Helpful PyTorch Tutorials
September 9, 2026 · IntelliSensei Team
Why MoE checkpoints break naive fine-tuning: how routing works, the load-balancing loss, adapter-based recipes that keep the router intact, expert-parallel sharding with FSDP2, and honest memory and throughput math for serving.
Read moreSeptember 9, 2026 · IntelliSensei Team
A practical PyTorch walkthrough of adapting an open VLM to your own documents: model and licence choice, chat-format datasets, LoRA on the language tower, TRL training, the evaluation metrics that matter, and serving with vLLM.
Read moreSeptember 9, 2026 · IntelliSensei Team
How to compress a large teacher model into a cheap production student: choosing the distillation type, building the transfer set with rejection sampling, the KL loss, training settings, agreement-based evaluation and quantizing afterwards.
Read moreSeptember 9, 2026 · IntelliSensei Team
When a custom GPU kernel is actually justified, how to write a fused RMSNorm in Triton, test and benchmark it honestly, and register it with torch.library so autograd and torch.compile treat it as a first-class operator.
Read moreSeptember 9, 2026 · IntelliSensei Team
How to build the evaluation harness that every fine-tune, RAG system and agent needs: golden datasets, deterministic metrics, calibrated LLM-as-judge scoring, bootstrap confidence intervals and a CI gate that blocks regressions.
Read more