adds harmony format , configurable gpt-oss parameters, launch.sh logic , improved templates for legml gpt-oss training, dynamic results directory and improve model pushing

adds quantization configuration correctly

adds memory optimized configuration

adds gpt-oss support

adds sft , quantization, better readmes

attempt to fix bfloat16 issue

fix config bug

fix launch script deploy and refactors

adds formatting fix

attempts to resolve training arguments issue

adds no think tag correctly

solves oom error with more reasonable configuration

adds max perf script

solves adam to adam_torch optimizer issue

improves requirements and dependencies

adds A100 large experiments

first commit