THUDM/slime

View on GitHub

slime is an LLM post-training framework for RL Scaling.

Stars: 8.5k
Forks: 1.3k
Open beginner issues: 0
Indexed issues: 228
Avg merge: 5h 36m
Merged PRs (30d): 22

Dominant language: Python
License: Apache-2.0
Last GitHub push: Sep 19, 2026
Latest indexed: Sep 19, 2026
Contributing guide: Contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed

Beginner friendly Everything

Open All

228 open issues indexed

[Bug] Qwen-VL patch bypasses SGLang token-ID preservation and breaks multi-turn TITO Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 48/100

THUDM/slime#2388 · Opened Sep 16, 2026
[Question] Would an adaptive in-reward KL controller fit Slime's PPO/RLHF scope? Open

question

Difficulty 5/5 Over a week Newbie friendliness 25/100

THUDM/slime#2387 · Opened Sep 15, 2026
Verify evals on Papers with Code Open

Difficulty 2/5 1-3 hours Newbie friendliness 35/100

THUDM/slime#2386 · Opened Sep 14, 2026
[Bug] Broadcasting full rollout_routed_experts data to every Megatron training rank will cause host OOM Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 45/100

THUDM/slime#2384 · Opened Sep 14, 2026
[Bug] shipped dynamic-sampling filters crash on fan-out groups (list[list[Sample]]) from multi-turn agent rollouts Open

Difficulty 2/5 1-3 hours Newbie friendliness 84/100

THUDM/slime#2374 · Opened Sep 9, 2026
[Bug] save_model fires unconditionally at the final step regardless of --save-interval; final optimizer-state write can kill the job Open

Difficulty 3/5 1-2 days Newbie friendliness 55/100

THUDM/slime#2370 · 1 comment · Opened Sep 8, 2026
[Bug] Fully-async multi-agent rollout crashes on nested groups Open

bug

Difficulty 3/5 1-2 days Newbie friendliness 76/100

THUDM/slime#2344 · Opened Sep 2, 2026
[Discussion] Support background evaluation while fully-async training rollout is running Open

question

Difficulty 5/5 Over a week Newbie friendliness 30/100

THUDM/slime#2339 · 2 comments · Opened Aug 31, 2026
[Question] realign 使用本轮 response 长度判断 Open

question

Difficulty 4/5 3-5 days Newbie friendliness 38/100

THUDM/slime#2338 · 1 comment · Opened Aug 31, 2026
[Bug] Qwen3.5-VL grounding RL silently wrong: bundled SGLang image drops H/W rows of [3,T] M-RoPE (upstream sgl-project/sglang#35345, fix #35744) Open

Difficulty 3/5 1-2 days Newbie friendliness 64/100

THUDM/slime#2336 · Opened Aug 28, 2026
[Bug] W&B initialization fails with wandb 0.28.2 because wandb.util.generate_id was removed Open

bug

Difficulty 1/5 1-3 hours Newbie friendliness 90/100

THUDM/slime#2332 · 1 comment · Opened Aug 26, 2026
[Question] Multiple critic updates per actor update (SAO-style faster value update)? Open

question

Difficulty 5/5 Over a week Newbie friendliness 35/100

THUDM/slime#2329 · Opened Aug 26, 2026
如何才能快速训练GLM 5.2 Open

question

Difficulty 4/5 3-5 days Newbie friendliness 35/100

THUDM/slime#2324 · 1 comment · Opened Aug 24, 2026
[Question] 什么时候支持ling3.0呢？ Open

question

Difficulty 5/5 Over a week Newbie friendliness 25/100

THUDM/slime#2314 · Opened Aug 24, 2026
[Bug] Agent adapters rewrite malformed tool-call arguments before the client sees them Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

THUDM/slime#2299 · Opened Aug 21, 2026
[Agent RL] Qwen3.5 history re-templating causes false trajectory forks Open

Difficulty 4/5 3-5 days Newbie friendliness 30/100

THUDM/slime#2288 · Opened Aug 19, 2026
[Question/Performance] Full message history is re-tokenized on every coding-agent turn Open

question

Difficulty 5/5 Over a week Newbie friendliness 35/100

THUDM/slime#2285 · 3 comments · Opened Aug 19, 2026
[Bug] Qwen3.5 MoE online weight update sends ETP-sharded TEGroupedLinear weights without all-gather Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 48/100

THUDM/slime#2280 · Opened Aug 17, 2026
Per-comm memory probe (available_memory/memory_stats) is O(#allocator segments) and slows weight sync & training under expandable_segments Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

THUDM/slime#2268 · 1 comment · Opened Aug 13, 2026
[Bug] Training/log-prob forward materializes full-vocab fp32 logits over the entire packed sequence → OOM for large-vocab, long multi-turn RL Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 45/100

THUDM/slime#2253 · 1 comment · Opened Aug 5, 2026
[Question] Proposal: Add Train–Inference Consistency Data Collection to slime Open

question

Difficulty 5/5 Over a week Newbie friendliness 25/100

THUDM/slime#2244 · 1 reaction · Opened Jul 27, 2026
[Question] 请问最新镜像（slimerl/slime:latest）支持A800吗 Open

question

Difficulty 1/5 Under an hour Newbie friendliness 25/100

THUDM/slime#2232 · 1 comment · Opened Jul 24, 2026
[Bug] Normalize reward advantages by explicit sample groups Open

Difficulty 3/5 1-2 days Newbie friendliness 66/100

THUDM/slime#2230 · Opened Jul 24, 2026
[Bug] Validate DAPO labels as exact integers without float coercion Open

Difficulty 3/5 1-2 days Newbie friendliness 42/100

THUDM/slime#2231 · Opened Jul 24, 2026
[Question]什么时候能支持国产昇腾芯片呢？ Open

question

Difficulty 5/5 Over a week Newbie friendliness 25/100

THUDM/slime#2225 · 1 comment · Opened Jul 21, 2026
[Question]将qwen3.5-27B转换为torch dist后，再转换回hf格式，使用sglang部署，在tp=4/8的情况下模型会输出且只输出大量‘！’ Open

question

Difficulty 4/5 3-5 days Newbie friendliness 30/100

THUDM/slime#2222 · 3 comments · Opened Jul 20, 2026
Your project is on StackMap — a curated map of the AI stack Open

Difficulty 1/5 Under an hour Newbie friendliness 25/100

THUDM/slime#2214 · 3 comments · Opened Jul 16, 2026
[Question] Do we support the SAO method used in the training of GLM-5.2 Open

question

Difficulty 5/5 Over a week Newbie friendliness 25/100

THUDM/slime#2212 · 5 reactions · Opened Jul 16, 2026
Delta weight sync (NCCL) produces NaN weights on Qwen3.5-122B MoE → "probability tensor contains inf/nan" crash Open

Difficulty 4/5 3-5 days Newbie friendliness 42/100

THUDM/slime#2209 · 1 comment · Opened Jul 15, 2026
[Question] Do we support general on-policy distillation? Open

question

Difficulty 5/5 Over a week Newbie friendliness 30/100

THUDM/slime#2202 · Opened Jul 13, 2026
GLM-5.2 SparseMLA TileLang backward returns NaN gradients from finite inputs Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

THUDM/slime#2201 · 2 comments · Opened Jul 13, 2026
[Question] qwen3.6-35B-A3b训练的时候只保存了文本权重，多模态权重没有保存，怎么办 Open

question

Difficulty 4/5 3-5 days Newbie friendliness 25/100

THUDM/slime#2194 · Opened Jul 10, 2026
[Bug] Colocate weight update fails with torch_memory_saver/offload because PyTorch CUDA IPC _share_cuda_ raises cudaErrorInvalidValue Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 45/100

THUDM/slime#2188 · 2 comments · Opened Jul 8, 2026
[Bug] `offload_train` train worker dies on CUDA 13 with `libcudart.so.12: cannot open shared object file` — the torch_memory_saver `LD_PRELOAD` `.so` is chosen by filename, not by CUDA runtime Open

bug

Difficulty 3/5 1-2 days Newbie friendliness 72/100

THUDM/slime#2186 · 1 comment · Opened Jul 7, 2026
[Bug] Using q instead of normalized q in Megatron's DSA MLA indexer for GLM 5 models Open

bug

Difficulty 3/5 1-2 days Newbie friendliness 58/100

THUDM/slime#2165 · 1 comment · Opened Jul 1, 2026
[RFC] Shard-level P2P weight sync for non-colocate training Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

THUDM/slime#2147 · Opened Jun 29, 2026
[Bug] When making minimax m2.7 hf checkpoint to torch_dist format, ran into error Open

bug

Difficulty 3/5 1-2 days Newbie friendliness 48/100

THUDM/slime#2129 · 3 comments · Opened Jun 24, 2026
Delta weight sync: failed engine apply is silently swallowed, leaving sender snapshot ahead of receiver state Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

THUDM/slime#2104 · Opened Jun 19, 2026
[Question] Need help to support Qwen3.5 dense(/moe) VLM megatron.bridge plugin together Open

question

Difficulty 5/5 Over a week Newbie friendliness 25/100

THUDM/slime#2073 · 2 comments · Opened Jun 13, 2026
[Question] code agent rl 数据格式问题 Open

question

Difficulty 3/5 1-2 days Newbie friendliness 35/100

THUDM/slime#2052 · 1 comment · Opened Jun 11, 2026
[Question] torch_memory_saver 报错only hook_mode=preload supports Open

question

Difficulty 4/5 3-5 days Newbie friendliness 28/100

THUDM/slime#2018 · 1 comment · Opened Jun 4, 2026
[Question] Any plans to support pipeline RL to avoid ramp down time during weight update in sglang servers Open

question

Difficulty 5/5 Over a week Newbie friendliness 30/100

THUDM/slime#2007 · 3 comments · Opened Jun 2, 2026
[Proposal] TCOD — extending slime's On-Policy Distillation to multi-turn agents Open

question

Difficulty 5/5 Over a week Newbie friendliness 25/100

THUDM/slime#2002 · Opened Jun 1, 2026
[RFC] Integrate TransferQueue into slime as an Optional Training Data Plane Open

question

Difficulty 5/5 Over a week Newbie friendliness 25/100

THUDM/slime#1971 · 3 comments · Opened May 29, 2026
[Question] retool example: compute_log_probs(logits.clone(), tokens, tp_group) torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 58.15 GiB. Open

question

Difficulty 4/5 3-5 days Newbie friendliness 42/100

THUDM/slime#1951 · 5 comments · Opened May 27, 2026
[Bug] TorchMemorySaver observes invalid LD_PRELOAD. when add --disable-weights-backuper Open

bug

Difficulty 3/5 1-2 days Newbie friendliness 48/100

THUDM/slime#1936 · Opened May 22, 2026
[Question] Support for DeepSeek V4? Open

question

Difficulty 5/5 Over a week Newbie friendliness 25/100

THUDM/slime#1910 · 3 comments · Opened May 14, 2026
[Bug] 保存模型hang住 Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 42/100

THUDM/slime#1908 · Opened May 14, 2026
[Question] I found that ois is numerically equivalent to tis during training Open

question

Difficulty 4/5 3-5 days Newbie friendliness 35/100

THUDM/slime#1907 · Opened May 13, 2026
Segfault in checkpoint save (cudaMemcpyAsync) — regression in v0.2.4, works in v0.2.3 Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

THUDM/slime#1893 · 1 comment · Opened May 6, 2026

THUDM/slime

Get new issues in your inbox