Multi NODE Slurm Skill
Convert single-node scripts to multi-node Slurm sbatch jobs 與 debug common multi-node failures。
Skill
這裡收錄 repo 裡找得到的完整 SKILL.md、awesome-agent-skills 上游索引，以及 skill repo 的技能商店與本地落地版。頁面上的標題、用途、說明與 Skill 內容都會整理成台灣慣用正體中文；來源連結會保留下來。
沒有符合條件的 Skill。
NEMO RL E2E Testing Skill
External NeMo-RL 端到端 validation 工作流 用於 Megatron-Bridge model/provider changes, including downstream compatibility checks, external RL lifecycle behavior, Megatron poli..。
Parity Testing Skill
Structured framework 用於 verifying numerical parity of HF<->MCore weight conversions。
PERF Activation Recompute Skill
Validate 與 use selective 與 full activation recompute in Megatron Bridge to reduce GPU memory usage at the cost of extra compute。
PERF CPU Offloading Skill
Validate 與 use CPU offloading in Megatron Bridge, including layer-level activation offloading 與 fractional optimizer state offloading 搭配 HybridDeviceOptimizer。
PERF CUDA Graphs Skill
Validate 與 use CUDA graph capture in Megatron Bridge, including local full-iteration graphs 與 Transformer Engine scoped graphs 用於 attention, MLP, 與 MoE modules。
PERF Expert Parallel Overlap Skill
Validate 與 use MoE expert-parallel communication overlap in Megatron-Bridge, including overlap_moe_expert_parallel_comm, delay_wgrad_compute, 與 flex dispatcher backends such as..。
PERF Hierarchical Context Parallel 設計 Skill
Operational 指南 用於 enabling hierarchical context parallelism in Megatron-Bridge, including config knobs, code anchors, pitfalls, 與 verification。
PERF Megatron FSDP 設計 Skill
Operational 指南 用於 enabling Megatron FSDP in Megatron-Bridge, including config knobs, code anchors, pitfalls, 與 verification。
PERF Memory Tuning Skill
Techniques 用於 reducing peak GPU memory in Megatron Bridge — expandable segments, parallelism resizing, activation recompute, CPU offloading constraints, 與 common OOM fixes。
PERF MOE COMM Overlap Skill
協助處理 PERF MOE COMM Overlap 相關工作，並依原始 Skill 說明完成設定與執行。
PERF MOE Dispatcher Selection Skill
Choose the right MoE token dispatcher (`alltoall`, DeepEP, 或 HybridEP) 用於 the hardware, EP degree, 與 optimization stage。
PERF MOE Hardware Configs Skill
Representative MoE training playbooks by hardware platform 與 model family。
PERF MOE LONG Context 設計 Skill
Long-context MoE training guidance 用於 Megatron Bridge。
PERF MOE Optimization Workflow Skill
Systematic 工作流 用於 MoE training optimization in Megatron Bridge, based on the Megatron-Core MoE paper。
PERF MOE VLM Training 設計 Skill
Practical guidance 用於 training MoE VLMs in Megatron Bridge。
PERF Parallelism Strategies 設計 Skill
Operational 指南 用於 choosing 與 combining parallelism strategies in Megatron Bridge, including sizing rules, hardware topology mapping, 與 combined parallelism configuration。
PERF Sequence Packing 設計 Skill
Validate 與 use packed sequences 與 long-context training in Megatron-Bridge, distinguishing offline packed SFT 用於 LLMs from in-batch packing 用於 VLMs, 與 applying the right CP..。
PERF TP DP COMM Overlap 設計 Skill
Operational 指南 用於 enabling TP, DP, 與 PP communication overlap in Megatron-Bridge, including config knobs, code anchors, pitfalls, 與 verification。
Recipe Recommender Skill
Recommend 與 customize Megatron Bridge recipes 用於 a user's model, GPU count, 與 training goal。
Resiliency Skill
Resiliency features in Megatron Bridge including fault tolerance, straggler detection, in-process restart, preemption, 與 re-run state machine。
Testing Skill
Testing 參考資料 用於 Megatron Bridge — unit 與 functional test layout, tier semantics (L0/L1/L2/flaky), script conventions, running 測試 locally, adding/moving/disabling 測試,..。
VERL E2E Testing Skill
External verl 端到端 validation 工作流 用於 Megatron-Bridge model/provider changes。
Build AND Dependency Skill
Container-based dev environment setup 與 dependency management 用於 Megatron-LM。
BUMP BASE Image 設計 Skill
協助處理 BUMP BASE Image 設計 相關工作，並依原始 Skill 說明完成設定與執行。
CI/CD 雲端部署 Skill
CI/CD 參考資料 用於 Megatron-LM。
Create Issue 雲端部署 Skill
Investigate a failing GitHub Actions run 或 job 與 create a GitHub issue 用於 the failure。
Linting AND Formatting Skill
Linting 與 formatting 用於 Megatron-LM。
Nightly SYNC Skill
Domain knowledge 用於 the nightly main-to-dev sync 工作流。
Onboard Gb200 1node Tests Skill
Onboard 1-node GitHub MR functional 測試 用於 GB200 from existing mr-scoped 2-node 測試。
Respond TO Issue Skill
Research 與 draft a response to a GitHub issue 或 question from an external contributor。
RUN ON Slurm Skill
協助處理 RUN ON Slurm 相關工作，並依原始 Skill 說明完成設定與執行。
Split PR 設計 Skill
協助處理 Split PR 設計 相關工作，並依原始 Skill 說明完成設定與執行。
Testing Skill
測試 system 用於 Megatron-LM。
Update Golden Values 雲端部署 Skill
Refresh golden values from a GitHub Actions 工作流 run (failing-only 或 all jobs), score the change 搭配 average normalized relative differences, 與 produce a PR-ready summary。
Accessing Mlflow 資料庫 Skill
Query 與 browse evaluation results stored in MLflow。
Debug 雲端部署 Skill
協助處理 Debug 雲端部署 相關工作，並依原始 Skill 說明完成設定與執行。
Deployment 雲端部署 Skill
Serve a quantized 或 unquantized LLM checkpoint as an OpenAI-compatible API endpoint 使用 vLLM, SGLang, 或 TRT-LLM。
Evaluation Skill
Evaluates accuracy of quantized 或 unquantized LLMs 使用 NeMo Evaluator Launcher (NEL)。
Launching Evals Skill
Run, monitor, analyze, 與 debug LLM evaluations via nemo-evaluator-launcher。
Monitor 雲端部署 Skill
Monitor submitted jobs (PTQ, evaluation, 部署) on SLURM clusters。
PTQ Skill
協助處理 PTQ 相關工作，並依原始 Skill 說明完成設定與執行。
Release Cherry PICK Skill
Cherry-pick merged PRs labeled 用於 a release branch into that branch, then open a PR 與 apply the cherry-pick-done label。
BYOB Skill
建立 custom LLM evaluation benchmarks 使用 the BYOB decorator framework。
Accessing Mlflow 資料庫 Skill
Query 與 browse evaluation results stored in MLflow。
Launching Evals Skill
Run, monitor, analyze, 與 debug LLM evaluations via nemo-evaluator-launcher。
NEL Assistant Skill
Interactive config wizard 用於 NeMo Evaluator Launcher (NEL)。
ADD Benchmark 設計 Skill
> 指南 用於 adding a new benchmark 或 training environment to NeMo-Gym。
NEMO GYM Debugging Skill
>- Use when debugging a Nemo Gym run 或 reward profiling job。
NEMO GYM DOCS Skill
> Maintain the NeMo Gym Fern docs site — add, update, move, 或 remove pages under fern/。
NEMO GYM Pivot Datasets Skill
>- Use when creating, validating, 或 documenting Nemo Gym pivot datasets from rollout, trajectory, chat-completion, Responses API, 或 tool-call artifacts。
NEMO GYM Reward Profiling Skill
>- Use to help users get started 搭配 Nemo Gym reward profiling。
AUTO Research 測試 Skill
Autonomous NeMo-RL research agent 工作流 用於 directed hypothesis testing 與 open-ended discovery。
BREV Etiquette Skill
Brev instance operating guidance 用於 NeMo-RL agents working in /home/ubuntu/RL 搭配 limited workspace disk, a larger /ephemeral volume, 與 optional /home/ubuntu/RL/.env secrets。
Build AND Dependency Skill
建置 and dependency management 用於 NeMo-RL。
CI/CD 雲端部署 Skill
CI/CD 參考資料 用於 NeMo-RL。
Config Conventions Skill
Configuration conventions 用於 NeMo-RL。
Contributing Skill
Contribution conventions 用於 NeMo-RL。
Copyright 設計 Skill
NVIDIA copyright header requirements 用於 NeMo-RL。
DOCS Skill
檔案 conventions 用於 NeMo-RL。