Search results

[Training Issue] Rewards, loss and AIME24 scores (1st time eval) are 0 when training Qwen3-30B-A3B-Instruct-2507 with GSPO

Status: Open.

#1527 In THUDM/slime;

· tiesanguaixia opened on Jan 31, 2026

working pyproject.toml for environment configuration with uv

Status: Open.

#1526 In THUDM/slime;

· leoleoasd opened on Jan 30, 2026

OOM during long-sequence RL training caused by entropy backward tensors

Status: Open.

#1523 In THUDM/slime;

· huanglei-lpd opened on Jan 30, 2026

OOM when training with long sequences despite using dynamic batch size and sequence parallel

Status: Open.

#1522 In THUDM/slime;

· Dodojordi opened on Jan 30, 2026

关于Tensorboard的观测指标 rollout /advantages 为0的问题

Status: Open.

#1521 In THUDM/slime;

· Mrkkew opened on Jan 30, 2026

Rollout generation结束后，报错

Status: Open.

#1514 In THUDM/slime;

· ray075hl opened on Jan 29, 2026

Featuring: Convert tool for Qwen3-VL moe model to hf format after training with Megatron

Status: Open.

#1510 In THUDM/slime;

· p1k0pan opened on Jan 28, 2026

自定义 async def batched_custom_rm(args, samples: list[Sample]) -> list[float]，的输入和输出都是什么啊？

Status: Open.

#1507 In THUDM/slime;

· Chunfeng1994 opened on Jan 28, 2026

Creating a new Dataset source throws in <code>rollout.py</code> because <code>len(data_source.dataset)</code> is used.

Status: Open.

#1498 In THUDM/slime;

· TSunny007 opened on Jan 27, 2026

CUDA_ERROR_OUT_OF_MEMORY

Status: Open.

#1497 In THUDM/slime;

· glennccc opened on Jan 27, 2026

Bug: sft and rl in qwen3-next do not support mtp training

Status: Open.

#1496 In THUDM/slime;

· huang3eng opened on Jan 26, 2026

More details about the examples of on-policy distillation

Status: Open.

#1493 In THUDM/slime;

· RikkiXu opened on Jan 26, 2026