labmlai / neox

Simple Annotated implementation of GPT-NeoX in PyTorch

⚠️ We have moved this implementation to https://nn.labml.ai/neox/index.html.

---

Simple Annotated implementation of GPT-NeoX in PyTorch

This is a simpler implementation of GPT-NeoX in PyTorch. We have taken out several optimizations in GPT-NeoX for simplicity.

Annotated implementation

Sample usages

• Notebook to download checkpoints and test the model
• Fine-tuning example
• Text generating on multi GPU
• Text generating on a small GPU (offloading to CPU)
• Text generating on a single 48GB GPU

Playground

Evaluation

Task	Metric	NeoX Impl (2 GPU)	This repo (1 GPU)
anli_r1	acc	0.3270	0.3360
	acc_stderr	0.0148	0.0149
anli_r2	acc	0.3410	0.3350
	acc_stderr	0.0150	0.0149
anli_r3	acc	0.3567	0.3525
	acc_stderr	0.0138	0.0149
hellaswag	acc	0.5351	0.5353
	acc_stderr	0.0050	0.0050
	acc_norm	0.7140	0.7145
	acc_norm_stderr	0.0045	0.0045
lambada	acc	0.7211	0.7204
	acc_stderr	0.0062	0.0063
	ppl	3.6760	3.6375
	ppl_stderr	0.0760	0.0747
piqa	acc	0.7748	0.7758
	acc_stderr	0.0097	0.0097
	acc_norm	0.7786	0.7845
	acc_norm_stderr	0.0097	0.0096
winogrande	acc	0.6598	0.6582
	acc_stderr	0.0133	0.0133
wsc	acc	0.5096	0.5000
	acc_stderr	0.0493	0.0493