Merge pull request #801 from ngc92/ngc92/fix-test

adjusted error thresholds

reorder checks to match actual tensor order

Merge pull request #766 from msaroufim/patch-1

Update README.md

cuda mode -> gpu mode

change default params: use tinyshakespeare and decrease LR

Merge pull request #724 from GaoYusong/master

Merge pull request #733 from zhangpiu/feature/llm.cpp

Merge pull request #735 from gordicaleksa/minor_refactor3

Merge pull request #744 from dengl11/pr

Merge pull request #745 from karpathy/feature/managed2

i misspelled reduced

reduce across GPUs nicer

fallback to memory allocation of m,v,master_weights on host automatically in case of OOM. will run slower but won't OOM