THUDM/slime

[FSDP] Try to optimize context parallelism

Open

#1062 aperta il 8 dic 2025

Vedi su GitHub
 (6 commenti) (0 reazioni) (0 assegnatari)Python (873 fork)github user discovery
good first issue

Metriche repository

Star
 (6010 star)
Metriche merge PR
 (Merge medio 1g 11h) (31 PR mergiate in 30 g)

Descrizione

Currently the context parallelism's training time is like 2 times when we use cp=2 than cp=1, we should try to optimize this for long sequence, maybe reduce the all gather entropy and logp and calculate the local loss to reduce the numbers of communication

Guida contributor