Traction scoreGitHub stars can be faked, so popularity alone can be misleading. Traction Score looks for broader signs of recent attention, adoption, and active maintenance.
This repository provides specialized skills for LLM infrastructure engineers tasked with optimizing serving frameworks like vLLM, SGLang, and TensorRT-LLM. It transforms historical pull request data into actionable intelligence for selecting fast paths and kernel fusion ideas.
By leveraging documented optimization history, engineers can better evaluate regression risks and validate source paths when tuning model performance or managing GPU memory capacity.
WHO IT'S FOR
LLM infrastructure engineers
optimizing serving framework performance and kernels
MLOps / Deployment engineers
planning GPU memory and request capacity
AI performance architects
benchmarking serving frameworks against latency SLAs
Compatible AgentsThe repository documents support for these agents. The skills may also work with other agents that can load SKILL.md files, but they may need some setup or small changes.
Access to the installed skill's resolver and diagram index
Make sure model-architecture-diagram is installed in your agent before trying this task.
Use model-architecture-diagram to find the indexed public original architecture diagram for GLM-5. Return the resolver's original image Markdown and source attribution. If there is no match, say so.
Replace <startup-log-path> with an existing SGLang or vLLM startup log containing memory and token-capacity lines
GPU type or enough log information to infer it
Optional model config.json and nvidia-smi memory output for cross-validation
Request token length: 4096
Make sure llm-serving-capacity-planner is installed in your agent before trying this task.
Use llm-serving-capacity-planner to analyze my SGLang or vLLM startup log at <startup-log-path>. Explain the logged weight, KV cache, CUDA graph, and framework memory use, and estimate concurrent request capacity for 4096-token requests. Use final logged capacity and rank-local evidence; identify missing information instead of assuming an absent mem-fraction-static value.
Installed model-pr-optimization-history knowledge base; run its query commands from that directory
Framework: SGLang; model-family slug: qwen3-core
An active run-artifact directory for the short evidence note
Make sure model-pr-history-knowledge is installed in your agent before trying this task.
Use model-pr-history-knowledge to search the SGLang qwen3-core history for fused QK normalization and RoPE. Read the matching history and any dated opening notes, then summarize the relevant implementation paths, PR evidence, and validation or regression risks. Distinguish manually audited evidence from discovery inventories, and save a short evidence note in my active run artifacts before proposing any patch.
Claude
Codex
DeepSeek