Repository Issue Activity (beta)

#10058 [RFC] Reward model (value-head RM) support in infer/deploy via vLLM

#10057 About qwen3.5 mtp sft, multi-modality data

#9973 enable_channel_loss crashes with IndexError when combined with sequence parallelism (sequence_parallel_size > 1)

#10050 `resume_from_checkpoint` loses the epoch seed: the rebuilt sampler replays the epoch-0 order, so already-trained samples are re-trained

#9991 GRPO训练，用colocate模式在rollout阶段会出现模型参数或者优化器参数没有被offload导致oom的问题

#10053 support for deepseek v4 flash vision exp

#8256 [Bug] predict_with_generate=True causes 1D dimension collapse of eval_prediction.predictions in custom metrics (DDP)

#9892 支持Ling-3.0-flash和Ling-3.0-tiny和Muse-Glimmer-30B

#10041 Qwen3-Omni使用add_special_tokens参数sft报错

#9508 Example for training on SWE (agentic software-engineering) tasks?

#9518 Logging Per-Sample Evaluation Loss During Inference in Swift PT for Qwen3-VL

#9240 Qwen3.5 9B async模式GRPO指定freeze vit时训练挂掉

#9371 fp8_param_gather does not reduce VRAM consumation comparing with fp16 in megatron

#9507 Multimodal SFT silently hangs forever on certain clips: load_audio (librosa→audioread) and decord video decode have no timeout (DataLoader deadlock)

#10013 [NPU][Megatron GRPO] 开启 PP 后 P2P 通信在 HcclGroupEnd 报错并导致进程崩溃

#10040 使用npu进行qwen35 27B megatron sft时，contex parallel设置是否能正确生效呢

#10024 Bug: unexpected behavior with special characters in input

#10022 求教：--loss_scale ignore_empty_think是否会影响SFT的think/no-think的混合训练？

#10014 Image augmentation is a bit tricky, tied to models, and hard to do for training data only

#10006 Data containing multiple images is not supported for the Qwen3.5 series.

#10016 Data loaders somehow pull in CUDA

#9399 dist_muon训练时卡住

#9454 grpo多轮训练中，没有设置dynamic_sample，但是报错Padding free mode is not supported for dynamic sample

#9435 【NPU】swift grpo qwen3-32b，开启sleep_level和vllm_tensor_parallel_size后报错

#9436 Consider using pre-built Flash Attention kernels via `kernels`