Instructions to use AlexWortega/tinyvla with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LeRobot
How to use AlexWortega/tinyvla with LeRobot:
- Notebooks
- Google Colab
- Kaggle
fix: yaml 1.1 parses '2250.0e12' as string -> float() cast in MFU calc (crashed first log step)
Browse files
tinyvla_b200/scripts/train_fast.py
CHANGED
|
@@ -319,8 +319,9 @@ def main():
|
|
| 319 |
torch.cuda.synchronize()
|
| 320 |
el = time.time() - t0
|
| 321 |
sps = seen * world / el # глобально по всем рангам
|
| 322 |
-
|
| 323 |
-
|
|
|
|
| 324 |
eta = (steps - step) * eff_batch / sps / 3600
|
| 325 |
log(f"step {step}/{steps} loss {info['loss']:.4f} | {sps:.0f} samples/s "
|
| 326 |
f"| {step/el:.2f} it/s | MFU~{mfu:.0f}% | eta {eta:.2f} h "
|
|
|
|
| 319 |
torch.cuda.synchronize()
|
| 320 |
el = time.time() - t0
|
| 321 |
sps = seen * world / el # глобально по всем рангам
|
| 322 |
+
# float(): yaml 1.1 парсит "2250.0e12" без знака экспоненты как строку
|
| 323 |
+
gflops = float(cfg.get("gflops_per_sample", 360.0))
|
| 324 |
+
mfu = sps * gflops * 1e9 / float(cfg.get("peak_flops", 989e12)) * 100
|
| 325 |
eta = (steps - step) * eff_batch / sps / 3600
|
| 326 |
log(f"step {step}/{steps} loss {info['loss']:.4f} | {sps:.0f} samples/s "
|
| 327 |
f"| {step/el:.2f} it/s | MFU~{mfu:.0f}% | eta {eta:.2f} h "
|