Out-of-Index Error when training by `Qwen2VLFlashAttention2`

[heli-qi]

System Info

• transformers version: 4.45.0.dev0
• Platform: Linux-5.15.0-94-generic-x86_64-with-glibc2.35
• Python version: 3.10.14
• Huggingface_hub version: 0.24.5
• Safetensors version: 0.4.4
• Accelerate version: 0.33.0
• Accelerate config: not found
• PyTorch version (GPU?): 2.2.0+cu121 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?:
• Using GPU in script?:
• GPU type: NVIDIA GeForce RTX 4090 D

Who can help?

@ArthurZucker @amyeroberts

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Hi, I'm finetuning the newly-released Qwen2VLForConditionalGeneration model by LoRA. I'm building the model by

Qwen2VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2-VL-7B-Instruct", attn_implementation="flash_attention_2", torch_dtype=torch.float16
)

I found attn_implementation="flash_attention_2" activates Qwen2VLFlashAttention2 which will throw a out-of-index error during training. When I switch to attn_implementation="sdpa", the error does not come out and training goes smoothly.

After some time of debugging, I located that the problem comes from this line where rotary_seq_len does not properly reflect the length of the input sequence but rather the real length minus 1. I modified this line to rotary_seq_len = cache_position[-1] + 1 in my local transformers offline package and it turns out that the training with flash_attention_2 goes smoothly.

My input batch to the model is as follow:

batch
    input_ids: Tensor (B, seq_len)
    attention_mask: Tensor (B, seq_len)
    labels: Tensor (B, seq_len)
    pixel_values: Tensor (B, res_h, res_w)  # res_h and res_w are the shape of image after processor()
    image_grid_thw: Tensor (B, 3)

I suspect that my input batch to the model has the correct shape, so I'm wondering whether my tiny workaround is the optimal solution to the problem. I really appreciate it if you could tell me some better solutions.

Expected behavior

As Reproduction section. Thanks for your patience for my issue.

[zucchini-nlp]

Will be fixed by #33161 🤗