Product Tool
PRODUCT MODE: You are using the actual tune.fyi tool. Running simulations here consumes GPU hours.
Live Tracker & Duration
Total Steps
-
Global Batch: -Remaining
-
Estimated ETA
-
Total Run Time: -
2. Data & Telemetry Drop
📂
Drop trainer_state.json here
adapter_config.json to auto-fill model
settings
Run: -
Steps: -
Final Loss: -
Epochs: -
Log Entries: -
Base Model: -
LoRA Rank: -
VRAM Estimator (Per GPU) Standard Estimation Formulas
Estimated VRAM / GPU-
0%
Weights:-
Gradients:-
Optimizer States:-
Activations:-
Master Weights:-
CUDA Overhead:~1.5 GB
Interactive Console (Simulation)
Ready...
Welcome to tune.fyi
Follow these steps to estimate and calibrate your training run:
1. Configure Model
Define architecture in Section 1 or drop a HuggingFace trainer_state.json into Section 2 to auto-fill.
Hardware & Data
In Section 1, set your dataset size limit, sequence length, and target GPU limit to set your compute ceiling.
3. Optimize VRAM
Toggle Gradient Checkpointing, LoRA, and low-bit Optimizers in Section 3 until Total VRAM fits your GPU.
4. Calibrate (Optional)
Paste raw logs into Section 2 to inject live speeds, dynamically recalibrating all estimates to reality.
Look for the [i] icon on every module for detailed contextual help and formulas.
Help
Calibration Audit DB
Throughput & Efficiency MFU = (tokens/s x 6N) / Peak TFLOPS
Tokens / Second
-
Samples / Second
-
Model FLOPs Utilization
-
Cost & Data Estimator Chinchilla: optimal tokens ~ 20 x params
Total GPU-Hours
-
Estimated Cost
-
Total Tokens
-
Learning Rate Schedule Preview
9. Validation & Test Suite Client-Side
Click "Run All Tests" to validate
calculations.