Open Dreaming-world opened 4 months ago
import torch from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
model_name = "deepseek-ai/deepseek-moe-16b-chat" tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto") model.generation_config = GenerationConfig.from_pretrained(model_name) model.generation_config.pad_token_id = model.generation_config.eos_token_id
messages = [ {"role": "user", "content": "Who are you?"} ] input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt") outputs = model.generate(input_tensor.to(model.device), max_new_tokens=100)
result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True) print(result)
按照上述方式推理，发现比llama-7B速度还慢？正常吗？
import torch from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
model_name = "deepseek-ai/deepseek-moe-16b-chat" tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto") model.generation_config = GenerationConfig.from_pretrained(model_name) model.generation_config.pad_token_id = model.generation_config.eos_token_id
messages = [ {"role": "user", "content": "Who are you?"} ] input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt") outputs = model.generate(input_tensor.to(model.device), max_new_tokens=100)
result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True) print(result)
按照上述方式推理，发现比llama-7B速度还慢？正常吗？