DeepSpeed
44216655 - Default gradient_clipping to 1.0 (#8068)

Commit
9 days ago
Default gradient_clipping to 1.0 (#8068) ## Summary - Change `GRADIENT_CLIPPING_DEFAULT` from `0.` (disabled) to `1.0`. ## Motivation With the old default, configs that omit `gradient_clipping` run unclipped. Most RL/LLM training (and the FSDP2 reference) clip at `1.0`; this avoids silently-unclipped runs. Isolated into its own PR since it is a default behavior change. ## Test plan - [ ] Init without `gradient_clipping` -> effective clip norm is `1.0`. - [ ] Explicit `gradient_clipping: 0.0` still disables clipping (override respected). Made with [Cursor](https://cursor.com) --------- Signed-off-by: Olatunji Ruwase <tunji.ruwase@snowflake.com> Co-authored-by: Cursor <cursoragent@cursor.com>
Parents
Loading