nn — Neural Network primitives for Almide

Transformer + signal-processing library written in Almide, built on the matrix stdlib.

Status: preview. Used to drive and validate matrix runtime work; Whisper tiny inference runs end-to-end (native + WASM).

Modules (src/)

File	Purpose
tensor.almd	Tensor load helpers (GGML f16/f32/f64 decoders)
gguf.almd / ggml_whisper.almd / whisper_loader.almd	GGUF/GGML weight loaders for Whisper
mel.almd / wav.almd / fft.almd	Audio preprocessing: WAV decode, mel spectrogram, FFT
activations.almd	gelu / softmax / layer_norm wrappers
attention.almd	Single- and multi-head attention, masked variants
transformer.almd	Encoder / decoder blocks (residual, LN, MLP, attention)
generate.almd	Autoregressive decode loop
whisper.almd	Full Whisper encoder+decoder pipeline
tokenizer.almd	BPE tokenizer for Whisper

Benchmarks (examples/)

Micro + end-to-end matmul benchmarks used to compare Almide against NumPy f32/f64:

• _bench_matmul.almd, _bench_matmul_small.almd, _bench_sweep.almd — matmul throughput 3²…1024²
• _bench_matmul_f32*.almd, _bench_fused_all.almd — f32 path and fused α·A·B
• _bench_linear_f32.almd — chained linear layers
• _bench_mlp_f32.almd — full MLP: gelu(X @ W1 + b1) @ W2 + b2
• _bench_attn_f32.almd — full attention block: LN + QKV + softmax + out
• _bench_encoder_block.almd — Whisper encoder block end-to-end
• _bench_ops_breakdown.almd — per-op dispatch profiling

Demo:

• examples/browser_demo/ — Whisper running in the browser via WASM + custom WASI shim

Running

almide build src/whisper.almd -o whisper          # native
almide build src/whisper.almd -o whisper.wasm --target wasm
almide run examples/_bench_attn_f32.almd
almide test                                       # nn/src tests

Depends on

• almide/almide — the compiler and matrix stdlib (f32/f64 paths, cblas dispatch)