skills.rest
sovr610
sovr610COMMUNITY
Explore all skills by this author →
WebsiteGitHub

Compute-Optimal Budget Planner (Chinchilla-style)

Plan compute-optimal training budgets for large language models with Chinchilla-style scaling.

Updated Feb 28, 2026
Search Tags:#planning#budget#transformer#gpu-utilization#FLOPs#tokens-per-param
One-click install
npx skills add https://github.com/sovr610/refffiy --skill compute-optimal-budget-planner-chinchilla-style
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: Compute-Optimal Budget Planner (Chinchilla-style)
Source: https://github.com/sovr610/refffiy/tree/main/brain-ai-dev/skills/compute-budget-planner
Command: npx skills add https://github.com/sovr610/refffiy --skill compute-optimal-budget-planner-chinchilla-style

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve?

Plan compute-optimal training budgets for large language model projects to predict tokens, FLOPs, wallclock time, and cost under a fixed compute budget.

Core Features & Use Cases

  • Three planner modes (validate run, compute required, solve optimal) to cover end-to-end budgeting scenarios.
  • Automatic GPU-spec lookup with MFU-based timing and cost estimation, plus fallback to user-supplied specs.
  • Outputs include budget.json, human-readable reports, and optional isoFLOPs visualization to understand the frontier.

Quick Start

Plan the budget for a 70B model with 1.4T tokens on 8 H100 GPUs and return the budget estimate.

Frequently Asked Questions about Compute-Optimal Budget Planner (Chinchilla-style)

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I plan a compute-optimal training budget for a large language model?▼

Plan a compute-optimal training budget by inputting model parameters and target tokens to predict FLOPs, wallclock time, and cost under a fixed compute budget using Chinchilla-style scaling.

What is Chinchilla-style scaling for LLM training?▼

Chinchilla-style scaling determines the compute-optimal ratio of training tokens to model parameters, ensuring you allocate your FLOPs budget efficiently rather than over-training or under-training a transformer model.

How do I estimate wallclock time and cost for training a transformer on H100 GPUs?▼

Estimate wallclock time and cost for H100 GPU training via automatic GPU-spec lookup with MFU-based timing, which calculates training duration and supports fallback to user-supplied specs.

Can I solve for the optimal token and parameter ratio given a fixed FLOPs budget?▼

Yes, the solve optimal mode calculates the ideal tokens-per-parameter ratio for a fixed FLOPs budget and optionally generates isoFLOPs visualizations to map the compute frontier.

Does the budget planner support validating an existing LLM training run configuration?▼

Yes, the validate run mode checks an existing configuration like a 70B model with 1.4T tokens on 8 H100 GPUs, returning a budget.json estimate and human-readable training report.

Related Skills

Explore More Software Engineering Skills→
kali20gakkikali20gakki
7

op-mfu-calculator

Compute MFU for matmul/GEMM operators from dimensions, time, and peak FLOPs.

Community
Intermediate
Monjyu1101Monjyu1101
1

lambda-labs-gpu-cloud

Launch and provision Lambda Labs GPU cloud instances with SSH access and persistent storage.

Community
Advanced
M4n5terM4n5ter
3

tinker-training-cost

Tokenize datasets and estimate Tinker fine-tuning training costs in USD.

Community
Advanced
hiyenwonghiyenwong
2

seizure-suppression-hub-stimulation

Identify epileptic brain network hubs and compute optimal electrical stimulation parameters.

Community
Advanced