Prepare and open VibeSys pull requests from local repo changes. Use when the user asks to create, open, publish, submit, or draft a PR for this repository, including tasks that need branch hygiene, targeted validation, PR intent reflection, PR template completion, commit/push, or GitHub pull request creation.
Create or update a VibeSys LLM model-serving example bundle for a new Hugging Face model, hardware target, and optimization workload. Use when the user wants to add an example under examples/model-serving, scaffold reference/accuracy_checker/benchmark inputs, adapt existing checker and HTTP benchmark scripts, or start an optimization run from a natural-language workload goal without adding new reusable framework code.
LLM and multimodal serving-system development. Activate when a task touches inference servers, latency / throughput / TTFT / TPOT, KV-cache, batching, attention kernels, CUDA graphs, speculative decoding, structured / grammar-constrained output, quantization, MoE serving, prefix caching, multi-modal (vision/speech/image/video) serving, porting a model to vLLM / SGLang / TensorRT-LLM, or serving on AWS Trainium (Neuron, torch-neuronx, NKI kernels).
This skill guides using the cli to generate NKI kernel profiles (NEFF + NTFF pairs) to analyze performance on Neuron hardware. Use when encountering "profile kernel", "capture execution trace", "generate NEFF", "get summary-json", or asking "how to profile NKI kernel".
Guide for writing and modifying NKI kernels. Covers new kernel creation from PyTorch/NumPy/natural language, editing existing kernels, adding shape/dtype support, refactoring tiling strategies, and implementing new features in NKI code. Use when user says "write NKI kernel", "convert PyTorch to NKI", "translate numpy to NKI", "create NKI kernel", "implement in NKI", "NKI version of", "how to write NKI kernel", "add support for <shape/dtype>", "modify this NKI kernel", "extend kernel to handle", "refactor tiling", "change tile size", "add batch dimension", "support variable length", "fix this kernel logic", "update kernel for gen4", or needs NKI API guidance for kernel changes.
This skill guides debugging NKI compilation errors on Neuron hardware. Use when encountering "compiler error on device", "debug NKI kernel", "test kernel on trn2/trn3", "neuronx-cc compilation failed", "validate kernel on hardware", "run kernel on trainium", or asking "how to debug NKI compilation errors on device".
Research NKI documentation for API lookups, tutorials, error codes, architecture, and optimization guides. Use when user asks "what does <API> do", "how to do <task>", "what is error <code>", "NKI <API> signature", "find NKI tutorial for <topic>", "look up <symbol>", or needs any NKI documentation reference.
Query and analyze NKI kernel profile data from neuron-explorer parquet files. Supports SQL queries via the neuron-explorer API and Python on parquet for advanced analysis. Works locally on trainium with NEFF/NTFF files on disk. Querying: start neuron-explorer, ingest profiles, run SQL against tables (Summary, Instruction, DmaPacket, DmaPacketAggregated, etc.), explore schemas. Use when user says "query profile", "run SQL on profile", "start neuron-explorer", or has NEFF+NTFF files and wants to query them. Analysis: compute performance bounds, identify bottleneck engines, measure gaps (idle time, inefficiency, excess traffic, transposes), and run investigations to localize inefficiencies to NKI source lines. Use when user says "analyze profile", "what's the bottleneck", "compute bounds", "why is my kernel slow", or wants profile-guided optimization guidance.