768 skills · sorted by relevance
Set up large language model pretraining with TorchTitan and PyTorch. Configure training across multiple GPUs, combine four ways of splitting the workload, enable Float8 calculations on H100 hardware, and manage checkpoints for saving and resuming training.
Helps organize PyTorch model training with LightningModule and Trainer, including validation, logging, model saving, and early stopping. It also guides distributed training across multiple GPUs and settings for reduced-precision computation.
Operate NVIDIA Nsight Graphics from the command line to capture application frames and GPU activity. Summarize performance traces, analyze existing capture files through ngfx-replay, and check installed versions and compatibility before starting an investigation.
Measure scripts, running processes, or whole systems to find performance bottlenecks. Select suitable tools, add measurement code when needed, and report CPU, GPU, memory, and communication findings with prioritized recommendations and a record of file changes.
Guide large language model training with OpenRLHF using human preference feedback. Covers training method selection, distributed GPU setup, reward model training, and memory tuning, with examples for PPO, GRPO, and DPO.
Use this skill when the user is doing hands-on DOCA GPUNetIO programming — wiring a CUDA kernel on an NVIDIA GPU to a doca-eth queue via doca_gpu_eth_rxq / doca_gpu_eth_txq, standing up the per-CUDA-device doca_gpu context, designing the persistent CUDA kernel that drains the GPU-visible queue, running the dual capability check (DOCA cap-query plus cudaGetDeviceProperties), registering cudaMalloc pools via doca_buf_arr_create_*, or debugging DOCA_ERROR_* returns from the GPUNetIO API. Trigger even when the user does not explicitly mention "DOCA GPUNetIO" or "persistent kernel" — typical implicit phrasings include "CUDA kernel reading packets directly from the NIC", "GPU-initiated networking on BlueField", "DOCA_ERROR_DRIVER on doca_gpu_create", "nvidia_peermem not loaded", "kernel-per-packet is too slow", or "which GPU supports GPU-side packet I/O". Refuse and route elsewhere for general CUDA programming, DOCA Ethernet queue bring-up, DOCA DPA, or DOCA install — those belong to other skills.
The English summary is being prepared. Showing the original description.
Use this skill when the user is building, running, or interpreting the doca/tools/gpunetio_ib_write_bw client+server benchmark — a CUDA kernel on the client posts RDMA WRITE work requests through the doca-gpunetio device-side surface to measure sustained GPU-driven WRITE bandwidth on a GPU+IB-device pair. Trigger even when the user does not explicitly mention "doca-gpunetio-ib-write-bw" or "GPUNetIO" — typical implicit phrasings include "measure WRITE BW when the GPU posts the WRs", "BW swings between runs on the same flags", "is the NIC saturated or am I CPU-bound on the CUDA kernel", "meson compile fails for the GPUNetIO bw tool", "nvidia_peermem isn't picking up my GPU buffer", or "GPU-initiated WRITE throughput vs CPU-initiated perftest". Refuse and route elsewhere for general doca-gpunetio library work, DOCA install, the GPU-initiated WRITE latency analog, the CPU-initiated upstream perftest, or application-level end-to-end throughput — those belong to other skills.
The English summary is being prepared. Showing the original description.
Helps you build similarity search with FAISS to find nearby vectors, numerical representations of data. Covers index selection, saving and loading, GPU use, and tuning search speed, accuracy, and memory consumption for large datasets.
Build Python workloads on Modal for cloud execution, including GPU model training and inference, web APIs, scheduled jobs, and parallel batch processing. Configure container environments, persistent storage, and automatic scaling through the Modal SDK.
Run language models locally with llama.cpp on CPUs, Apple Silicon, or AMD and Intel GPUs. The skill guides model setup, memory-saving GGUF format choices, interactive chat, and serving models through an OpenAI-compatible API.
Deploy AI models on GKE with suitable GPUs or TPUs and model servers. Generate deployment configuration, adjust automatic scaling, and tune inference—the process of producing model responses—for memory use, throughput, and response time.
Launch machine learning training jobs on local or remote GPUs. The skill checks available GPU memory, synchronizes code, saves logs, and verifies startup, with optional experiment tracking and support for Vast.ai and Modal.
Deploy language model APIs and run batch generation with vLLM. Configure GPU memory use, quantization—lower precision model storage—and distribution across GPUs, then test response times and monitor how many requests the server handles.
Add PyTorch FSDP2 to training scripts so model state can be split across GPUs. Covers distributed initialization, mixed precision, CPU offloading, and checkpoint saving and loading for models that exceed a single GPU's memory.
Train or fine-tune language and vision models on Hugging Face Jobs without a local GPU. Covers dataset checks, hardware and cost planning, training monitoring, Hub storage, and GGUF conversion for local use.
Use this skill when the user is measuring GPU-kernel-initiated RDMA WRITE latency through doca-gpunetio — building and running the `gpunetio_ib_write_lat` client + server pair under `doca/tools/gpunetio_ib_write_lat/`, checking GPU-NIC pairing, reading the half-iter / full-iter / CUDA-side usec columns, characterizing median / p99 / jitter for a real-time control loop, picking GPUNetIO vs GPI vs CPU-initiated `perftest`, or weighing the latency-vs-batching trade-off. Trigger even without 'GPUNetIO' or 'ib_write_lat': 'GPU kernel RDMA latency benchmark', 'how fast can a CUDA kernel post a WRITE', 'p99 RDMA latency on H100 + ConnectX', 'kernel-launched WR tail latency', or 'compare GPU-init vs CPU-init perftest'. Route elsewhere for bandwidth runs (doca-gpunetio-ib-write-bw), the GPI surface (doca-gpi), library debugging (doca-gpunetio), or DOCA install.
The English summary is being prepared. Showing the original description.
Helps write and review PyTorch models, training loops, and data loading code. Covers repeatable experiments, saving training progress, and techniques for reducing GPU memory use or improving training speed, with practical code examples.
Clarify the availability and requirements of model serving on booked Itô GPUs. The supplied implementation does not support serving, so it reports that limitation and documents future booking verification, configuration review, and explicit confirmation requirements.
Check whether machine learning training is available for a completed Itô GPU booking. Training is currently unavailable; the skill explains this limitation and documents the requirements for a future confirmed handoff of a training plan.
Fine-tune large language models with LoRA and QLoRA, methods that train small additions to a base model while limiting GPU memory use. Configure training, save task-specific adapters, switch between them, and merge them for deployment.
Compress large language models to 4-bit weights to reduce GPU memory use and accelerate inference. Covers loading compressed models, configuring quantization, and using Transformers or vLLM, with guidance on hardware requirements and backend choices.
Design and review desktop interfaces built with GPUI Kit. Applies rules for layouts, colors, components, interaction states, and interface wording, with checks for keyboard access, focus behavior, and accessibility before finishing the work.
Run TVM's TIRx compiler tests alongside tirx-kernels import and correctness checks. It verifies the local build, selects GPUs, runs parallel tests, saves full logs, and guides failure investigation and result reporting.
Evaluate NVIDIA Cosmos Policy for robot manipulation in LIBERO and RoboCasa simulations. Follow GPU environment setup, configure rendering without a display, run small checks before larger benchmarks, inspect success rates, and profile inference latency.