Helpful PyTorch Tutorials
September 28, 2026 · IntelliSensei Team
Regulated data blocks most enterprise fine-tuning projects. Here is how DP-SGD with Opacus and federated LoRA actually work in PyTorch: threat models, DP-LoRA configuration, the hyperparameters that invert, FedAvg aggregation, and the accuracy you should expect to pay.
Read moreSeptember 21, 2026 · IntelliSensei Team
Reasoning models make the token bill variable. How to cap and force-close thinking, route easy traffic to a non-thinking path, plan capacity for a long-tailed output distribution, handle traces safely, and find the accuracy-versus-budget knee on your own evals.
Read moreSeptember 21, 2026 · IntelliSensei Team
How to move a PyTorch training or serving stack onto AMD MI300-class GPUs: auditing CUDA lock-in, capability-based device code, the dependency substitution table, HIPify vs Triton, RCCL tuning, and benchmarking cost per million tokens before you sign anything.
Read moreSeptember 21, 2026 · IntelliSensei Team
Three fine-tunes, one serving budget. How weight averaging, task arithmetic, TIES and DARE combine fine-tuned checkpoints into one model in minutes of CPU time — with PyTorch code, a mergekit recipe, LoRA-specific rules and the evaluation sweep that keeps a merge honest.
Read moreSeptember 18, 2026 · IntelliSensei Team
Chronos-Bolt and TimesFM will forecast a series they have never seen, in ten lines and no training run. Here is how zero-shot compares to a trained PatchTST-style global model, with rolling backtests, quantile loss, covariates and a decision table.
Read more