sgl-project

Prepare batches of merged SGLang pull requests for a release branch. It checks conflicts and previous attempts, starts the cherry-pick workflow, monitors completion, and reports created pull requests or failure reasons.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Prepare multiple merged PRs for a release branch
• Check conflicts before starting workflows
• Detect previous cherry-pick PRs

Analyze LLM serving traces to identify GPU time usage, opportunities to run operations concurrently, and known patterns for combining operations. Produce three consistent tables from saved traces or live captures on supported servers.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Identify GPU operations that dominate inference time
• Compare input processing and token generation
• Trace GPU operations back to Python code

Develop lightweight CUDA operations for SGLang using JIT, which compiles code on first use. A tensor-scaling example explains implementation conventions, input validation, memory access helpers, and when to choose this approach.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Add a lightweight CUDA operation to SGLang
• Choose between JIT and ahead-of-time compilation
• Validate tensor inputs in a kernel launcher

Collect and review SGLang startup logs to identify unnecessary output and trace its source. Get saved logs and specific cleanup proposals while preserving useful diagnostics, with runtime changes left for user approval.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Investigating repeated messages in startup logs
• Checking which startup warnings need attention
• Comparing startup logs across models and GPU counts

Measure denoising latency in SGLang diffusion models and identify slow operations with profiler traces. Compare performance and output quality on the same GPUs, check existing optimizations, and gather evidence before starting low-level kernel work.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Measuring diffusion denoising latency
• Ranking slow operations from profiler traces
• Checking whether existing optimizations address a bottleneck

Choose SGLang Diffusion settings for a specific model, GPU, and memory budget. Compare generation speed, memory use, and output quality to select command-line options, including CPU offloading, multiple GPUs, caching, and reduced precision.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Choosing generation settings for a specific model and GPU
• Reducing GPU memory use to fit a memory budget
• Comparing performance settings across multiple GPUs

Convert an existing SGLang cookbook page to shared templates driven by model configuration. Preserve its deployment commands and interactive choices, transfer eligible benchmark measurements, and check equivalence, links, and rendered behavior.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Migrating a legacy SGLang cookbook page
• Moving deployment choices into model configuration
• Transferring reproducible benchmark measurements

Investigate CUDA crashes in SGLang using logs and inputs saved before execution. Inspect tensor shapes, data types, and numerical anomalies, with guidance for separate process logs and additional CUDA debugging tools.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Investigating illegal memory access and device-side assertions
• Tracing NaN or Inf values through kernel calls
• Saving inputs before a CUDA crash

Attribute GPU time in one sglang-omni stage to specific lines of `sglang_omni/` python source. Drives the `llm-torch-profiler-analysis` backend with a mapping/formal trace pair plus steady-state gates, and carries the rules for reading its three tables on an omni stage. Use when a stage is slower than expected and the question is which python code is the bottleneck, not which kernel is hot.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Run all reference benchmarks under benchmarks/eval/ and refresh the reference-table cells in benchmark_*.py. Auto-detects host hardware (H200/H100/H800/...). For each row's matching hw entry: replace cells in place. For new hardware or a new workload tag: append a new row to the section's table. Local-Pipeline-Result tables are skipped. One slash command runs everything end-to-end and auto-commits.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Generate a bounded task plan (and the agent prompt for it) that runs the 5-layer profiling methodology in .claude/skills/model-profiling/METHODOLOGY.md against one model, stops for human confirmation before any GPU work starts, then delegates the actual run to a background agent, verifies its completion claim, and routes findings into a gitignored .profiling-runs/<model>/ directory — durably tracked as a sub-issue under the tracking issue named in SKILL.md's "Result tracking" section — plus the methodology doc.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Use when running throughput/latency (serving speed) benchmarks against an already-running model server. Uses sgl_jax.bench_serving to sweep batch-size/concurrency and measure throughput and latency, producing summary.csv and per-point bs_<N>/result.jsonl. Speed only — accuracy/eval belongs to model-accuracy-benchmark. Triggers on throughput, latency, bench_serving, batch-size sweep, tok/s, TTFT, TPOT — not accuracy/eval.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Use when evaluating a newly adapted model's accuracy/eval score. Runs evalscope eval (or the test/srt/run_eval.py fallback) against an OpenAI-compatible endpoint on datasets like gsm8k / aime, producing per-dataset reports and a score. Accuracy only — throughput/latency benchmarks belong to model-speed-benchmark. Triggers on accuracy, eval, evalscope, gsm8k, aime, run_eval — not throughput.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Use when you need to CAPTURE / PRODUCE XProf (jax.profiler) profiling artifacts from sgl-jax (SGLang on JAX/TPU) for later performance analysis: driving a live server (/start_profile, stage-separated prefill/decode), offline bench_one_batch --profile, or kernel microbench. This skill produces the trace files; it does not analyze them.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

RBG controller stress test - deploy kwok in existing cluster, run create/update/delete load tests with real controller pod, collect pprof profiling and logs, generate analysis report with recommendations

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Implements Manus-style file-based planning to organize and track progress on complex tasks. Creates task_plan.md, findings.md, and progress.md. Use when asked to plan out, break down, or organize a multi-step project, research task, or any work requiring >5 tool calls. Supports automatic session recovery after /clear.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Step-by-step tutorial for adding a new lightweight SYCL JIT kernel for Intel XPU to sgl-kernel-xpu's python/sgl_kernel/jit module (SYCL source + Python wrapper + tests + benchmark)

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Upgrade the vendored NeMo-Skills slice to a newer upstream commit. Use when the user asks to "bump vendored", "sync NeMo-Skills", "update vendored sha", "upgrade NS", or wants to pull a specific upstream fix. Inside the sgl-eval repo.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Add a new benchmark to sgl-eval by vendoring its NeMo-Skills dataset module and registering it in `_TABLE`. Use when the user asks to "add <benchmark>" / "support <benchmark>" / "register <benchmark>" inside the sgl-eval repo.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

Audit whether all score-deciding logic in sgl-eval is vendored from NeMo-Skills, or whether some has crept into SE code. Use when the user asks "are we vendoring enough", "review vendor coverage", "audit vendoring", "is our vendoring complete", or before a release.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.