 4
-    - 8
-    - 12
-  command_prefix:
-  - sglang
-  - serve
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8
-  max_concurrency:
-  - null
-  - 2
-  - 4
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 2.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/llama-4-maverick-17b-128e-instruct-fp8
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE3
-  draft_model_path: lmsys/sglang-EAGLE3-Llama-4-Maverick-17B-128E-Instruct-v1
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/llama-4-scout-17b-16e-instruct.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for llama 4 scout 17b 16e instruct.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/llama4scout.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    enable_multimodal: true
-    context_length: 65536
-    dtype: bfloat16
-    trust_remote_code: true
-    model_path: meta-llama/Llama-4-Scout-17B-16E-Instruct
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 8
-    - 16
-    - 24
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: meta-llama/Llama-4-Scout-17B-16E-Instruct
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/llama-4-scout-17b-16e-instruct
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE3
-  draft_model_path: lmsys/sglang-EAGLE3-Llama-4-Scout-17B-16E-Instruct-v1
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/mimo-v2-flash.yaml

@@ -1,91 +0,0 @@
-# Cookbook auto benchmark config for mimo v2 flash.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/MiMoConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    trust_remote_code: true
-    max_running_requests: 128
-    chunked_prefill_size: 16384
-    model_loader_extra_config: '{"enable_multithread_load": "true","num_threads":
-      64}'
-    attention_backend: fa3
-    reasoning_parser: qwen3
-    tool_call_parser: mimo
-    model_path: XiaomiMiMo/MiMo-V2-Flash
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-  command_prefix:
-  - python3
-  - -m
-  - sglang.launch_server
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: XiaomiMiMo/MiMo-V2-Flash
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/mimo-v2-flash
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/minimax-m2.1.yaml

@@ -1,77 +0,0 @@
-# Cookbook auto benchmark config for minimax m2.1.
-# Cookbook does not expose an H100/H200 default for this model; this config follows the cookbook MI300X baseline instead.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/MiniMaxM2ConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3
-  base_flags:
-    tp_size: 4
-    trust_remote_code: true
-    model_path: MiniMaxAI/MiniMax-M2.1
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-  command_prefix:
-  - sglang
-  - serve
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: MiniMaxAI/MiniMax-M2.1
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 8.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/minimax-m2.1
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/minimax-m2.5.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for minimax m2.5.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/MiniMaxM25ConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3
-  base_flags:
-    tp_size: 4
-    trust_remote_code: true
-    model_path: MiniMaxAI/MiniMax-M2.5
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-    ep_size:
-    - 1
-    - 4
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: MiniMaxAI/MiniMax-M2.5
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/minimax-m2.5
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/ministral-3-8b-instruct-2512.yaml

@@ -1,77 +0,0 @@
-# Cookbook auto benchmark config for ministral 3 8b instruct 2512.
-# Cookbook does not expose an H100/H200 default for this model; this config follows the cookbook MI300X baseline instead.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/Ministral3ConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: '0'
-  base_flags:
-    trust_remote_code: true
-    tool_call_parser: mistral
-    model_path: mistralai/Ministral-3-8B-Instruct-2512
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-  command_prefix:
-  - sglang
-  - serve
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: mistralai/Ministral-3-8B-Instruct-2512
-  max_concurrency:
-  - null
-  - 16
-  - 32
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 16.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/ministral-3-8b-instruct-2512
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/mistral-small-4-119b-2603.yaml

@@ -1,79 +0,0 @@
-# Cookbook auto benchmark config for mistral small 4 119b 2603.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.8/mistral-small-4.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1
-  base_flags:
-    tp_size: 2
-    model_path: mistralai/Mistral-Small-4-119B-2603
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: mistralai/Mistral-Small-4-119B-2603
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 6.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/mistral-small-4-119b-2603
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/nemotron-3-nano-30b-a3b-bf16.yaml

@@ -1,81 +0,0 @@
-# Cookbook auto benchmark config for nemotron 3 nano 30b a3b bf16.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/nemotron.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: '0'
-  base_flags:
-    tp_size: 1
-    trust_remote_code: true
-    kv_cache_dtype: fp8_e4m3
-    model_path: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
-  max_concurrency:
-  - null
-  - 16
-  - 32
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 16.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/nemotron-3-nano-30b-a3b-bf16
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/nemotron-3-super-120b-a12b-bf16.yaml

@@ -1,81 +0,0 @@
-# Cookbook auto benchmark config for nemotron 3 super 120b a12b bf16.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.8/nemotron-super.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3
-  base_flags:
-    tp_size: 4
-    trust_remote_code: true
-    kv_cache_dtype: fp8_e4m3
-    model_path: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 6.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/nemotron-3-super-120b-a12b-bf16
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/qwen3-235b-a22b.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for qwen3 235b a22b.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/qwen.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    model_path: Qwen/Qwen3-235B-A22B
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-    - 8
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: Qwen/Qwen3-235B-A22B
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/qwen3-235b-a22b
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/qwen3-coder-480b-a35b-instruct.yaml

@@ -1,82 +0,0 @@
-# Cookbook auto benchmark config for qwen3 coder 480b a35b instruct.
-# Cookbook does not expose an H100/H200 default for this model; this config follows the cookbook B200 baseline instead.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: src/components/autoregressive/Qwen3CoderConfigGenerator/index.js
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    ep_size: 2
-    moe_runner_backend: triton
-    model_path: Qwen/Qwen3-Coder-480B-A35B-Instruct
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - flashinfer
-    decode_attention_backend:
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 2
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: Qwen/Qwen3-Coder-480B-A35B-Instruct
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/qwen3-coder-480b-a35b-instruct
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/qwen3-coder-next.yaml

@@ -1,79 +0,0 @@
-# Cookbook auto benchmark config for qwen3 coder next.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.8/qwen3codernext.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1
-  base_flags:
-    tp_size: 2
-    model_path: Qwen/Qwen3-Coder-Next
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: Qwen/Qwen3-Coder-Next
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 12.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/qwen3-coder-next
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/qwen3-next-80b-a3b-instruct.yaml

@@ -1,82 +0,0 @@
-# Cookbook auto benchmark config for qwen3 next 80b a3b instruct.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.6/qwen3next.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1
-  base_flags:
-    tp_size: 2
-    model_path: Qwen/Qwen3-Next-80B-A3B-Instruct
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-    ep_size:
-    - 1
-    - 2
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: Qwen/Qwen3-Next-80B-A3B-Instruct
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 12.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/qwen3-next-80b-a3b-instruct
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/qwen35-397b-a17b-fp8.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for qwen35 397b a17b fp8.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.8/qwen35.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3
-  base_flags:
-    tp_size: 4
-    model_path: Qwen/Qwen3.5-397B-A17B-FP8
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-    ep_size:
-    - 1
-    - 4
-    - 8
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: Qwen/Qwen3.5-397B-A17B-FP8
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 4.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/qwen35-397b-a17b-fp8
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/ring-2.5-1t.yaml

@@ -1,79 +0,0 @@
-# Cookbook auto benchmark config for ring 2.5 1t.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.8/ring25.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
-  base_flags:
-    tp_size: 8
-    model_path: inclusionAI/Ring-2.5-1T
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 32
-    - 48
-    - 64
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: inclusionAI/Ring-2.5-1T
-  max_concurrency:
-  - null
-  - 4
-  - 8
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 2.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/ring-2.5-1t
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/cookbook-llm/step-3.5-flash.yaml

@@ -1,83 +0,0 @@
-# Cookbook auto benchmark config for step 3.5 flash.
-# Baseline follows the cookbook H200 default command.
-# If you run on larger-memory or stronger GPUs, you can often reduce tp/ep/pp scale or total GPU count.
-# Replace server.base_flags.model_path with a local model path if the weights are already present on disk.
-# These configs default to synthetic random traffic. Use your real workload data if you want production-faithful tuning.
-# Each input_len/output_len pair defines one benchmark scenario, for example [1000, 1000] and [8000, 1000].
-# Source: data/models/generated/v0.5.8/step35.yaml
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: 0,1,2,3
-  base_flags:
-    tp_size: 4
-    trust_remote_code: true
-    model_path: stepfun-ai/Step-3.5-Flash
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  search_space:
-    prefill_attention_backend:
-    - fa3
-    - flashinfer
-    decode_attention_backend:
-    - fa3
-    - flashinfer
-    chunked_prefill_size:
-    - 4096
-    - 8192
-    max_running_requests:
-    - 64
-    - 96
-    - 128
-    ep_size:
-    - 1
-    - 4
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names:
-  - chat
-  - summarization
-  input_len:
-  - 1000
-  - 8000
-  output_len:
-  - 1000
-  - 1000
-benchmark:
-  backend: auto
-  tokenizer: stepfun-ai/Step-3.5-Flash
-  max_concurrency:
-  - null
-  - 8
-  - 16
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 0.25
-    upper: 8.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/cookbook-llm/step-3.5-flash
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  draft_model_path: ''
-  search_space:
-    speculative_num_steps:
-    - 3
-    - 5
-    speculative_eagle_topk:
-    - 1
-    - 2
-    speculative_num_draft_tokens:
-    - 4
-    - 8

.claude/skills/sglang-auto-benchmark/references/qwen3-32b.yaml

@@ -1,62 +0,0 @@
-server:
-  launch: true
-  host: 127.0.0.1
-  port: 30000
-  env:
-    CUDA_VISIBLE_DEVICES: "0"
-  # Optional parallel search sugar:
-  # parallel:
-  #   # dp_size is auto-derived as visible_gpus / (tp_size * pp_size).
-  #   tp: [4, 2]
-  #   # pp_size: [1, 2]
-  base_flags:
-    # Replace with a local model path if you have one.
-    model_path: Qwen/Qwen3-32B
-    tp_size: 1
-    trust_remote_code: true
-    # Keep the baseline close to a pure-TP launch command.
-    mem_fraction_static: 0.82
-    schedule_policy: lpm
-  # Keep CUDA graph enabled for normal performance tuning.
-  search_space:
-    prefill_attention_backend: [fa3, flashinfer]
-    decode_attention_backend: [fa3, flashinfer]
-    chunked_prefill_size: [4096, 8192]
-    max_running_requests: [64, 96, 128]
-dataset:
-  kind: random
-  num_prompts: 80
-  scenario_names: [chat, summarization]
-  input_len: [1000, 8000]
-  output_len: [1000, 1000]
-
-benchmark:
-  backend: auto
-  # Replace with a local tokenizer/model path if you have one.
-  tokenizer: Qwen/Qwen3-32B
-  max_concurrency: [null, 16, 32]
-  extra_request_body:
-    temperature: 0.0
-  qps:
-    lower: 1.0
-    upper: 12.0
-    tolerance: 0.1
-  sla:
-    max_ttft_ms: 1500
-    max_tpot_ms: 30
-  output_dir: ./auto_benchmark_results/qwen3-32b
-
-search:
-  tier: 2
-  max_candidates: 8  # Cap search breadth so full searches stay tractable.
-  resume: true
-
-speculative:
-  enabled: false
-  algorithm: EAGLE
-  # Fill this only if you explicitly want a second-stage speculative search.
-  draft_model_path: ""
-  search_space:
-    speculative_num_steps: [3, 5]
-    speculative_eagle_topk: [1, 4]
-    speculative_num_draft_tokens: [4, 8]

.claude/skills/sglang-prod-incident-triage/SKILL.md

@@ -0,0 +1,291 @@
+---
+name: sglang-prod-incident-triage
+description: Replay-first debug flow for SGLang serving problems. Use when a live or recent server shows health-check failures, latency or throughput regressions, queue growth, timeouts, distributed stalls, crash dumps, wrong outputs after deploys, or PD/EP/HiCache issues, and the job is to turn the problem into a replay plus the right next debug tool.
+---
+
+# SGLang Serving Debug
+
+## Overview
+
+Use this skill to turn a live serving problem into a debug path you can replay.
+
+Use one loop:
+
+- collect a baseline bundle
+- save the failing request or crash dump
+- replay on a clean target
+- only then switch tools
+
+Do not start with profiling.
+
+This skill should work with more focused skills instead of re-implementing them:
+
+- `debug-cuda-crash` when replay plus coredump points to a CUDA crash path
+- `debug-distributed-hang` when the problem is clearly a TP/PP/DP/EP hang
+- `llm-torch-profiler-analysis` when the issue is already narrowed to a
+  compute-side path
+
+Three examples are included:
+
+- TTFT spike with low queue time
+- replay-first CUDA crash flow
+- request-shaped distributed hang flow
+
+## Output Contract
+
+Return:
+
+- problem class
+- what was checked
+- strongest signal so far
+- current best guess
+- what was ruled out
+- next step
+- production risk
+
+## When To Use It
+
+- `/health` or `/health_generate` is unhealthy
+- latency or throughput regressed under serving load
+- queue size grows while health still looks green
+- one request class times out or hangs
+- the server crashes only after some requests
+- outputs changed after a deploy, topology change, or weight switch
+- one older commit is known-good and a newer commit is known-bad
+
+## Workflow
+
+### 1. Collect a baseline bundle
+
+If a live server is reachable, collect a read-only bundle before anything more
+intrusive:
+
+```bash
+python3 scripts/incident_artifact_tool.py collect-bundle \
+  --base-url http://127.0.0.1:30000 \
+  --outdir /tmp/incident_bundle
+
+python3 scripts/incident_artifact_tool.py summarize-bundle \
+  /tmp/incident_bundle
+```
+
+If the server is protected:
+
+```bash
+python3 scripts/incident_artifact_tool.py collect-bundle \
+  --base-url http://127.0.0.1:30000 \
+  --token "$SGLANG_BEARER_TOKEN" \
+  --outdir /tmp/incident_bundle
+```
+
+The bundle script collects:
+
+- `/health`
+- `/health_generate`
+- `/model_info`
+- `/server_info`
+- `/v1/loads?include=all`
+- `/v1/loads?include=core,queues,disagg,spec`
+- `/metrics`
+- `/hicache/storage-backend` on a best-effort basis
+
+Use the summary for a quick read on:
+
+- health vs. active health state
+- topology and runtime flags
+- point-in-time queue and token usage
+- TTFT / E2E / queue-time heuristics from Prometheus metrics
+
+If the summary says the bundle was captured while the server was idle, recollect
+it during traffic or move quickly to dump plus replay.
+
+If no live server is reachable, start from the best dump or log already available:
+
+- crash dump
+- request dump
+- logs
+- CUDA coredump
+- OTel trace
+- torch profile
+
+### 2. Save the failing request
+
+Read [references/decision-tree.md](references/decision-tree.md) only if the
+problem class is still unclear:
+
+- server down or unhealthy
+- latency or throughput regression
+- wrong output or behavior regression
+- intermittent timeout or hang
+
+Then preserve the request payload that actually triggers the problem:
+
+- crash path: use `--crash-dump-folder`
+- non-crash path: enable request dump or save the exact trigger request
+
+Do not jump straight from a live symptom to low-level debugging without first
+saving something you can replay.
+
+### 3. Replay on a clean target
+
+Read [references/endpoints-and-signals.md](references/endpoints-and-signals.md)
+when you need help reading the baseline bundle or the replay target.
+
+Read [references/replay-trace-profile.md](references/replay-trace-profile.md)
+when you need the replay, trace, profile, or bisect paths.
+
+Standard order:
+
+1. collect baseline bundle
+2. capture request dump or crash dump
+3. restart a clean debug target if needed
+4. replay the same issue
+5. collect replay-time logs and dumps
+
+### 4. Only go deeper after replay
+
+#### Replay
+
+Use replay when:
+
+- a crash dump exists
+- a request dump exists
+- the problem depends on request shape or workload mix
+
+If a crash dump exists, summarize it first:
+
+```bash
+python3 scripts/incident_artifact_tool.py summarize-dump \
+  --input-file /path/to/crash_dump.pkl
+```
+
+Then replay:
+
+```bash
+python3 /path/to/sglang/scripts/playground/replay_request_dump.py \
+  --input-file /path/to/crash_dump.pkl \
+  --host 127.0.0.1 \
+  --port 30000 \
+  --parallel 128
+```
+
+If `safe_pickle_load` blocks a locally captured trusted dump, use:
+
+```bash
+python3 scripts/replay_trusted_request_dump.py \
+  --input-file /path/to/request_dump.pkl \
+  --host 127.0.0.1 \
+  --port 30000 \
+  --parallel 1
+```
+
+If replay indicates a CUDA crash path, restart the same build with coredumps
+enabled before reproducing again:
+
+```bash
+SGLANG_CUDA_COREDUMP=1 \
+SGLANG_CUDA_COREDUMP_DIR=/tmp/sglang_cuda_coredumps \
+python -m sglang.launch_server \
+  --model-path ... \
+  --crash-dump-folder /tmp/sglang_crash_dump \
+  ...
+```
+
+Then inspect the generated coredump:
+
+```bash
+cuda-gdb "$(which python3)" \
+  -ex "target cudacore /tmp/sglang_cuda_coredumps/cuda_coredump_<host>.<pid>.<ts>"
+```
+
+For a replay-first crash example, read
+[references/case-studies.md](references/case-studies.md).
+
+#### OTel trace
+
+Use tracing when:
+
+- request-stage timing is unclear
+- router vs. worker attribution is unclear
+- PD prefill/decode transfer may be implicated
+
+If tracing was enabled at startup, you can change the level without restart:
+
+```bash
+curl "http://127.0.0.1:30000/set_trace_level?level=1"
+curl "http://127.0.0.1:30000/set_trace_level?level=2"
+```
+
+#### Torch profile
+
+Use profiling when:
+
+- the issue is already narrowed to compute-side ownership
+- replay already reproduces the problem
+- metrics and loads do not explain the regression
+
+At that point, switch to `llm-torch-profiler-analysis`. Do not duplicate
+its profiling workflow here.
+
+For a low-noise latency example, read
+[references/case-studies.md](references/case-studies.md).
+
+#### Distributed hang
+
+If this looks like a collective stall, save the failing request, replay it on a
+clean target, collect the replay-time bundle and stacks, then switch to
+`debug-distributed-hang`.
+
+For an example of that flow, read
+[references/case-studies.md](references/case-studies.md).
+
+#### Regression between two commits
+
+If one commit is known-good and another is known-bad, build a deterministic
+harness before doing deeper manual debugging:
+
+1. choose a stable reproducer: request replay, benchmark command, or correctness check
+2. make the harness return `0` on good behavior and non-zero on bad behavior
+3. run `git bisect start <bad> <good>`
+4. run `git bisect run <harness>`
+5. return here only after a candidate commit is isolated
+
+Prefer replay-backed bisect when the regression depends on request shape or
+long-running serving state.
+
+### 6. Switch tools when the boundary is clear
+
+Switch tools once the fault class is clear:
+
+- `llm-torch-profiler-analysis` for kernel and overlap attribution
+- `debug-distributed-hang` for collective or rank-divergence hangs
+- `debug-cuda-crash` for CUDA crash reproduction and kernel API logging
+
+Do not switch tools before collecting the first bundle unless the user already has
+decisive logs or dumps.
+
+## References
+
+Load only what the current step needs:
+
+- [references/decision-tree.md](references/decision-tree.md)
+  - problem classes, tool switch points, return shape
+- [references/endpoints-and-signals.md](references/endpoints-and-signals.md)
+  - endpoint behavior, auth notes, field reading
+- [references/replay-trace-profile.md](references/replay-trace-profile.md)
+  - request dump, crash dump, replay, trace, profiler step, bisect
+- [references/case-studies.md](references/case-studies.md)
+  - compact examples for replay-first CUDA crash, latency, and distributed-hang triage
+
+## Scripts
+
+- [scripts/incident_artifact_tool.py](scripts/incident_artifact_tool.py)
+  - collect a read-only live bundle
+  - summarize a collected bundle into a compact debug note
+  - summarize a trusted request dump or crash dump before replay
+- [scripts/replay_trusted_request_dump.py](scripts/replay_trusted_request_dump.py)
+  - replay a trusted request dump when `safe_pickle_load` blocks stock replay
+
+If a live bundle was collected, include its path.
+
+If replay, trace, or profiling was chosen, say why bundle plus dump were not enough.

.claude/skills/sglang-prod-incident-triage/references/case-studies.md

@@ -0,0 +1,81 @@
+# Case Studies
+
+Use these examples only after the live bundle and request dump point toward the
+same class of failure. They are patterns for how to reason from replayable
+evidence, not recipes to copy blindly.
+
+## CUDA Crash: Upstream Top-K Corruption, Downstream MoE OOB
+
+Use when a replayed CUDA crash lands in a MoE align or shared-memory kernel but
+the suspicious data was produced by an earlier routing kernel.
+
+Shape that made the original case useful:
+
+- model family: Qwen3 MoE
+- visible crash: `moe_align_block_size_kernel`
+- likely producer: `topkGatingSoftmax` / MoE top-k routing
+- evidence path: crash dump -> replay -> CUDA coredump -> walk one kernel
+  upstream from the visible fault
+
+Triage loop:
+
+```text
+summarize crash dump
+  -> replay the exact request
+  -> enable CUDA coredump on the replay target
+  -> identify the failing kernel
+  -> inspect the immediately preceding producer kernel and tensors
+```
+
+Key lesson: a consumer kernel can be the first one to fault even when the bad
+index was produced earlier. Preserve the request shape before changing prompts.
+
+## Latency: TTFT Spike With Low Queue Time
+
+Use when `/health` and `/health_generate` are green, queue depth is low, but TTFT
+is still high.
+
+Signals from the original case:
+
+- `waiting=0`
+- average queue time was tiny
+- TTFT was high
+- scheduler stage timing pointed to prefill forward time
+
+Triage loop:
+
+```text
+collect live bundle
+  -> save the slow request
+  -> replay the same request on a clean target
+  -> profile only after replay reproduces compute-side ownership
+```
+
+Key lesson: rule out queue pressure with `/v1/loads`, `/metrics`, and stage
+timing before opening a profiler trace.
+
+## Distributed Hang: Request-Shaped TP Collective Mismatch
+
+Use when one request hangs, ranks stop making progress differently, and the
+failure looks like a generic serving stall until replay isolates it.
+
+Shape that made the original case useful:
+
+- a prompt tokenized to a specific extend length
+- one TP rank skipped a logits `all_gather`
+- the peer rank still entered the real collective
+- the request never returned
+
+Triage loop:
+
+```text
+collect healthy bundle
+  -> save the trigger request
+  -> replay on a clean target
+  -> collect rank stacks and replay-time bundle
+  -> switch to debug-distributed-hang
+```
+
+Key lesson: once the symptom looks like rank divergence or a collective mismatch,
+do not keep profiling kernels. Preserve the replay and move to distributed-hang
+debugging.

.claude/skills/sglang-prod-incident-triage/references/decision-tree.md

@@ -0,0 +1,197 @@
+# SGLang First Checks
+
+Use this reference when the problem class is still unclear and you need a fast
+starting point.
+
+## Default Order
+
+1. classify the symptom
+2. collect the fastest useful signal
+3. save the failing request or dump
+4. replay before you profile
+
+Do not start with `torch.profiler` unless the issue is already clearly
+compute-side.
+
+If one commit is known-good and another is known-bad, turn the problem into a
+stable `git bisect run <harness>` first.
+
+## Problem Classes
+
+### Server down or unhealthy
+
+Check:
+
+- `/health`
+- `/health_generate`
+- `/server_info`
+- recent stderr/stdout
+- crash dump status if `--crash-dump-folder` is enabled
+
+Likely directions:
+
+- startup or weight-load failure
+- deadlock or blocked scheduler
+- CUDA crash or OOM
+- auth or routing mismatch
+
+### High latency or low throughput
+
+Check:
+
+- `/v1/loads?include=all`
+- `/metrics`
+- `/server_info`
+- the exact request shape or benchmark command
+
+Likely directions:
+
+- queueing or capacity pressure
+- cache hit rate collapse
+- PD or EP topology mismatch
+- speculative decoding disabled or ineffective
+- kernel or backend regression
+
+### Wrong output or behavior regression
+
+Check:
+
+- exact request and expected output
+- `/model_info`
+- `/server_info`
+- current weights or recent config change
+
+Likely directions:
+
+- wrong weights or wrong revision
+- chat template, parser, or tool config drift
+- multimodal preprocessing drift
+- quantization or kernel correctness bug
+
+### Timeout or hang
+
+Check:
+
+- `/health`
+- `/health_generate`
+- `/v1/loads?include=all`
+- request dumps if enabled
+- per-rank logs
+- OTel trace if already enabled
+
+Likely directions:
+
+- distributed divergence or collective hang
+- queue starvation or retraction storm
+- PD transfer stall
+- storage or HiCache backend stall
+
+## Quick Paths
+
+### TTFT spike
+
+Start with:
+
+- `/v1/loads?include=all`
+- `/metrics`
+- `/server_info`
+
+Watch for:
+
+- `num_waiting_reqs` growth
+- `token_usage` saturation
+- `cache_hit_rate` drop
+- PD queue buildup
+
+If queue pressure does not explain the slowdown, save the slow request and
+replay it.
+
+### Throughput collapse
+
+Start with:
+
+- `/v1/loads?include=all`
+- `/metrics`
+- benchmark reproduction if available
+
+Watch for:
+
+- low `gen_throughput`
+- queue growth
+- low cache hit rate
+- speculative metrics collapse
+- PD transfer or decode prealloc queues backing up
+
+### Crash after some requests
+
+Start with:
+
+- crash dump folder
+- stderr/stdout
+- request dump folder if available
+
+Then replay the crash dump or recent request dump.
+
+### Regression between two commits
+
+Start with:
+
+- known-good commit
+- known-bad commit
+- one stable pass/fail harness
+
+Best move:
+
+- `git bisect run <harness>`
+
+### One request class fails
+
+Start with:
+
+- exact request payload
+- request dump if available
+- smallest reproduction request
+
+Typical categories:
+
+- multimodal edge case
+- parser or structured output bug
+- model-specific kernel path
+- tool-call formatting issue
+
+## When To Switch Tools
+
+### Use replay when
+
+- a crash dump or request dump already exists
+- the issue depends on request shape or workload mix
+- you need one stable reproducer before going deeper
+
+### Use OTel trace when
+
+- request-stage timing is unclear
+- router vs. worker ownership is unclear
+- PD boundaries may be involved
+
+### Use torch profiler when
+
+- replay already reproduces the issue
+- queueing and routing are mostly ruled out
+- you need kernel-level attribution
+
+At that point, switch to `llm-torch-profiler-analysis`.
+
+### Use lower-level debug paths when
+
+- replay plus trace still leave ambiguity
+- the problem looks like a specific crash, hang, or correctness bug
+
+## What To Return
+
+- problem class
+- what was checked
+- strongest signal so far
+- current best guess
+- what was ruled out
+- next step
+- production risk