RuntimeError: The size of tensor a (2048) must match the size of tensor b (4241) at non-singleton dimension 1

[azzedineA]

I am getting this error even though the same code was running successfully. Any help, please?

---

RuntimeError Traceback (most recent call last)
in <cell line: 0>()

12 frames
/usr/local/lib/python3.11/dist-packages/unsloth/models/llama.py in LlamaModel_fast_forward(self, input_ids, causal_mask, attention_mask, position_ids, past_key_values, inputs_embeds, use_cache, output_attentions, output_hidden_states, return_dict, *args, **kwargs)
701 inputs_embeds.requires_grad_(False)
702 pass
--> 703 inputs_embeds *= attention_mask.unsqueeze(0).transpose(0, 1).transpose(1, 2)
704 if inputs_requires_grad: inputs_embeds.requires_grad_(True)
705 pass

RuntimeError: The size of tensor a (2048) must match the size of tensor b (4241) at non-singleton dimension 1

[SomeoneKong]

+1

trl==0.15.0
unsloth==2025.2.12
unsloth_zoo==2025.2.5
triton==3.2.0

[YannickRuppenthal]

Had a similar problem, but it turns out I had to increase max_seq_length in FastLanguageModel.from_pretrained(...)

[fyj3266098]

+1

[danielhanchen]

Apologies is this for normal finetuning or DPO?

[danielhanchen]

I'm assuming it's something to do with the max sequence length?

[SomeoneKong]

Apologies is this for normal finetuning or DPO?

GRPO

[janglichao]

I'm assuming it's something to do with the max sequence length?

+1, I also got same error,pls help
Name Version Build Channel
unsloth 2025.2.12 pypi_0 pypi
unsloth-zoo 2025.2.5 pypi_0 pypi

[delbyte]

I'm assuming it's something to do with the max sequence length?

Check if your input_ids and attention_mask align:

print("Input IDs shape:", input_ids.shape)
print("Attention Mask shape:", attention_mask.shape)

They should have the same first dimension, e.g., (batch_size, sequence_length).

[ZhangYuHao1]

+1, pls help

[Damonzhl]

+1, pls help

[linyang-ant]

+1, pls help