-
Notifications
You must be signed in to change notification settings - Fork 793
Pull requests: THUDM/slime
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
fix: add eval-before-train to train_async.py (parity with train.py)
#1906
opened May 13, 2026 by
Taosheng-ty
Loading…
4 tasks done
feat: filter logits by loss_mask before log_probs/entropy computation
#1905
opened May 13, 2026 by
Taosheng-ty
Loading…
5 of 6 tasks
fix: preserve fused 3D expert tensors for Qwen3.5 MoE in torch_dist→H…
#1904
opened May 12, 2026 by
rouchenzi
Loading…
fix: restore actor weights after loading OPD teacher checkpoint
#1903
opened May 12, 2026 by
canlin03
Loading…
Neutralize zero-advantage samples to skip wasted forward compute
#1901
opened May 11, 2026 by
nanjiangwill
Collaborator
Loading…
fix: align correct-sample rewards with DP-local lengths
#1900
opened May 10, 2026 by
miamia0
Loading…
fix: add fallback for --save-hf when Megatron-Bridge lacks model support
#1881
opened Apr 30, 2026 by
WangHong-yang
Contributor
Loading…
3 tasks done
feat(profile): safer torch.profiler defaults + per-grad-step capture
#1879
opened Apr 29, 2026 by
leofan-lab
Contributor
Loading…
Add Megatron-Bridge LoRA support for GRPO actor training
#1865
opened Apr 26, 2026 by
taivu1998
Loading…
fix: guard DP-imbalance empty micro-batches under dynamic batching
#1860
opened Apr 24, 2026 by
leofan-lab
Contributor
Loading…
fix: rebind asyncio Semaphore and HTTP client on event-loop change
#1858
opened Apr 24, 2026 by
leofan-lab
Contributor
Loading…
fix: make Megatron one-shot train() assumptions idempotent across slime rollouts
#1857
opened Apr 24, 2026 by
leofan-lab
Contributor
Loading…
feat(gemma4): add Gemma4 26B-A4B MoE and 31B dense support
#1855
opened Apr 24, 2026 by
leofan-lab
Contributor
Loading…
Fix double prepare_grads / loss-scaler-double-update in train_one_step
#1842
opened Apr 17, 2026 by
jthomy
Loading…
fix(sft): enable max-length filtering for messages datasets
#1841
opened Apr 17, 2026 by
none0663
Contributor
Loading…
3 tasks done
Fix missing activation checkpointing (recompute) parameters in bridge mode
#1833
opened Apr 14, 2026 by
XJL010622
Loading…
[build] Add A100 support: patch set, offline-friendly conda build, and examples
#1832
opened Apr 14, 2026 by
jason9693
Loading…
fix(gemma3): use GeGLU activation instead of SwiGLU
#1825
opened Apr 10, 2026 by
leofan-lab
Contributor
Loading…
fix: auto-fallback to flash_attn for Qwen3.5 on pre-Hopper GPUs (head_dim=256)
#1808
opened Apr 6, 2026 by
dadiaomengmeimei
Loading…
Previous Next
ProTip!
no:milestone will show everything without a milestone.