mlc-ai
Official@mlc-ai
High-performance distributed training optimization for MoE architectures using Nsight Systems profiling and SLURM-based cluster resource management.
Agent Skills by mlc-ai
Showing 8 vetted skills indexed across 1 GitHub repositories.
mlc-aianalyze-nsys-profile
Query Nsight Systems profiles to measure compute/communication overlap and inspect per-rank streams.
mlc-aicapture-nsys-profile
Capture Nsight Systems traces for PithTrain runs across parallelism configurations.
mlc-aiadd-memory-prints
Add memory profiling instrumentation to distributed MoE training workflows.
mlc-aisetup-benchmark-inputs
Prepare tokenized corpus shards and HF-DCP checkpoints for MoE benchmarks.
mlc-aiestimate-memory
Estimate peak GPU memory usage for DualPipeV MoE training runs.
mlc-aiadd-new-model
Create MoE model skeletons for PithTrain and wire them into training and inference workflows.
mlc-aivalidate-correctness
Run paired base and feature branch training validations and diff per-step metrics.
mlc-ailaunch-with-slurm
Dispatch SLURM workloads by wrapping srun inside the current allocation.
Frequently Asked Questions About mlc-ai
FAQPage SchemaWhat specific performance tasks can be performed using these capabilities?▼
These capabilities enable granular performance analysis of distributed training, including Nsight Systems trace capture, memory usage estimation for MoE architectures, and per-step metric validation. Users can measure compute-communication overlap and instrument memory profiling to optimize large-scale training runs.
Which engineering personas benefit from these training optimization skills?▼
These skills are designed for machine learning infrastructure engineers and performance researchers focused on distributed training. They are specifically intended for those managing large-scale MoE model development, cluster resource allocation, and deep-dive performance debugging on GPU-accelerated hardware.
What are the prerequisites for executing these training and profiling tasks?▼
Execution requires an existing SLURM-managed cluster environment and access to Nsight Systems for trace generation. Users must have their training environment configured with PithTrain and DualPipeV dependencies, along with tokenized corpus shards and HF-DCP checkpoints prepared for the target model architecture.