inference skills

sgl-project/sglang

Skill Claude CodeCodex

Step-by-step tutorial for adding a new lightweight JIT CUDA kernel to sglang's jitkernel module.

sgl-project/sglang

Skill Claude CodeCodex

How SGLang's runtime configuration and process-global state are organized (RuntimeContext tiers, publish + namespace config bags, the pristine ServerArgs seed, override entry points, resource/stream/buffer leases, per-forward flags), the CI guardrails that enforce the design, and the idioms for developing and testing…

sgl-project/sglang

Skill Claude CodeCodex

Use when adding a new diffusion model or Diffusers pipeline to SGLang.

kvcache-ai/Mooncake

Skill Claude CodeCodex

Help users work with the Mooncake Python APIs for distributed storage and high-performance data transfer. Use when working with Mooncake Store (distributed KV cache), Transfer Engine (RDMA/TCP transfers), service setup (master, metadata server), PyTorch tensors in the Store, zero-copy/buffer management, batch…

kvcache-ai/Mooncake

Skill Claude CodeCodex

Run Mooncake pre-PR local validation through scripts/runcitest.sh. Use this skill whenever the user wants to validate a branch before opening or submitting a PR, run local CI, run ci test, check changes before PR, reproduce GitHub Actions locally, or force a full pre-submit verification. Trigger on phrases like "提交 PR…

kvcache-ai/Mooncake

Skill Claude CodeCodex

Automatically diagnose Mooncake deployment and runtime issues. Checks services (mooncakemaster, metadata server), RDMA devices, environment variables, connectivity, memory limits, object integrity, and analyzes logs for common error patterns. Use when Mooncake deployment fails, services won't start, connections fail…

hao-ai-lab/FastVideo

Skill Claude CodeCodex

Manual /add-model workflow for implementing a FastVideo model or first-class component port after add-model-01-prep has staged reference code and weights. Organizes the port into numbered phases with conversion rules, component policies, parity gates, and handoff checks.

hao-ai-lab/FastVideo

Skill Claude CodeCodex

Re-seed the HF performance-tracking baseline for an intentional runtime, dependency, environment-caused benchmark shift, or reviewed v2 calibration using one or more reviewed normalized performance JSONs. Use when performance CI fails because metrics such as latency, throughput, component time, or peak memory changed…

hao-ai-lab/FastVideo

Skill Claude CodeCodex

Seed HF reference artefacts for a single newly-added SSIM test (pixel .mp4 for runtexttovideosimilaritytest-style tests, or latent .pt for runtexttolatentsimilaritytest-style tests). Runs the test on Modal L40S, downloads the generated artefacts via modal volume get, pauses for the user to verify (visual eyeball for…

raullenchai/Rapid-MLX

Skill Claude CodeCodex

Autonomous performance optimization: research, PoC, benchmark, implement, review, PR.

xLLM-AI/xllm

Skill Claude CodeCodex

Add or update xLLM unit tests in the repository. Use when Codex needs to create a new C++/CUDA/NPU/MLU unit test, place a test under tests/, wire it into CMake with cctest, update an existing test target, choose platform gates, or validate test naming and dependencies against current xLLM test conventions.

xLLM-AI/xllm

Skill Claude CodeCodex

Review code changes for quality, security, performance, and correctness following project-specific standards. Use when reviewing pull requests, examining git diffs, or when the user asks for a code review. This skill should be used proactively — when the user asks for a review without specifying commits, automatically…

xLLM-AI/xllm

Skill Claude CodeCodex

Use when the user wants to add, modify, debug, or review an xLLM TileLang Ascend kernel or specialization, including Python kernel definitions, generated Ascend-C source, runtime wrapper dispatch, TileLang CMake wiring, and NPU tests.

ddalcu/mlx-serve

Skill Claude CodeCodex

Skill "bench" from ddalcu/mlx-serve, covering benchmarking and comparison traps (these cost real days).

ddalcu/mlx-serve

Skill Claude CodeCodex

Timings measured 2026-07-16 on the M4 Max 128 GB, AFTER the stopallengines port-wait fix (before it, everything below was 2.2× slower — see the gotcha in Benchmarking).

graphsignal/graphsignal-profiler

Skill Claude CodeCodex

Set up the Graphsignal Profiler for inference workloads — vLLM, SGLang, PyTorch, and dstack services. Use when the user wants GPU profiling, tracing, or monitoring for inference, asks about graphsignal-run or graphsignal.watch(), or asks about CUPTI / Prometheus / OTLP setup.

NVIDIA/nvcf

Skill Claude CodeCodex

Install, operate, and tear down self-hosted NVIDIA Cloud Functions (NVCF) deployments with nvcf-cli. Use for control-plane or compute-plane install, status checks, cluster registration, function deploy/invoke, task create/list/cancel/delete, API keys, admin tokens, JWKS rotation, failed-install diagnosis, and…

NVIDIA/nvcf

Skill Claude CodeCodex

Install the prerequisites the NVCA operator / compute plane needs before nvcf-nvca-install can succeed: the operator tool nvcf-cli (required by the compute-plane stack's register-cluster step), KAI Scheduler (for the KAIScheduler feature gate), and the SMB CSI driver (for the sharedStorage Samba sidecar PVCs). The two…

NVIDIA/TensorRT-Model-Connect

Skill Claude CodeCodex

Use when TensorRT output diverges from a model reference, model-first validation fails, generated text or media is wrong, or a family change introduces a numerical mismatch. Routes the investigation by model modality and escalates from the first divergent boundary to the smallest responsible family-owned operation.

NVIDIA/TensorRT-Model-Connect

Skill Claude CodeCodex

Use for documentation maintenance scans that keep the canonical website journey, repo-local skills, commands, API reference, architecture and design, extension guides, feature context, ADRs, and traceability status aligned with the current GitHub main branch. Covers the Source/Internal CI boundary and model-owned…

NVIDIA/TensorRT-Model-Connect

Skill Claude CodeCodex

Use when monitoring GitHub pull request CI, diagnosing failed checks, rebasing branches onto github/main, applying narrowly scoped fixes, and updating PRs until their latest checks are green or a human blocker is identified.

cloudrift-ai/emmy

Skill Claude CodeCodex

Onboard or periodically reverify and benchmark a Hugging Face model on an exact target GPU platform. Use when asked to add a model recipe, refresh a maintained recipe on a supplied GPU server, benchmark serving, create reproducible experiments and a durable results report, fully qualify and tune the model's Emmy…

cloudrift-ai/emmy

Skill Claude CodeCodex

Use this skill when the user asks to re-run an article's benchmarks, reproduce blog post numbers, validate that an article URL still holds, check whether the latest code still performs like a published post, or otherwise compare re-measured Emmy results with published results. It fetches the article, finds its…

cloudrift-ai/emmy

Skill Claude CodeCodex

Tune Emmy kernels for a Hugging Face model, traced IR, or golden YAML. Use when asked to tune a model or golden set, seed MCTS with model-proposed knob configurations, compare hybrid proposals against MCTS-only search, diagnose slow or failing kernels, refresh per-GPU goldens, or produce a per-kernel tuning findings…