DeepSpeed版本训练报错RuntimeError: The size of tensor a (780) must match the size of tensor b (781)

[tize-72]

不是unsloth版本的也会训练报维度错误，这个不知道有人遇到过吗

[hellobiek]

一样遇到，但是我的是 unsloth 版本

[Mrkkew]

遇到了如下的错误：
[rank3]: File "/home/jovyan/work/tanzichang/miniconda/envs/vl_vllm/lib/python3.10/site-packages/transformers/trainer.py", line 2171, in train
[rank3]: return inner_training_loop(
[rank3]: File "/home/jovyan/work/tanzichang/miniconda/envs/vl_vllm/lib/python3.10/site-packages/transformers/trainer.py", line 2531, in _inner_training_loop
[rank3]: tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
[rank3]: File "/home/jovyan/work/tanzichang/miniconda/envs/vl_vllm/lib/python3.10/site-packages/transformers/trainer.py", line 3675, in training_step
[rank3]: loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
[rank3]: File "/home/jovyan/work/tanzichang/miniconda/envs/vl_vllm/lib/python3.10/site-packages/trl/trainer/grpo_trainer.py", line 495, in compute_loss
[rank3]: rewards_per_func[:, i] = torch.tensor(output_reward_func, dtype=torch.float32, device=device)
[rank3]: RuntimeError: The expanded size of the tensor (10) must match the existing size (11) at non-singleton dimension 0. Target sizes: [10]. Tensor sizes: [11]
请问是怎么回事

[SwayDy]

我也遇到了：RuntimeError: The size of tensor a (1024) must match the size of tensor b (1025) at non-singleton dimension 1

[SwayDy]

我也遇到了：RuntimeError: The size of tensor a (1024) must match the size of tensor b (1025) at non-singleton dimension 1

train_Datawhale-R1_unsloth.py中第201行:
max_seq_length=training_args.max_completion_length, # 设置最大序列长度
改成:
max_seq_length=training_args.max_prompt_length + training_args.max_completion_length, # 设置最大序列长度
就能跑了

[xiaoan17]

我也遇到了：RuntimeError: The size of tensor a (1024) must match the size of tensor b (1025) at non-singleton dimension 1我也遇到了：RuntimeError：张量 a 的大小（1024）必须与张量 b 的大小（1025）在非单元素维度 1 上匹配

train_Datawhale-R1_unsloth.py中第201行: max_seq_length=training_args.max_completion_length, # 设置最大序列长度 改成: max_seq_length=training_args.max_prompt_length + training_args.max_completion_length, # 设置最大序列长度 就能跑了

十分感谢，修改之后可以正常运行了。

[anine09]

我也遇到了：RuntimeError: The size of tensor a (1024) must match the size of tensor b (1025) at non-singleton dimension 1

train_Datawhale-R1_unsloth.py中第201行: max_seq_length=training_args.max_completion_length, # 设置最大序列长度 改成: max_seq_length=training_args.max_prompt_length + training_args.max_completion_length, # 设置最大序列长度 就能跑了

我没有 get 到为什么会出现这种情况， @tize-72 你这样修改后能在 DeepSpeed 的版本上运行吗？