AGPT — Aggregated-Gradient Pretraining

Research project on aggregated-gradient pretraining for autoregressive language models. Trains a transformer on a prefix-trie representation of the corpus, factoring the gradient over branching subtrees rather than sliding context windows. Built on top of the µGPT Crystal/CUDA components kit.

What's here

• Paper — the gradient-factorization theorem, memory-scalable implementation, and empirical results on Shakespeare.
• CUDA training engine (src/cuda/agpt_train.cu, bin/agpt_train) — the GPU trainer. Radix-compressed trie input, per-subtree KV-cache scoping, bigram partitioning, auto-LR scaling, frequency-based pruning, and several sampler modes (L1 uniform, L2 root-child uniform, L3 mass-weighted, L4 path).
• Trie builders — bin/agpt_build_index produces a leveled per-depth trie from a corpus; bin/agpt_build_radix compresses unary chains into multi-character edges.
• Wrap-around corpus synthesis (bin/synth_wrap_corpus) — sample arbitrary-length token sequences from a depth-D trie via leaf→root wrapping with bridge-token sampling.
• Diagnostic tools — bin/radix-verify, bin/trie-profile, bin/bayesian-posterior, bin/convergence, bin/check_weights.

Building

shards install              # resolves the µGPT shard dependency
just build-all              # AGPT-native binaries
just build-microgpt-tools   # reference binaries from the µGPT shard

CUDA kernels are sourced from the µGPT shard at lib/microgpt/. nvcc on PATH (or at /opt/cuda/bin/nvcc) is required for the GPU trainer.

Dependency boundary

AGPT is its own repo, but it intentionally depends on the µGPT shard for:

• model/runtime primitives (Mat, MiniGPT, backends, RoPE)
• shared CUDA kernels / stubs
• reference comparison tools (bin/microgpt, bin/perplexity)

AGPT owns the trie/radix code, AGPT trainers, AGPT research tools, notes, and rnd/ experiment history.

Quick start (Shakespeare, depth 32)

# 1. Build a depth-32 leveled trie from the corpus.
bin/agpt_build_index --corpus data/input.txt --max-depth 32

# 2. Compress it to a radix trie.
bin/agpt_build_radix --leveled /tmp/agpt_input_d32

# 3. Train.
cp data/input.random.model /tmp/run.model
bin/agpt_train \
  --model /tmp/run.model --trie-dir /tmp/agpt_input_d32_radix \
  --save /tmp/run.model --epochs 3 --lr 3e-3 \
  --optimizer rmsprop --rmsprop-beta 0.999 \
  --lr-schedule warmup-cosine --warmup-epochs 1 \
  --entropy-lambda 1.0 --mass-weight linear --no-accumulate

# 4. Evaluate held-out perplexity.
bin/perplexity --model /tmp/run.model --file data/input.txt \
  --max-positions 4096 --backend openblas

Tests

just test          # AGPT-native specs + AGPT foundational parity tests
just test-crystal  # AGPT-native Crystal specs
just test-agpt     # AGPT foundational parity tests

Foundational tests require bin/microgpt and bin/perplexity from the µGPT shard — just build-microgpt-tools builds them.

Layout

src/agpt/        Crystal: trie, radix, samplers, KV store, walkers
src/cuda/        agpt_train.cu (GPU trainer; kernels.cu lives in µGPT)
src/tools/       Crystal CLIs (builders, synthesis, diagnostics)
spec/            Crystal specs for AGPT-only modules
tests/           Foundational shell tests
notes/      Design notes, paper drafts, status
notes/grants/    Grant pitch
rnd/             Research logs (per-experiment subdirectories)

License

Released under the PolyForm Noncommercial License 1.0.0 — see LICENSE. Academic and research use is permitted and encouraged. Commercial licensing available — see COMMERCIAL_LICENSE.md or contact transfire@gmail.com.