i got room service. acai bowl and some sausages and a smoothie.
had a chat with e about the plan and updates. the goal was to speed up and try to fit 70b in one h100 instead of two. tried out online dynamic quantization with vllm (fp8). it was actually slower than before and also still required the same amount of memory. pointless.
next explored qlora training with bitsandbytes. kept facing 0.0 grad_norm issues and tried to debug the entire day and couldn't figure out why. a problem for another day to solve.
at night all of us went to the rooftop and everyone shared about their past failures and projects and i felt lucky to be here at a stage where things are more stable and where growth is skyrocketing, but also wished i was part of the early experience and just building fun experimental features and making mistakes. i'm at a point where i feel like i'm the new guy that doesn't fit in anywhere, and it's hard to bond when i have little in common with everyone. but i'm sure i will find my place eventually. insecurity and overthinking stems from thinking too much about myself. i just need to relax and by myself.