fix(protrain): auto-offload memory-bound VL models; compose lora_mlp_kernel with offload#39

thad0ctor merged 2 commits into

fix/protrain-costmodel-followupsthad0ctor/axolotl:fix/protrain-costmodel-followupsfrom

fix/protrain-vl-automode-offloadthad0ctor/axolotl:fix/protrain-vl-automode-offloadCopy head branch name to clipboard