A city pond-census intern fine-tunes a small head on Friday and cannot find the weights on Monday when Triton should serve them. What handoff did the project miss?
Select an answer to reveal the explanation.
Short Explanation
Friday’s refined head is gone on Monday if it lived only in RAM. Save it with save_pretrained, or an equivalent local folder, and reload that directory for Triton. Random retrains and NCCL topology files are not a checkpoint.
Full Explanation
A framework project persists the refined checkpoint in a reloadable folder, typically with save_pretrained, and serving starts from that artifact. Weights left only in Friday's RAM are gone on Monday. Retraining from random weights each week throws away the Friday work. An NCCL topology file is not a model artifact Triton can load.