30 个 · 按相关度排序
Configure large language model training with NVIDIA Megatron-Core across multiple GPUs. This skill helps choose how to split computation, set training parameters, monitor GPU efficiency, and address memory limits or unstable training.
中文简介正在准备中。当前显示英语简介。
Configure large language model training across multiple NVIDIA GPUs with Megatron-Core. The skill helps choose how to split models and workloads, set training parameters, monitor performance, and address memory shortages or low GPU utilization.
中文简介正在准备中。当前显示英语简介。
Build and train Mixture of Experts models, which activate selected specialist networks for each input. Use DeepSpeed or HuggingFace workflows to configure routing, balance workloads, distribute experts across GPUs, and optimize inference.
中文简介正在准备中。当前显示英语简介。
Train Mixture of Experts models, which activate selected specialist networks for each input. The skill guides architecture, input routing, workload balancing, and GPU distribution using DeepSpeed or HuggingFace, with examples for training configuration and inference optimization.
中文简介正在准备中。当前显示英语简介。
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.
中文简介正在准备中。当前显示英语简介。
Official NVIDIA-authored guidance for PhysicsNeMo ShardTensor domain parallelism — integrate domain parallelism into training/inference scripts (new or existing) with DDP or FSDP2, write and register shard patches to enable new layers/ops, and bootstrap multi-GPU correctness tests. Use when working with ShardTensor, scatter_tensor, domain parallelism, sequence/spatial sharding, ring attention, DeviceMesh + DDP/FSDP2 hybrid parallelism, or physicsnemo.domain_parallel. Do NOT use for generic PyTorch DDP/FSDP setup without domain parallelism, picking a PhysicsNeMo model or example (use physicsnemo-discover), or non-distributed training questions.
中文简介正在准备中。当前显示原文说明。
Train Mixture of Experts (MoE) models using DeepSpeed or HuggingFace. Use when training large-scale models with limited compute (5× cost reduction vs dense models), implementing sparse architectures like Mixtral 8x7B or DeepSeek-V3, or scaling model capacity without proportional compute increase. Covers MoE architectures, routing mechanisms, load balancing, expert parallelism, and inference optimization.
中文简介正在准备中。当前显示原文说明。
Deploy language models with vLLM through an OpenAI-compatible API or offline batch processing. Tune response speed, throughput, and GPU memory use with quantization, which reduces numerical precision, and execution across multiple GPUs.
中文简介正在准备中。当前显示英语简介。
Helps configure and troubleshoot DeepSpeed for training large models across multiple devices. Covers memory optimization, parallel training, and DeepNVMe storage transfers, with examples for tuning data movement and configuring models that divide work among specialized components.
中文简介正在准备中。当前显示英语简介。
Implement and troubleshoot large-model training with DeepSpeed across multiple GPUs. The skill covers memory optimization, parallel training, and numerical precision choices, alongside examples for storage transfers and models that divide work among specialized components.
中文简介正在准备中。当前显示英语简介。
Treinar modelos de Mixture of Experts (MoE) usando DeepSpeed ou HuggingFace. Use ao treinar modelos em larga escala com computação limitada (redução de 5× em custos vs modelos densos), implementar arquiteturas esparsas como Mixtral 8x7B ou DeepSeek-V3, ou escalar capacidade de modelo sem aumento proporcional de computação. Cobre arquiteturas MoE, mecanismos de roteamento, balanceamento de carga, paralelismo de especialistas e otimização de inferência.
中文简介正在准备中。当前显示原文说明。
Treina modelos de linguagem grandes (2B-462B de parâmetros) usando NVIDIA Megatron-Core com estratégias avançadas de paralelismo. Use ao treinar modelos >1B de parâmetros, precisar de máxima eficiência de GPU (47% MFU em H100), ou necessitar de paralelismo tensor/pipeline/sequência/contexto/expert. Framework pronto para produção usado em Nemotron, LLaMA, DeepSeek.
中文简介正在准备中。当前显示原文说明。
Configure large language model pretraining with torchtitan across multiple GPUs. Covers ways to split training work, launch jobs on computing clusters, enable Float8 numerical processing, and save or resume training through distributed checkpoints.
中文简介正在准备中。当前显示英语简介。
Configure large language model pretraining with torchtitan across multiple GPUs. Covers ways to split training work, launch jobs on computing clusters, enable Float8 numerical processing, and save or resume training through distributed checkpoints.
中文简介正在准备中。当前显示英语简介。
Guide for selecting and configuring distributed training strategies in NeMo AutoModel, including FSDP2, Megatron FSDP, DDP, and parallelism settings.
中文简介正在准备中。当前显示原文说明。
Use when training models across multiple GPUs or nodes, handling large models that don't fit in memory, or optimizing training throughput - covers DDP, FSDP, DeepSpeed ZeRO, model/data parallelism, and gradient checkpointingUse when ", " mentioned.
中文简介正在准备中。当前显示原文说明。
MANDATORY for Flink or Amazon Managed Service for Apache Flink (MSF) questions. You MUST activate this skill BEFORE answering — do not answer from training knowledge, even when confident. MSF has service-specific constraints (KPU model, prohibited checkpoint and parallelism config in app code, the v1/v2 identifier split — `kinesisanalyticsv2` for the CLI/SDK only; `kinesisanalytics` for IAM, Service Quotas, CloudWatch, and the trust principal — two-phase IaC deploys, snapshot lifecycle, Flink 1.x→2.x migration) that override generic Flink knowledge. Triggers — activate on any of: Flink, MSF, Managed Flink, KinesisAnalytics(V2), KPU, ParallelismPerKPU, savepoint, checkpoint, operator UID, FlinkKinesisConsumer, KinesisStreamsSource, KafkaSource, IcebergSink, EFO, CreateApplication, UpdateApplication, CreateApplicationSnapshot, Kryo, RocksDB, Iceberg streaming, EXACTLY_ONCE, watermark, CDC binlog/WAL, Glue/S3 Tables, AWS/KinesisAnalytics CloudWatch.
中文简介正在准备中。当前显示原文说明。
Write or rewrite prose so it sounds like a human wrote it — not a frontier model. Named in homage to Rabindranath Tagore, whose prose carried what frontier models reach for and miss: a point of view, specificity over abstraction, and restraint over puffery. Merges two complementary approaches: a 29-pattern catalog of AI tells (from humanizer) plus an 8-rule operating system with an 8-dimension scoring gate (extending stop-slop). Use when drafting, editing, or reviewing any prose: essays, posts, docs, reports, emails. Detects and removes inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary, passive voice, negative parallelisms, filler phrases, inanimate-verb constructions, narrator-from-a-distance voice, and metronomic rhythm. Adds back the things AI writing usually lacks: point of view, stakes, specificity, restraint, varied rhythm, and trust in the reader.
中文简介正在准备中。当前显示原文说明。
Molecular featurization hub with one consistent interface over 100+ featurizers. Fingerprints (ECFP/Morgan, MACCS, atom pair, topological torsion, Avalon, RDKit, ERG), RDKit and Mordred descriptor sets, pharmacophore and 3D shape descriptors, scaffold keys, and pretrained embeddings (ChemBERTa, ChemGPT, MolT5, GIN, Graphormer) through a common transformer API with caching and parallelism. Use this skill to convert SMILES into model-ready feature matrices for QSAR, virtual screening, and molecular ML, and to choose between featurizer families. Also trigger on molfeat, MoleculeTransformer, FPVecTransformer, PretrainedHFTransformer, molfeat model store, or featurizer selection.
中文简介正在准备中。当前显示原文说明。
分布式并行策略选型与实测（USP / CP 通信掩盖 / CFG / TP/RSP/PP 概览；含拓扑相关选型 （按实测拓扑分域条件化：域内 bulk vs 跨域 head-parallel 翻转）与 AlltoAllV 缺陷绕过）。在 model-auto-optimization 中承担 S3：优先 USP、结合拓扑带宽差异选 CP，少量 step + 多 rank 验证特性开启与掩盖（产物：并行方案 + 多 rank 证据）。当用户需要多卡并行 策略选择、**序列并行形态抉择**（纯 Ulysses vs 复合 AllGather-KV×Ulysses：按 GQA / 跨域带宽 / 形态 plumbing 条件化定胜负）、**并行 × 稀疏叠加**（seam 契约：先汇聚后稀疏、 窗口偏移、块对齐、per-head 掩码；含「稀疏看似生效实则未生效」判定）、通信掩盖调优 （含**掩盖率上限**：1-1/n 何时成立、c/f 决定的真实上限、没生效的排查）、**并行方案 差异归因**（阶段 Δ 分解 / 集合通信按 communicator 归属 / 4→8 卡线性度），或排查多卡 跑不动 / 通信暴露大 / 换卡组 / 端口 bind / HCCL 带宽验证问题时使用；**并收编原并行作用域诊断**： 改了 SP/CP/Ulysses/AllGather-KV 后**不报错但结果没变/性能没变**、或小规模能跑大规模崩 （如 Ascend EE1003 coreDim 超限）时，用本技能证明"改动到底有没有生效"（判别量逐层收窄 + 两侧对照 + 日志≠生效，见 `references/scope-effectiveness-check.md`）； 即使用户只说 "多卡跑不动""通信暴露大""为什么没达到 6/7 的掩盖率""CP 和 USP 该选哪个""CP 叠稀疏 怎么不生效"而未说并行，也应触发。特性档位/接口事实见 `docs/zh/features/parallelism.md` /`usp.md`（仓内真源），框架侧开启见 framework-integration； 本技能承载选型决策、monkey-patch 掩盖与多卡诊断实测。 由 dev-workflow 多卡场景触发，亦由 model-auto-optimization 的 S3 阶段触发。
中文简介正在准备中。当前显示原文说明。
Serve LLMs com alta throughput usando PagedAttention do vLLM e continuous batching. Use ao fazer deploy de APIs LLM em produção, otimizar latência/throughput de inferência, ou servir modelos com memória GPU limitada. Suporta endpoints compatíveis com OpenAI, quantização (GPTQ/AWQ/FP8) e tensor parallelism.
中文简介正在准备中。当前显示原文说明。
Serves LLMs with high throughput using vLLM's PagedAttention and continuous batching. Use when deploying production LLM APIs, optimizing inference latency/throughput, or serving models with limited GPU memory. Supports OpenAI-compatible endpoints, quantization (GPTQ/AWQ/FP8), and tensor parallelism.
中文简介正在准备中。当前显示原文说明。
Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP). Use when pretraining Llama 3.1, DeepSeek V3, or custom models at scale from 8 to 512+ GPUs with Float8, torch.compile, and distributed checkpointing.
中文简介正在准备中。当前显示原文说明。
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.
中文简介正在准备中。当前显示英语简介。