estimate-memory

What problem does it solve?

This memory estimation skill provides precise peak GPU memory predictions for DualPipeV MoE training runs, enabling users to plan resources and avoid OOMs.

Core Features & Use Cases

• Analyze natural-language training configurations to produce exact CLI invocations for memory estimation.
• Supports common configurations (pp/ep/cp, total GPUs, batch sizes, sequence length, and GPU memory) and yields a per-component memory breakdown and a pass/fail verdict.
• Use case: before launching a training job on 32 GPUs, estimate peak memory to ensure sufficient headroom and adjust batch sizes or parallelism accordingly.

Quick Start

Run the memory estimator by providing the model config, pipeline, and batch parameters to generate a memory usage estimate.