apply_rotary_pos_emb_flashatt_fp32() takes 2 positional arguments but 4 were given

[black-tea01]

I follow the install pip install git+https://github.com/huggingface/transformers accelerate and the pip list:
transformers==4.49.0.dev0
tokenizers==0.21.0
accelerate== 0.34.0
flash-attn==2.5.8

apply_rotary_pos_emb_flashatt_fp32 question from training

File ""/mnt/beegfs_lfs3/training/riderapi/al_team/qian.ye/miniconda3/envs/qwen2/lib/python3.10/site-packages/torch/nn/modules/module.py"", line 1844, in _call_impl"
"1739588804.641397953","[rank7]:     return inner()"
"1739588804.641407251","[rank7]:   File ""/mnt/beegfs_lfs3/training/riderapi/al_team/qian.ye/miniconda3/envs/qwen2/lib/python3.10/site-packages/torch/nn/modules/module.py"", line 1790, in inner"
"1739588804.641416549","[rank7]:     result = forward_call(*args, **kwargs)"
"1739588804.641424894","[rank7]:   File ""/mnt/beegfs_lfs3/training/riderapi/al_team/qian.ye/miniconda3/envs/qwen2/lib/python3.10/site-packages/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py"", line 201, in forward"
"1739588804.641432762","[rank7]:     q, k = apply_rotary_pos_emb_flashatt(q.unsqueeze(0), k.unsqueeze(0), cos, sin)"
"1739588804.641441583","[rank7]: TypeError: apply_rotary_pos_emb_flashatt_fp32() takes 2 positional arguments but 4 were given"

[2U1]

It's a bit odd, I don't get it how the arguments are passed. I can't reproduce the issue.
Can you show your dataset example and the scripts you are running?

[black-tea01]

I have solved the question, the lasted transformers have update the cos/sin embeding and the code
def apply_rotary_pos_emb_flashatt_fp32(tensor: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor: tensor_ = tensor.float() cos = freqs.cos().float() sin = freqs.sin().float() output = apply_rotary_emb(tensor_, cos, sin).type_as(tensor) return output
need to be replace to be
def apply_rotary_pos_emb_flashatt_fp32( q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor ) -> Tuple[torch.Tensor, torch.Tensor]: cos = cos.chunk(2, dim=-1)[0].contiguous() sin = sin.chunk(2, dim=-1)[0].contiguous() cos = cos.float() sin = sin.float() q_embed = apply_rotary_emb(q.float(), cos, sin).type_as(q) k_embed = apply_rotary_emb(k.float(), cos, sin).type_as(k) return q_embed, k_embed
for the lasted transformers.

[2U1]

Oh I see, it was updated again. I'll fix the README for now until it is relesead.
Thanks for letting me know!

pip install git+https://github.com/huggingface/transformers.git@9d2056f12b66e64978f78a2dcb023f65b2be2108
Maybe this version should work with the code.