Helpful PyTorch Tutorials
September 9, 2026 · IntelliSensei Team
Most input-bound training runs are fixable in a day: measure the loader on its own, reshard the dataset, stream it with a resumable StatefulDataLoader, tune workers and pinning, pack sequences and move decode to the GPU.
Read moreSeptember 9, 2026 · IntelliSensei Team
How speculative decoding cuts LLM latency without changing output quality: acceptance rates, draft models vs n-gram lookup vs EAGLE-3, vLLM config, a PyTorch implementation, and the throughput trade to watch.
Read moreSeptember 9, 2026 · IntelliSensei Team
A hands-on torch.profiler workflow: time steps correctly, classify the bottleneck into input, launch, kernel or communication bound, read a Perfetto trace, profile memory with the snapshot viewer and fix the dataloader.
Read moreSeptember 9, 2026 · IntelliSensei Team
A working GRPO post-training run in PyTorch with TRL: when RL beats SFT, how the group baseline replaces a critic, reward function design, vLLM-backed generation, and how to spot reward hacking.
Read moreSeptember 9, 2026 · IntelliSensei Team
Train a FLUX or SD 3.5 LoRA on a few hundred images: dataset and captioning, a flow-matching training loop, evaluation metrics that beat eyeballing, and per-request adapter serving.
Read more