-
Notifications
You must be signed in to change notification settings - Fork 1.2k
Pull requests: THUDM/slime
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
feat(nemotron_h): native support for the Nemotron-3 hybrid stack
#2378
opened Sep 11, 2026 by
shifangx
Loading…
feat(sft): a VL SFT rollout, so a -vl model config can be fine-tuned
#2377
opened Sep 11, 2026 by
shifangx
Loading…
fix(qwen3.5-vl): cp_size 1 needs no packed-sequence divisibility check
#2376
opened Sep 10, 2026 by
shifangx
Loading…
fix: place PPO terminal rewards on the owning CP rank
#2375
opened Sep 9, 2026 by
MrCapricornLiu
Loading…
fix(rollout): measure each engine's host instead of deriving it from its rank
#2373
opened Sep 9, 2026 by
kaysonyu
Contributor
Loading…
fix: make GenerateState.semaphore a lazy property to avoid event loop mismatch between train and eval
#2372
opened Sep 9, 2026 by
LucQueen
Loading…
fix(update-weight): release rollout engine lock when distributed transfer fails
#2371
opened Sep 9, 2026 by
wangxianyu0825
Loading…
[ROCm] Update and unify ROCm support for gfx942 and gfx950
#2369
opened Sep 8, 2026 by
LZ-QWQ
Contributor
Loading…
fix(args): reject --kl-coef with GRPO-family advantage estimators
#2368
opened Sep 8, 2026 by
QJSQJS
Loading…
fix: use group-level concurrency in fully-async rollout
#2367
opened Sep 7, 2026 by
JayPritchet
Loading…
fix(data): fill sample requests across multiple epochs
#2365
opened Sep 6, 2026 by
feng1201
Loading…
[docs] debug guide: tell a cold Triton compile from a colocate hang (en + zh)
#2364
opened Sep 6, 2026 by
David553
Loading…
[Performance] Offload multimodal rollout preprocessing from the asyncio event loop
#2363
opened Sep 6, 2026 by
Qi2yU
Loading…
4 tasks done
fix(weight-sync): compose engine waves, credits and timeline
#2362
opened Sep 5, 2026 by
0z5a
Loading…
fix(rollout): reject aborted samples before reward tensor
#2361
opened Sep 5, 2026 by
YeonwooSung
Loading…
fix(agent): realign from divergence using previous response length
#2360
opened Sep 5, 2026 by
YeonwooSung
Loading…
fix(args): reject --kl-coef for GRPO-style estimators
#2359
opened Sep 5, 2026 by
YeonwooSung
Loading…
1 task done
fix(sglang): tolerate missing CUDA graph memory tag
#2358
opened Sep 5, 2026 by
YeonwooSung
Loading…
1 task done
fix(rm): stop double-extracting boxed answers for boxed_math
#2357
opened Sep 5, 2026 by
YeonwooSung
Loading…
fix(whiten): scale-only path when shift_mean is false
#2356
opened Sep 5, 2026 by
YeonwooSung
Loading…
docs: fix flag help and fault-tolerance default values
#2355
opened Sep 5, 2026 by
YeonwooSung
Loading…
Previous Next
ProTip!
Follow long discussions with comments:>50.