Tiny-moe is a Mixture of Experts (MoE) made with the following models using
LazyMergekit:
1base_model: Corianas/Tiny_Test
2gate_mode: cheap_embed # one of "hidden", "cheap_embed", or "random"
3dtype: bfloat16 # output dtype (float32, float16, or bfloat16)
4## (optional)
5# experts_per_token: 2
6experts:
7 - source_model: Corianas/Tiny_Test
8 positive_prompts:
9 - "↨once upon a time"
10 ## (optional)
11 # negative_prompts:
12 # - "This is a prompt expert_model_1 should not be used for"
13 - source_model: Corianas/TinyTask-minipaca
14 # ... and so on
15 positive_prompts:
16 - "↨please"
1!pip install -qU transformers bitsandbytes accelerate
2
3from transformers import AutoTokenizer
4import transformers
5import torch
6
7model = "Corianas/Tiny-moe"
8
9tokenizer = AutoTokenizer.from_pretrained(model)
10pipeline = transformers.pipeline(
11 "text-generation",
12 model=model,
13 model_kwargs={"torch_dtype": torch.float16, "load_in_4bit": True},
14)
15
16messages = [{"role": "user", "content": "Explain what a Mixture of Experts is in less than 100 words."}]
17prompt = pipeline.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
18outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
19print(outputs[0]["generated_text"])
