"megatron/legacy/data/data_samplers.py" did not exist on "bcd605f8570ebeeb0436c115ebbfafc3c5a40ae5"
- 16 Feb, 2022 1 commit
-
-
Michael Figurnov authored
Previously, setting num_recycle=0 in a pretrained recycling model skipped creating some Linears and LayerNorms. This meant that the the offsets of these modules were not applied leading to a degraded performance in that case. PiperOrigin-RevId: 429075328 Change-Id: I9257f859521799f45e2deef3803c249311051225
-
- 02 Nov, 2021 1 commit
-
-
Augustin-Zidek authored
PiperOrigin-RevId: 407076987
-
- 30 Jul, 2021 1 commit
-
-
DeepMind authored
PiperOrigin-RevId: 387766802 Change-Id: Ic838537513fe1d5bf41facffffd44046e91c3fa3
-
- 27 Jul, 2021 1 commit
-
-
Tom Ward authored
PiperOrigin-RevId: 387085679 Change-Id: I73287fcd0a29e899543b64c596e306195a2f435e
-
- 22 Jul, 2021 1 commit
-
-
Saran Tunyasuvunakool authored
PiperOrigin-RevId: 386228948
-
- 15 Jul, 2021 1 commit
-
-
Augustin-Zidek authored
PiperOrigin-RevId: 384954738
-