EvaluatedApplications.AlgFormer 2.18.0

AlgFormer

The reusable phasor-transformer engine from PrismFormer: define a model, train it data-parallel through an EvalApp pipeline, and run inference (predict / full logits / streaming generation with a KV cache). Ships two param-for-param comparable attention cores — AlgFormer (softmax dot-product) and HoloFormer (holographic bind/bundle/unbind resonance, O(T·d) with O(1)-per-token KV-serve) — plus the supporting pieces migrated from the full PrismFormer app: a subword tokenizer, the frozen production spec (PrismSpec), HoloShape capacity-law sizing, data-source adapters, and MQTT-based distributed/ swarm training. Pure double[] + SIMD, no external ML runtime or native dependency.

Features

• Two param-for-param comparable attention cores: AlgFormer (softmax dot-product) and HoloFormer (holographic bind/bundle/unbind resonance).
• Data-parallel training (PrismTrainer, resource-gated and adaptively tuned via EvalApp) for full epochs over your dataset, or train the model directly (TrainStep/TrainEpoch/Train).
• Streaming inference with an incremental KV cache (NewCache/Prime/Step/Generate) for low-latency, close-to-constant-work-per-token generation.
• Checkpoint save/load (Serialize/Deserialize, Save/Load) with format versioning, plus in-place growth: extend context length, attention shifts, or vocabulary on an already-trained model without retraining from scratch.
• A capacity-law sizing module (HoloShape) for choosing model dimensions from measured, not guessed, tradeoffs.
• A built-in subword tokenizer and several ready-made data-source adapters (raw corpus windows, prompt/answer pairs, source mixing).
• Distributed/swarm training over MQTT, from full-model relay to data-free position-manifest relay.
• Built-in gradient-check oracles so the training math itself is verified, not just assumed correct.
• HoloFormer only, opt-in, default OFF: a per-layer Adam gate ("ADAM & EVE") that nudges a layer's update toward a codebook-derived target norm range instead of letting it drift unchecked. This is new and has not yet run a real training step — it does nothing unless you explicitly construct HoloFormer.EveOptions with a Mode and pass it to the gated Step/TrainEpochGated overloads.

Quick start

dotnet add package EvaluatedApplications.AlgFormer

using PrismFormer;

// Define a model — tokens are ints; bring your own tokenizer/encoding.
var model = new AlgFormer(vocab: 16, shifts: 4, layers: 2, maxContext: 6, dModel: 32, frozenPrefix: 0);

// Train — data is a list of (context tokens, target token) pairs.
var data = new List<(int[] Ctx, int Target)>
{
    (new[] { 1, 2, 3 }, 4),
    (new[] { 2, 3, 4 }, 5),
};
var trainer = new PrismTrainer(model);                 // data-parallel via EvalApp
double loss = trainer.TrainEpoch(data, batchSize: 2, lr: 5e-2, shuffleSeed: 1);

// Run inference.
var context = new[] { 1, 2, 3 };
int next        = model.Predict(context);                                    // argmax next token
double[] logits = model.LogitsFor(context);                                  // full logits
int[] generated = model.Generate(context, maxNewTokens: 5, temperature: 0.8, seed: 1);

Console.WriteLine($"loss={loss:F3} next={next} logits.Length={logits.Length} generated=[{string.Join(",", generated)}]");

Swapping in the holographic core is the same shape — construct a HoloFormer instead of an AlgFormer with matching dimensions, and train/serve it through the same API.

License

Proprietary — this package ships the compiled library only; source is not distributed. Every capability described above is free to use today; nothing in it is gated behind a tier or a license key. PrismTrainer's evalAppKey constructor parameter (and PrismEval.DefaultKey) exist but have no effect on which capabilities are available. See LICENSE. Copyright (c) 2026 Dongyang Stephen Chen, trading as Evaluated Applications. All rights reserved.