oxfrug
/

chronos-2-int8-torchao

Time Series Forecasting

8-bit precision

Model card Files Files and versions

oxfrug commited on 23 days ago

Commit

f288c67

·

verified ·

1 Parent(s): 8db7cbc

Update README.md

Files changed (1) hide show

README.md +2 -2

README.md CHANGED Viewed

@@ -134,9 +134,9 @@ Half-precision (FP16 / BF16) did **not** help this 120M model on a 3090.
 ## Did INT8 change the forecasts?
-We did **not** rerun GIFT-Eval. That bench measures Chronos-2 itself. This pack only compresses weights.
-What we did: take a few public series, hide the last 12–168 points, forecast them with the original model and with this INT8, compare.
 - **German electricity (hourly)** — INT8 stays close. 24h median error vs original +1.8%; 168h actually −3%. Correlation of the two forecasts ≈ 0.99.
 - **M4 hourly / daily / monthly** — same story on shape (high correlation), median error a bit worse (about +9% to +15% MASE on those short holds).

 ## Did INT8 change the forecasts?
+This pack only compresses weights.
+What i did: take a few public series, hide the last 12–168 points, forecast them with the original model and with this INT8, compare.
 - **German electricity (hourly)** — INT8 stays close. 24h median error vs original +1.8%; 168h actually −3%. Correlation of the two forecasts ≈ 0.99.
 - **M4 hourly / daily / monthly** — same story on shape (high correlation), median error a bit worse (about +9% to +15% MASE on those short holds).