How-to
Staged Model Pipeline
Compose Python-owned pretraining and post-training with immutable artifact handoff and native-backed local serving.
Read guideHow-to guides
Complete focused tasks in training, fine-tuning, memory management, and serving.
How-to
LoRA Fine-Tuning
Attach low-rank adapters, optimize only adapter parameters, compare ranks without test leakage, and merge for serving.
Read guideHow-to
QLoRA Fine-Tuning
Fine-tune LoRA adapters over an NF4-packed frozen base while preserving packed storage during computation.
Read guideHow-to
Incremental Serving
Run incremental generation with per-request decoder state, paged KV storage, and the local C or Python serving surfaces.
Read guideHow-to
Activation Checkpointing
Trade replay computation for lower retained autograd state at transformer-block boundaries.
Read guideHow-to
Resume training from an exact checkpoint
Capture, validate, and restore exact FP32 full-training or LoRA Adam state, RNG state, and built-in batch-source position.
Read guideHow-to
Convert model artifacts safely
Convert Riftco models through strict SafeTensors, Hugging Face-style, GGUF v3, and ONNX boundaries without hiding architecture incompatibility.
Read guideHow-to
Validate the Cloud TPU Backend
Run the Linux fake-PJRT source gate or the non-skippable real Cloud TPU hardware acceptance preset and record reproducible evidence.
Read guideHow-to
Troubleshooting
Diagnose configuration, build, backend, shape, training, artifact, Python loading, and documentation failures.
Read guide