Helpful PyTorch Tutorials
September 14, 2026 · IntelliSensei Team
Why GPU utilisation is the wrong scaling signal, how to cut a multi-minute GPU cold start to seconds, when scale-to-zero actually pays, and how to track cost per million tokens for a PyTorch inference fleet.
Read moreSeptember 11, 2026 · IntelliSensei Team
Long contexts make the KV cache, not the weights, your capacity limit. How to size KV bytes per token, make prompts cache-friendly for prefix caching, tier KV to DRAM/NVMe, when prefill/decode disaggregation pays off, and where context parallelism fits on the training side.
Read moreSeptember 10, 2026 · IntelliSensei Team
Long multi-GPU runs fail. This tutorial covers checkpoint economics, sharded saves with torch.distributed.checkpoint, async staging that hides save cost, resharding on a different world size, elastic restarts with torchrun, and the determinism details that make a resume actually correct.
Read moreSeptember 9, 2026 · IntelliSensei Team
A complete PyTorch recipe for modern recommendation: a two-tower retrieval model with in-batch softmax and logQ correction, hard negative mining, a cross-feature ranker, offline metrics that predict online behaviour, and an ANN serving path.
Read moreSeptember 9, 2026 · IntelliSensei Team
How to turn an open PyTorch model into a reliable tool-calling agent: schema design, trajectory data, loss masking, constrained decoding in vLLM, a hardened runtime loop and trajectory-level evaluation.
Read more