Li Xiaozhe

[Question] flash-attn question

2

### Question I follow readme to create my environment, but when I using lora to finetune llava1.5 model（finetune_task_lora.sh）, I facing this problem: import flash_attn_2_cuda as flash_attn_cuda ImportError: /home/lxz/miniconda3/envs/llava/lib/python3.10/site-packages/flash_attn_2_cuda.cpython-310-x86_64-linux-gnu.so: undefined symbol:...

[Question] I finetuning llava1.5-7b model using lora on llava_mix_665k dataset but get 545 on MME perception score. Has anyone else encountered this issue?

10

### Question I finetuning llava1.5-7b model using lora on llava_mix_665k dataset on 4*A100-40g. However, model has bad performance on MME benchmark. perception score:545, cognition score 197. Has anyone else encountered...

I used the code provided in the gitHub repository for inference. Why am I getting the error: 'ValueError: total_frames cannot be zero'?

Great job! Could you please release the pretrained checkpoint of video-llava so that we can use it in fine-tuning stage? Thanks!

How to load lora weights using StableVideoDiffusionPipeline？

1

How to load lora weights using StableVideoDiffusionPipeline？ I using lora_pipeline = StableVideoDiffusionPipeline.from_pretrained( "/home/data/MODEL/stable-diffusion/stable-video-diffusion-img2vid-xt-1-1", torch_dtype=torch.float16, ) lora_pipeline.load_lora_weights("/home/code/Motion-LoRA/LoRA_forward_onepoint/pytorch_lora_weights.safetensors"), and it turns out "AttributeError: 'StableVideoDiffusionPipeline' object has no attribute 'load_lora_weights'"? Could you tell...

model support

1

Thanks for your great work! Can dola support more LLM model? Such as llama3.1, llama2, qwen2 or Mistral serious?

Thanks for your excellent work on OCRBenchV2. We are currently attempting to add support for OCRBenchV2 within VLMEvakKit. However, based on your publicly available codebase, we have been unable to...

qwen3_vl.Qwen3VLModel'> does not support tensor parallel yet!

1

您好，我在使用grpo训练qwn3VL模型，当tensor_model_parallel_size开4，报了如下错误（tp=1 可以正常训练）```qwen3_vl.Qwen3VLModel'> does not support tensor parallel yet!```,进入代码中发现是源于``` tp_plan = getattr(self.model.config, "base_model_tp_plan", None) or {} if not tp_plan and tp_size > 1: raise ValueError( f"{type(self.model)} does not support tensor parallel...

- Add PRED_FORMAT and EVAL_FORMAT environment variables - Support xlsx/json/tsv for predictions, csv/json for evaluations - Fix DataFrame ↔ JSON conversion with type detection

No image input question

3

您好，使用grpo训练qwen3-VL-8B-Instruct模型的时候，遇到VL模型看不见图片的问题，具体表现为在每一步模型Rollout的输出```no image provided```请问这是什么原因？rollout的engine是```sglang```.