← Back
Home › Systems › AI and ML infrastructure

nnue-training

NNUE training pipeline

Stockfish nnue-pytorch

SystemTechniques · 4Primitives · 27 nnue-training Alpha-beta searchgame-ai · medium: Minimax that skips branches which cannot change the result; with iterative deepening, PVS and aspiration windows it is the core of classical adversarial search.›Data loading pipelinesmachine-learning · medium: Turn files on disk into shuffled, sharded, resumable GPU batches fast enough that the accelerator never waits.›Knowledge distillationml-training · medium: Train a small student to match a large teacher's softened outputs (plus the hard labels) so it keeps most of the teacher's accuracy at a fraction of the cost.›Sequential probability ratio teststatistics · medium: Add up the log-likelihood ratio per observation and stop when it crosses a fixed bound; the least data of any test with the same error rates (for two point hypotheses).›Loss functionsmachine-learning · low: The scalar a model minimizes: cross-entropy for classes and tokens, MSE or Huber for values, each with a stable form and simple gradient.›Compact state encodingdata-structures · low: Pack a game position into a few fixed bytes (24 for chess: occupancy bitboard plus nibbles) for training data and dedupeBitboardsgame-ai · low: One 64-bit word per set of up to 64 items, so set algebra becomes shifts, ANDs and table lookups; magic bitboards give ray queries in one multiply.›NNUE incremental evaluationgame-ai · low: Efficiently updatable neural network: a huge sparse first layer whose output (the accumulator) is patched per move, then tiny int8 layers; fast neural eval.›Elo ratinggame-networking · low: One number per competitor, moved by K × (actual − expected score) after each pairwise result; expected score is logistic in the rating difference.›SIMD programminghardware-os · low: Process 4-16 lanes per instruction with AVX2, AVX-512 or NEON via Rust std::arch, with runtime dispatch and a portable fallback.›AdamWml-training · low: Adam's per-parameter adaptive steps with weight decay applied directly to the weights; the default optimizer for transformers and LLM pretraining.›Learning-rate warmup and cosine dec…ml-training · low: Ramp the learning rate up linearly for the first ~1% of steps, then decay it along a half cosine to ~10% of peak.›Quantization aware trainingml-training · low: Fake-quantize weights/activations in training with straight-through gradients so int8/int4 models keep accuracyWeight quantizationml-training · low: Store weights as 8/4-bit integers with per-group scales; decode is memory-bound, so 4-bit weights give about 3-4x tokens/s and fit 4x larger models.›Sparse matrices (CSR)numerics · low: Compressed sparse row stores only nonzeros as row pointers, column indices and values; SpMV streams them once, so it is memory-bandwidth bound.›Candidate ordering and pruninggame-ai · low: The heuristics that make alpha-beta and branch-and-bound fast: try likely-best candidates first (TT, captures, killers, history) and skip or reduce the rest (null move, LMR, futility).›Search time managementgame-ai · low: Allocate thinking time per move from clock, increment and search instabilityTransposition tablegame-ai · low: A fixed-size, lossy hash table of search results keyed by a state hash; gives ordering and cutoffs for repeated states, shared lock-free between threads.›Zobrist hashinggame-ai · low: A state's hash is the XOR of one random 64-bit key per (item, slot) fact that holds in it, updated in a few XORs per change.›Object storagecloud · low: S3-compatible blob storage at cents per GB-month with 11 nines durability; pick by egress price, request price and latency, not by storage price alone.›+ 11 more primitives

Similar systems

Share the most building blocks with this one.

self-play-training 26 %chess-engine 22 %model-compression 18 %