litgpt
litgpt copied to clipboard
Use FSDP everywhere
Replaces DeepSpeed with FSDP
Requires https://github.com/Lightning-AI/lightning/pull/17845
Closes #116 Closes #177 Closes #169
Falcon 7b takes 32 GB max memory allocated using 2 devices and 32-true or bf16-mixed precision. Loss is NaN, requires further debugging into why.