zhang00

I added my own data to the first stage for 10 rounds of training based on the trained weights given by the author. The semantic segmentation effect became better, but...

LLM权重放在哪里呢？

1

你好： 我观察到yaml里有llm_model，这在您的说明中表示llm_model的ckpt，但是这/data/llava-v1.5-7b似乎是一个文件夹，从代码来看似乎需要tokenizer和llm的ckpt self.llm_tokenizer = LlamaTokenizer.from_pretrained(llm_model, use_fast=False, truncation_side="left") self.llm_model = LlamaForCausalLM.from_pretrained(llm_model, torch_dtype=torch.float16, low_cpu_mem_usage=True)