karpathy / llm.c

LLM training in simple, raw C/CUDA
MIT License
23.6k stars 2.64k forks source link

Relax grad tensor thresholds in tests #615

Closed gordicaleksa closed 3 months ago

gordicaleksa commented 3 months ago

Our CI is using A4000 GPUs that didn't caught some of the tight grad tensor thresholds I've set in https://github.com/karpathy/llm.c/pull/614

Relaxing a few of these reported by @rosslwheeler

On a side note: @karpathy would be cool to understand exactly how many GPUs we've got as CI runners, and potentially ask if we can trade a few of these for some different GPUs. :) (one can but hope)

rosslwheeler commented 3 months ago

Passes my tests now. Thank you for the quick response/fix!