Professional LLM Training Telemetry & VRAM Calculator

Professional LLM Training Telemetry

Industry-Standard Formulas · Client-Side Processing · Click log entries to override with Ground Truth

Session

1. Live Tracker & Duration

Input Mode

Dataset Size

Target Epochs

Number of GPUs

Micro Batch

Grad Accumulation

Current Speed (s/it)

Steps Completed

Total Steps

—

Global Batch: —

Remaining

—

Estimated ETA

—

Total Run Time: —

2. Training Log Import HuggingFace trainer_state.json

📂

Drop trainer_state.json here

or adapter_config.json to auto-fill model settings

3. VRAM Estimator (Per GPU) Industry-Standard Formulas

Model Params (B)

Seq Length

Hidden Dim

Num Layers

Attn Heads

Method

Distributed

Optimizer

LoRA Rank

Target Modules

Hardware

Grad Checkpointing

Flash Attention 2

Mixed-Prec Master Copy

Estimated VRAM / GPU—

Limit: — GB

Weights:—

Gradients:—

Optimizer States:—

Activations:—

Master Weights:—

CUDA Overhead:~1.5 GB

4. Interactive Console (Simulation)

Ready...

5. Calibration Audit DB

6. Throughput & Efficiency MFU = (tokens/s × 6N) / Peak TFLOPS

Tokens / Second

—

Samples / Second

—

Model FLOPs Utilization

—

7. Cost & Data Estimator Chinchilla: optimal tokens ≈ 20 × params

Cost per GPU-hour ($)

Total GPU-Hours

—

Estimated Cost

—

Total Tokens

—

8. Learning Rate Schedule Preview

Peak LR

Min LR

Warmup Steps

Schedule

9. Validation & Test Suite Client-Side

Click "Run All Tests" to validate calculations.