Skip to content

feat: --trim-loss-positions (compute teacher/logits/loss only at supe… #705

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
julyanghar wants to merge 2 commits into
base: main
Choose a base branch
from
Open

feat: --trim-loss-positions (compute teacher/logits/loss only at supe… #705

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
329003a
feat: trim_loss_positions — compute teacher/logits/loss only at super…
julyanghar Jul 20, 2026
54bb777
feat: SP-native trim_loss_positions under USP sequence parallelism
julyanghar Aug 1, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions examples/configs/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -233,6 +233,7 @@ Common fields:
| `training.resume_from` | `null` | Full-run checkpoint/run root: draft, optimizer/scheduler, counters, data position, and RNG. Mutually exclusive with `model.draft_checkpoint_path`. |
| `training.compact_teacher` | `false` | Exact lower-peak-memory teacher projection for offline text EAGLE3. |
| `training.compact_teacher_chunk_size` | `null` | Positive vocabulary chunk size; requires `compact_teacher: true`. |
| `training.trim_loss_positions` | `false` | Compute the teacher target_p, draft logits, and loss only at supervised positions (batch size 1, plain KL loss); mathematically equivalent to the full-length path. |
| `training.role` | `all` | Use `all` for local offline training; disaggregated entrypoints select `auto`, `producer`, or `consumer`. |
| `training.seed` | `42` | Run and per-rank RNG seed. |

Expand Down
Loading
Loading