Search results

Pure BF16配置后，lora层的梯度为什么还是fp32?

bugSomething isn't workingSomething isn't working

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9591 In hiyouga/LLaMA-Factory;

· poryfly opened on Dec 9, 2025

Too many dataloader workers: 16 (max is dataset.num_shards=2). Stopping 14 dataloader workers.

bugSomething isn't workingSomething isn't working

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9590 In hiyouga/LLaMA-Factory;

· ZZQ987 opened on Dec 9, 2025

RoPE scaling configuration not applied when using mcore_adapter for training

bugSomething isn't workingSomething isn't working

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9589 In hiyouga/LLaMA-Factory;

· sijyy opened on Dec 9, 2025

LLaMA-Factory 与 pip 版 unsloth 存在硬性依赖冲突：LF 要求 trl<=0.9.6，但 unsloth(2025.*) 要求 trl>=0.18.2，导致无法同时使用

bugSomething isn't workingSomething isn't working

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9583 In hiyouga/LLaMA-Factory;

· York-Gold opened on Dec 8, 2025

RuntimeError: CUDA error: device-side assert triggered

bugSomething isn't workingSomething isn't working

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9578 In hiyouga/LLaMA-Factory;

· jiangxinufo opened on Dec 6, 2025

get_cosine_with_min_lr_schedule_with_warmup

bugSomething isn't workingSomething isn't working

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9574 In hiyouga/LLaMA-Factory;

· passionvery opened on Dec 5, 2025

inference gemma-7b model error

bugSomething isn't workingSomething isn't working

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9573 In hiyouga/LLaMA-Factory;

· jackmego opened on Dec 4, 2025

希望支持 HunyuanOCR。Add support for fine-tuning HunyuanOCR.

enhancementNew feature or requestNew feature or request

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9571 In hiyouga/LLaMA-Factory;

· huangxinping opened on Dec 4, 2025

Qwen2.5-1.5B 进行SFT训练 grad_norm固定为1.73205 loss一开始很大后面降为0

bugSomething isn't workingSomething isn't working

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9567 In hiyouga/LLaMA-Factory;

· luoqishuai opened on Dec 4, 2025

Mixed thinking and non-thinking training

enhancementNew feature or requestNew feature or request

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9566 In hiyouga/LLaMA-Factory;

· agenticgogogo opened on Dec 4, 2025

训练启动后卡死不动

bugSomething isn't workingSomething isn't working

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9565 In hiyouga/LLaMA-Factory;

· jiangxinufo opened on Dec 4, 2025

适配Ministral-3

enhancementNew feature or requestNew feature or request

pendingThis problem is yet to be addressedThis problem is yet to be addressed

Status: Open.

#9564 In hiyouga/LLaMA-Factory;

· xming521 opened on Dec 3, 2025