houliangxue

F1 79.96 on ontonotes 5 with your pretrained spanbert_large

1

Is F1 79.6 the highest score you've ever got on Ontonotes? Why I got F1 79.95 without anything changes for your experiments.conf. Did you set a seed in your experiment?

May I know your training time on TPU，and how many TPUs did you use？Thanks

qwen1.5-7b-chat模型4090内存溢出

2

配置：4090 ，24G显存 加载方式： self.model = AutoModelForCausalLM.from_pretrained(model_path,torch_dtype="auto",device_map="auto").eval() self.tokenizer = AutoTokenizer.from_pretrained(model_path) self.generation_Config = GenerationConfig.from_pretrained(model_path, trust_remote_code=True) 当token达到6k时，出现内存溢出 torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 1.14 GiB (GPU 1; 23.65 GiB total capacity;...

when train gap task ,I got this error ,is it a bug?

1

when I run: python /main.py --task_name gap --do_eval --do_train --eval_batch_size 10 --data_dir "autodl-tmp/bert-commonsense-master/data/" --bert_model bert-large-uncased --max_seq_length 128 --train_batch_size 16 --alpha_param 20 --beta_param 0.2 --learning_rate 5.0e-6 --num_train_epochs 1.0 --output_dir autodl-tmp/bert-commonsense-master/model_output/ I...

another question :why your c2f+spanbert result (80.2) is better than Joshi 's paper(79.6) . That’s your own implementation for c2f+spanbert？

单机单卡4090，lora微调deepseek-r1-distill-1.5B卡死，稳定复现，会导致整个机器卡死，nvdia-smi命令失效，

2

### Reminder - [x] I have read the above rules and searched the existing issues. ### System Info  ### Reproduction ， ### Others _No response_

### 🔎 Search before asking - [x] I have searched the PaddleOCR [Docs](https://paddlepaddle.github.io/PaddleOCR/) and found no similar bug report. - [x] I have searched the PaddleOCR [Issues](https://github.com/PaddlePaddle/PaddleOCR/issues) and found no...

paddleocr_VL漏识别问题

2

### 🐛 Bug (问题描述) 1、部署方案：docker部署 paddlex genai_server --model_name PaddleOCR-VL-0.9B --backend vllm --port 8080 2 、推理方式 推理方式1：通过OPENAI方式请求 self.client = OpenAI(api_key='xx', base_url="http://192.168.2.115:8080/v1",) messages = [{ "role": "user", "content": [ { "type": "text",...

输入长文本时，开头读的总是断断续续，是怎么回事，

2

推理代码如下 if request.stream: print("#################流式生成#################") def generate(): # 调用模型生成音频（注意：需根据语速调整，若模型支持） for chunk in cosyvoice.inference_zero_shot( request.input, # 使用OpenAI的"input"参数 prompt_text, # 提示文本（可根据voice调整音色） prompt_speech_16k, stream=True ): audio_tensor = chunk["tts_speech"] # 根据response_format转换格式（支持mp3/wav） buffer = io.BytesIO() torchaudio.save(...

非流式比流式更快吗

6

4090部署cosyvoice3, 非流式rtf 0.1左右，流式0.2左右， 非流式 流式