Phi-3：小模型，大未来！（附魔搭社区推理、微调实战教程）

from modelscope import snapshot_downloadmodel_dir = snapshot_download("LLM-Research/Phi-3-mini-128k-instruct")

import torchfrom modelscope import snapshot_downloadfrom transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
torch.random.manual_seed(0)
model_dir = snapshot_download("LLM-Research/Phi-3-mini-128k-instruct")
model = AutoModelForCausalLM.from_pretrained(model_dir, device_map="cuda", torch_dtype="auto", trust_remote_code=True, )tokenizer = AutoTokenizer.from_pretrained(model_dir)
messages = [{"role": "system", "content": "You are a helpful digital assistant. Please provide safe, ethical and accurate information to the user."},{"role": "user", "content": "Can you provide ways to eat combinations of bananas and dragonfruits?"},{"role": "assistant", "content": "Sure! Here are some ways to eat bananas and dragonfruits together: 1. Banana and dragonfruit smoothie: Blend bananas and dragonfruits together with some milk and honey. 2. Banana and dragonfruit salad: Mix sliced bananas and dragonfruits together with some lemon juice and honey."},{"role": "user", "content": "What about solving an 2x + 3 = 7 equation?"},]
pipe = pipeline("text-generation",model=model,tokenizer=tokenizer,)
generation_args = {"max_new_tokens": 500,"return_full_text": False,"temperature": 0.0,"do_sample": False,}
output = pipe(messages, **generation_args)print(output[0]['generated_text'])

git clone https://www.modelscope.cn/LLM-Research/Phi-3-mini-4k-instruct-onnx.git

pip install --pre onnxruntime-genaicurl https://raw.githubusercontent.com/microsoft/onnxruntime-genai/main/examples/python/model-qa.py -o model-qa.py

python model-qa.py -m Phi-3-mini-4k-instruct-onnx/cpu_and_mobile/cpu-int4-rtn-block-32 -l 2048

git clone https://github.com/modelscope/swift.gitcd swiftpip install -e .[all]

CUDA_VISIBLE_DEVICES=0,1,2,3NPROC_PER_NODE=4 \swift sft \--model_type phi3-4b-4k-instruct \--dataset ms-agent-for-agentfabric-default alpaca-en ms-bench ms-agent-for-agentfabric-addition coig-cqia-ruozhiba coig-cqia-zhihu coig-cqia-exam coig-cqia-chinese-traditional coig-cqia-logi-qa coig-cqia-segmentfault coig-cqia-wiki \--batch_size 2 \--max_length 2048 \--use_loss_scale true \--gradient_accumulation_steps 16 \--learning_rate 5e-5 \--use_flash_attn true \--eval_steps 500 \--save_steps 500 \--train_dataset_sample -1 \--dataset_test_ratio 0.1 \--val_dataset_sample 10000 \--num_train_epochs 2 \--check_dataset_strategy none \--gradient_checkpointing true \--weight_decay 0.01 \--warmup_ratio 0.03 \--save_total_limit 2 \--logging_steps 10 \--sft_type lora \--lora_target_modules ALL \--lora_rank 8 \--lora_alpha 32

--custom_train_dataset_path xxx.jsonl \--custom_val_dataset_path yyy.jsonl \

# Experimental environment: A100CUDA_VISIBLE_DEVICES=0 \swift infer \--ckpt_dir "/path/to/output/phi3-4b-4k-instruct/vx-xxx/checkpoint-xxx" \--load_dataset_config true \--max_new_tokens 2048 \--temperature 0.1 \--top_p 0.7 \--repetition_penalty 1. \--do_sample true \--merge_lora false \