张宇

icon indicating an email [email protected]

icon company Alibaba Group icon location Machine Learning Engineer @Alibaba Group

Results 5 comments of 张宇

Authentication token does not exist, failed to access model unknown which may not exist or may be private. Please login first.

Same error comes together with `TypeError: Accelerator.__init__() got an unexpected keyword argument 'use_seedable_sampler'`, update accelarator solves it: `pip install accelerate==0.27.2`

zero3 training hangs with mixed multimodal dataset

> Is the problem solved now? I come across the same problem with zero 3 on mix-modality training. The training always hangs at the make_experience stage, and the progress is...

训练PPO出现问题：ValueError: Target module ModuleDict( (default): Identity() (reward): Identity() ) is not supported. Currently, only the following modules are supported: `torch.nn.Linear`, `torch.nn.Embedding`, `torch.nn.Conv2d`, `transformers.pytorch_utils.Conv1D`.

The target_modules pattern ```"^(?!.*visual).*(?:o_proj|k_proj|gate_proj|up_proj|down_proj|q_proj|v_proj).*"``` caused this error. PEFT match it to lora layers in the model, tries to add lora for `Identity()` layer. The ```patch_target_modules ``` function should be fixed...

训练PPO出现问题：ValueError: Target module ModuleDict( (default): Identity() (reward): Identity() ) is not supported. Currently, only the following modules are supported: `torch.nn.Linear`, `torch.nn.Embedding`, `torch.nn.Conv2d`, `transformers.pytorch_utils.Conv1D`.

The line ```src\llamafactory\train\trainer_utils.py", line 146, in create_reward_model model.pretrained_model.load_adapter(finetuning_args.reward_model, "reward") ``` seems not loading pretrained rm as intended, since load_adapter degraded to add_adapter, which does not load any state_dict. Repeat the...

[Bug]: When using the VLLM framework to load visual models, CPU memory overflow occurs while continuously processing data with images.

> [@ywang96](https://github.com/ywang96), I have tried with latest main branch > > 1. disable_mm_preprocessor_cache=True > 2. export VLLM_MM_INPUT_CACHE_GIB=4 > > still facing the same issue Have you solved the problem? I'm...