Quantization made by Richard Erkhov.
PowerMoE-3b - bnb 8bits
- Model creator: https://huggingface.co/ibm/
- Original model: https://huggingface.co/ibm/PowerMoE-3b/
Original model description:
pipeline_tag: text-generation inference: false license: apache-2.0 library_name: transformers model-index: - name: ibm/PowerMoE-3b results: - task: type: text-generation dataset: type: lm-eval-harness name: ARC metrics: - name: accuracy-norm type: accuracy-norm value: 58.1 verified: false - task: type: text-generation dataset: type: lm-eval-harness name: BoolQ metrics: - name: accuracy type: accuracy value: 65.0 verified: false - task: type: text-generation dataset: type: lm-eval-harness name: Hellaswag metrics: - name: accuracy-norm type: accuracy-norm value: 71.5 verified: false - task: type: text-generation dataset: type: lm-eval-harness name: OpenBookQA metrics: - name: accuracy-norm type: accuracy-norm value: 41.0 verified: false - task: type: text-generation dataset: type: lm-eval-harness name: PIQA metrics: - name: accuracy-norm type: accuracy-norm value: 79.1 verified: false - task: type: text-generation dataset: type: lm-eval-harness name: Winogrande metrics: - name: accuracy-norm type: accuracy-norm value: 65.0 verified: false - task: type: text-generation dataset: type: lm-eval-harness name: MMLU (5 shot) metrics: - name: accuracy type: accuracy value: 42.8 verified: false - task: type: text-generation dataset: type: lm-eval-harness name: GSM8k (5 shot) metrics: - name: accuracy type: accuracy value: 25.9 verified: false - task: type: text-generation dataset: type: lm-eval-harness name: math (4 shot) metrics: - name: accuracy type: accuracy value: 14.8 verified: false - task: type: text-generation dataset: type: bigcode-eval name: humaneval metrics: - name: pass@1 type: pass@1 value: 20.1 verified: false - task: type: text-generation dataset: type: bigcode-eval name: MBPP metrics: - name: pass@1 type: pass@1 value: 32.4 verified: false
Model Summary
PowerMoE-3B is a 3B sparse Mixture-of-Experts (sMoE) language model trained with the Power learning rate scheduler. It sparsely activates 800M parameters for each token. It is trained on a mix of open-source and proprietary datasets. PowerMoE-3B has shown promising results compared to other dense models with 2x activate parameters across various benchmarks, including natural language multi-choices, code generation, and math reasoning. Paper: https://arxiv.org/abs/2408.13359
Usage
Note: Requires installing HF transformers from source.
Generation
This is a simple example of how to use PowerMoE-3b model.
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
device = "cuda" # or "cpu"
model_path = "ibm/PowerMoE-3b"
tokenizer = AutoTokenizer.from_pretrained(model_path)
# drop device_map if running on CPU
model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device)
model.eval()
# change input text as desired
prompt = "Write a code to find the maximum value in a list of numbers."
# tokenize the text
input_tokens = tokenizer(prompt, return_tensors="pt")
# transfer tokenized inputs to the device
for i in input_tokens:
input_tokens[i] = input_tokens[i].to(device)
# generate output tokens
output = model.generate(**input_tokens, max_new_tokens=100)
# decode output tokens into text
output = tokenizer.batch_decode(output)
# loop over the batch to print, in this example the batch size is 1
for i in output:
print(i)