EvaluatedApplications.AlgFormer 2.13.0

ILGPU/CUDA acceleration for the AlgFormer package's HoloFormer model: GpuHoloTrainer runs batched GPU forward/backward training (CPU Adam by default, with an opt-in fused on-GPU Adam step) and GPU inference (Logits, StackIterLogits). GpuDevice auto-detects a CUDA device at startup and otherwise runs the SAME kernels on ILGPU's own CPU accelerator, so one build works with or without a CUDA card; genuinely over-length windows still run on the CPU HoloFormer directly. Every kernel is gradchecked against the CPU HoloFormer as its correctness oracle (float-close fp32, not bit-identical). Also ships an opt-in, default-OFF "ADAM & EVE" per-layer directional gradient gate layered on the fused on-GPU Adam step, live in production training (not an unused experiment) since 2026-09-24, plus preference (DPO) training on the GPU: per-sequence log-probabilities (SequenceLogProb) and a signed-weight all-positions training step (TrainBatchAllPosWeighted), at the same weight-tied depth the trainer trains at, with a frozen reference model (NewReference); exact bit-identical GPU page transforms for Phasor's Pleroma key band (PleromaPageTransform: Forward, Inverse, Dft over a 61-bit or sub-40-bit prime field, with an ILGPU CPU and a managed backend) and the Pleroma Cascade Newton pass (PleromaNewtonGpu); and a deprecated non-holo GpuModel/GpuTrainer twin for the older softmax-attention AlgFormer core. Add this package only if you want GPU acceleration; the core AlgFormer package stays dependency-free and AOT/trim-safe without it.

Grammar-driven synthetic corpus generator: mines nouns, verbs and adjectives and the syntax that joins them from real text with the rules of grammar, stores them in HoloDb, and recombines them into new grammatical sentences built only from words and pairings the text used, plus reading-comprehension Q&A pairs (single-sentence, multi-sentence passage, and two-turn). Rules-first (deterministic, zero-training) part-of-speech tagging, chunking and role assignment; mining fans out across cores through EvalApp (generation does too, opt-in); an optional mini-HoloFormer scorer re-ranks generated candidates by next-lemma likelihood once trained (POS tagging stays 100% rules-based).

Beta: A production standard for AlgFormer/HoloFormer models: an append-only lineage journal that answers "what was true at this point in training" without ever re-deriving it from current state, a confidence-gated decode policy (resonance-floor or nucleus sampling, with a degeneracy guard), and a growth surface that grows a model's context, layers, shifts, or vocabulary correctly by construction while stamping the journal automatically. For teams running AlgFormer/HoloFormer models in production who want checkpoint lineage and decode behaviour to stay correct through every growth step, not re-derived from whatever is live at the time.

An always-learning neural concurrency tuner for EvalApp. A small holographic policy model (HoloFormer, d48/L2/K2/shifts8) replaces the built-in heuristic on your app's resource gates and keeps learning from its own decisions, so it re-converges when the best concurrency moves. Ships with a warm-start model, makes at most one decision per 25 ms, and keeps learning in production. Opt in with .WithNeuralTuning(), or pass in a model you own with .WithNeuralTuning(model).