How We Launched YandexGPT
summary
A talk describing how YandexGPT was taken from pretraining to product release, outlining the stages of model development and the practical problems encountered. It explains debugging and fixing a bug in the NCCL distributed-communication framework and how that fix accelerated pretraining by about 30%.
AI-written summary · about
chapters (8)
- 0:00YandexGPT introduction
- 2:55LLM resource intensity issues
- 6:31Megatron-LM framework problem
- 9:02Efficient fine-tuning methods
- 13:59Model inference optimization
- 15:25Knowledge distillation
- 18:55Model quantization
- 22:47Conclusion & Q&A
related talks
45mHow to use GPT in your projects: A Case Study
48mHow we brewed the Gigachat Pretrain data
43mState of GPT
23mUnveiling the Journey of NLP
28mGPT or how I learned to stop worrying and love the generative AI
02h 00mChatGPT & Prompt Engineering
43mTraining GigaChat MAX
35mChatGPT is lying, how can we fix it?