Sparse Finetuning MPT
Collection
Explore our breakthrough in sparse fine-tuning LLMs! Our novel method maintains downstream accuracy even with >70% sparsity.
•
13 items
•
Updated
•
3
Paper: Sparse Finetuning for Inference Acceleration of Large Language Models
Code: https://github.com/neuralmagic/deepsparse/tree/main/research/mpt
This model was produced from a MPT-7B base model finetuned on the GSM8k dataset with pruning applied using SparseGPT and retrain for 2 epochs with L2 distillation. Then it was exported for optimized inference with DeepSparse.
GSM8k zero-shot accuracy with lm-evaluation-harness : 30.33% (FP32 baseline is 28.2%)
from deepsparse import TextGeneration
model_path = "hf:neuralmagic/mpt-7b-gsm8k-pruned40-quant" # or use a sparsezoo stub (zoo:mpt-7b-gsm8k_mpt_pretrain-pruned40_quantized)
model = TextGeneration(model=model_path)
model("There are twice as many boys as girls at Dr. Wertz's school. If there are 60 girls and 5 students to every teacher, how many teachers are there?", max_new_tokens=50)
All MPT model weights are available on SparseZoo and CPU speedup for generative inference can be reproduced by following the instructions at DeepSparse
Model Links | Compression |
---|---|
neuralmagic/mpt-7b-gsm8k-quant | Quantization (W8A8) |
neuralmagic/mpt-7b-gsm8k-pruned40-quant | Quantization (W8A8) & 40% Pruning |
neuralmagic/mpt-7b-gsm8k-pruned50-quant | Quantization (W8A8) & 50% Pruning |
neuralmagic/mpt-7b-gsm8k-pruned60-quant | Quantization (W8A8) & 60% Pruning |
neuralmagic/mpt-7b-gsm8k-pruned70-quant | Quantization (W8A8) & 70% Pruning |
neuralmagic/mpt-7b-gsm8k-pruned70-quant | Quantization (W8A8) & 75% Pruning |
neuralmagic/mpt-7b-gsm8k-pruned80-quant | Quantization (W8A8) & 80% Pruning |
For general questions on these models and sparsification methods, reach out to the engineering team on our community Slack.