Machine Learning

Launch, monitor and debug NeMo-RL training recipes on Kubernetes with nrl-k8s

Launch distributed Megatron-LM training jobs on SLURM with sbatch and torchrun

Export medical AI inference metadata, metrics and evidence packs to MLflow

Choose and configure FSDP2, Megatron FSDP and DDP for NeMo AutoModel training

Configure NeMo AutoModel launches for interactive, Slurm and SkyPilot runs

Onboard a new model architecture into NeMo AutoModel with registration and tests

Create and modify NeMo AutoModel training and eval recipes in YAML

Run Megatron-LM and Megatron Bridge training with mock or real data

Convert single-node training scripts into multi-node Slurm jobs and fix NCCL timeouts

Use selective or full activation recompute in Megatron Bridge to save GPU memory

Offload activations and optimizer state to CPU in Megatron Bridge training

Capture CUDA graphs for attention, MLP and MoE modules in Megatron Bridge

Overlap MoE expert-parallel communication in Megatron Bridge with DeepEP or HybridEP

Enable hierarchical context parallelism for long sequences in Megatron Bridge

Enable Megatron FSDP in Megatron Bridge with config knobs and checks

Cut peak GPU memory in Megatron Bridge and fix common out-of-memory errors

Overlap MoE dispatch and combine communication in Megatron Bridge

Pick an MoE token dispatcher (alltoall, DeepEP or HybridEP) for your hardware

Get starting MoE training configs for each GPU type in Megatron Bridge

Size context parallelism and recompute for long-context MoE training

Run an evidence-gated MoE performance tuning workflow in Megatron Bridge

Train MoE vision-language models with FSDP or 3D parallelism in Megatron Bridge

Choose and combine tensor, data and pipeline parallelism in Megatron Bridge

Pack sequences for efficient long-context training in Megatron Bridge

Enable TP, DP and PP communication overlap in Megatron Bridge training

Recommend and resize Megatron Bridge recipes for your model and GPU count

Add fault tolerance, straggler detection and restarts to Megatron Bridge training

Run autonomous NeMo-RL experiments from hypothesis to baseline and analysis

Plan and chain Nemotron curation, SFT, RL alignment and checkpoint conversion steps

Finetune NV-Generate brain MRI models on your own NIfTI datalists

Finetune the NV-Generate MAISI VAE on CT and MRI NIfTI datalists

Finetune NV-Segment-CT VISTA3D with softmax on CT image and label sets

Optimize an existing NVFLARE federated job with agent-assisted Auto-FL

Generate a final report and progress plot for an NVFLARE Auto-FL campaign

Convert Hugging Face Trainer or SFTTrainer code into an NVFLARE federated job

Convert PyTorch Lightning training code into an NVFLARE federated job

Convert plain PyTorch training code into an NVFLARE federated job

Diagnose failed, stalled or timed-out NVFLARE jobs from bounded evidence

Route NVFLARE questions and conversions to the right converter skill

Shared NVFLARE conversion references and templates used by other skills

Generate defect images for PCB, metal and glass inspection with Cosmos AnomalyGen

Run video data augmentation and auto-labeling workflows on OSMO end to end

Find the right PhysicsNeMo model, datapipe or example for a scientific ML task

Add ShardTensor domain parallelism to PhysicsNeMo training and inference scripts

Run root-cause analysis on Visual ChangeNet runs using image evidence

Compute per-class precision, recall and AP for object detection results

Find weak object detection images from per-class recall, precision and AP50

Rank the weakest Visual ChangeNet samples and pick the best decision threshold

Define and validate JSON schemas for TAO spec bundles, job records and results