kasahare 0.1.0

A self-contained LLaMA-style inference engine with a Python binding.

simple-llm / kasahare

kasahare (which means "speak fast" in Twi) is the PyPI distribution name for simple-llm. A small, self-contained C++17 inference engine for LLaMA-style transformer language models. No external dependencies beyond the C++ standard library and a recent g++ / clang++. Designed to be read end-to-end in an afternoon, then hacked on.

What's in the box

file	purpose
include/simplellm/simd.h	AVX2 + FMA kernels (matmul, RMSNorm, SwiGLU, vec ops)
include/simplellm/tensor.h	Lightweight matrix view + matmul / linear kernels
include/simplellm/config.h	Static model description
include/simplellm/tokenizer.h	Byte-level BPE tokenizer (no HF tokenizers dep)
include/simplellm/sampler.h	Greedy / temperature / top-k / top-p sampler
include/simplellm/model_data.h	On-disk model format reader / writer
include/simplellm/model.h	Public surface: Model::load, forward, generate
src/model.cpp	LLaMA block: RMSNorm, GQA attention, RoPE, SwiGLU
src/main.cpp	sl-llm CLI
tools/create_tiny_model.cpp	Synthesises a tiny model on disk for smoke testing
tools/bench.cpp	Micro-benchmark: make bench
tests/test_basic.cpp	End-to-end smoke test: make test

Architecture

For each block (LLaMA-2 style):

x_n              = x_{n-1}                                  # residual
x_n              = attn_norm(x_n)
q, k, v          = Wq x_n, Wk x_n, Wv x_n                  # no biases
q, k             = RoPE(q, k, pos)
attn             = softmax(q . K^T / sqrt(d_head)) . V
x_n              = Wo(attn) + x_n                           # residual
x_n              = ffn_norm(x_n)
x_n              = W2( silu(W1 x_n) * (W3 x_n) ) + x_n      # SwiGLU, residual

Optimisations:

• matmul and linear are tiled over the inner dim with 8-wide AVX2
  • FMA. Compiles to a single fused vfmadd231ps chain.
• rmsnorm, vec_add, vec_mul, silu_mul are SIMD-ised.
• A pre-allocated KV cache (KvSlot per head per layer) lets forward() be O(1) allocation cost.
• Tied input / output embeddings (LLaMA style) — one copy of the embedding matrix is read once per token.
• Decoder-time attention materialises only the per-step logits; we re-score every cached key on each new token. For long contexts a Flash-Attention style kernel would beat this by ~2x, but the current implementation is dramatically simpler.

The block uses GQA (grouped query attention) so it accepts a config where n_kv_heads < n_heads (or equal, for MHA, or 1, for MQA).

Build

make            # builds build/bin/{sl-llm, create_tiny_model, bench}
make test       # builds and runs the smoke test
make bench      # runs the benchmark
make clean

Requirements: g++ ≥ 9 (we use g++ 12.2), AVX2 + FMA support. Override the SIMD flags with make CFLAGS="-O3 -march=native" to autodetect, or make CFLAGS="-O3 -msse4.2" to disable AVX2.

Quick start

# 1) Generate a tiny model + tokenizer (~110 KB total)
./build/bin/create_tiny_model --out build

# 2) Run it
./build/bin/sl-llm \
    --model build/tiny.sllm \
    --tokenizer build/tiny.tok.txt \
    --prompt "hi" \
    --max-new 16 \
    --temperature 0.8

# 3) Show timing
./build/bin/sl-llm \
    --model build/tiny.sllm \
    --tokenizer build/tiny.tok.txt \
    --prompt "" \
    --max-new 64 --temperature 0 --timing

(The model produced by create_tiny_model has random weights; the output is deterministic per --seed but not meaningful text. Real text needs real weights — see "Plugging in a real model" below.)

CLI flags

--model <file.sllm>        model file
--tokenizer <file.txt>     tokenizer file
--prompt <text>            input text
--max-new <n>              number of tokens to generate (default 32)
--temperature <f>          sampling temperature, 0 = greedy (default 0.8)
--top-k <n>                top-k filtering (default 40)
--top-p <f>                top-p filtering (default 0.95)
--seed <n>                 PRNG seed (0 = random)
--show-config              print model config and exit
--no-print-prompt          don't echo the prompt before the generated tail
--timing                   print per-step timing to stderr
--stream                   flush stdout after each generated token
--help                     this message

On-disk model format (.sllm)

+---------------------------------------------------+
| magic   (u32 LE, 0x534C4C4D == "SLLM")             |
| version (u32 LE, currently 1)                      |
| blen    (u32 LE)                                   |
| config blob  (blen bytes, key=value text)          |
| weights blob (u32 × N floats, little-endian)       |
+---------------------------------------------------+

The config blob is a human-readable key=value text. Weights are stored in a single contiguous float32 array in this order:

token_emb                (vocab_size * dim)
for each block:
    attn_norm            (dim)
    wq                   (dim * dim)
    wk                   (kv_dim * dim)
    wv                   (kv_dim * dim)
    wo                   (dim * dim)
    ffn_norm             (dim)
    w1                   (hidden * dim)   # SwiGLU gate
    w3                   (hidden * dim)   # SwiGLU up
    w2                   (dim * hidden)   # SwiGLU down
output_norm              (dim)

tied_embeddings reuses token_emb for the output projection (LLaMA style). The file is mmap-friendly, so you can load a model lazily in a follow-up patch.

Tokenizer format

A line-oriented text file:

# simple-llm tokenizer
vocab_size <N>
bos <id>
eos <id>
token <id> "<csv-escaped-payload>"
merge <rank> "<csv-escaped-left>" "<csv-escaped-right>"

Payloads are byte strings; CSV escape handles ", \, control chars and arbitrary bytes via \xHH. The BPE merge rank is determined by file order (lower rank = higher priority). A tokenizer with zero merges falls back to a longest-match scan over the vocab, which is enough for byte-level pre-tokenization.

Performance

The numbers below are from make bench on a single CPU core (g++ 12.2, AVX2 + FMA, FP32). They are not competitive with llama.cpp — we don't have quantisation, batching, GPU, threading, or FlashAttention — but they show the hot path is reasonable.

config	tokens/s
dim=256, layers=4, heads=4, hidden=512, seq=128	~380
dim=384, layers=6, heads=6, hidden=1024, seq=256	~60
dim=512, layers=8, heads=8, hidden=1536, seq=256	~24

The decode loop is O(seq_len) per token, so a 2x longer context halves the tok/s. Most of the time is spent in the linear() calls for Q/K/V projection and FFN; an OpenMP / threadpool pass over those would be the biggest single win.

Python binding

The engine is also a pip install-able Python package available on PyPI as kasahare:

pip install kasahare

Build it as editable for development, or build a wheel for distribution:

# Editable install (rebuilds on C++ changes — the binding .so lives in
# python/simplellm/_core*.so and is patched in by setuptools).
pip install -e .

# Wheel (good for shipping to a cluster with no compiler).
pip install build
make py-wheel              # produces dist/simplellm-0.1.0-*.whl
pip install dist/simplellm-0.1.0-*.whl

Run the Python tests:

make py-test

Then in Python:

import numpy as np
import simplellm

m = simplellm.Model.load("tiny.sllm", "tiny.tok.txt")
ids  = m.tokenizer.encode("hello world", add_bos=True)
log  = m.forward(ids[0], reset_kv=True)         # -> np.ndarray (vocab_size,)
out  = m.generate(ids, max_new=32,
                  options=simplellm.SampleOptions(temperature=0.7, top_k=40))
print(m.tokenizer.decode(out))

# Streaming:
def on_token(tid): print(m.tokenizer.decode_token(tid), end="", flush=True)
m.reset_kv()
m.generate(ids, max_new=16, options=simplellm.SampleOptions(temperature=0.8), stream=on_token)

# In-memory models (no temp files, no paths):
m2 = simplellm.Model.from_bytes(open("tiny.sllm","rb").read(),
                                open("tiny.tok.txt","rb").read())

The binding is zero-copy on the C++ → Python transition for the logits array: forward() returns a fresh np.ndarray of shape (vocab_size,) that you can pass straight to np.argmax, torch.softmax, etc. projection and FFN; an OpenMP / threadpool pass over those would be the biggest single win.

Plugging in a real model

tools/create_tiny_model.cpp shows the on-disk layout. To use a trained model, the steps are:

1. Train or download a LLaMA-2 / Mistral / Qwen2 model.
2. Convert its HF / safetensors / GGUF weights to a single float32 buffer in the order listed above.
3. Write the config blob (a few key=value lines) with the matching dimensions.
4. Run sl-llm --model ... --tokenizer ....

We do not ship a converter for a specific upstream format; the layout is small enough that numpy.tofile() + a few header bytes does the trick. See tools/create_tiny_model.cpp for a working example of writing a model from scratch.

Limitations (intentional)

• FP32 only. No FP16, BF16, INT8 or INT4. Adding INT4 would 8x memory and ~3-4x decode throughput.
• Single-threaded. No SIMD beyond AVX2/FMA. No GPU.
• Decoder-only. No training.
• No Flash-Attention — attention is O(seq_len) per token.
• No KV cache compression (quantised cache, sliding window).
• No beam search. Sampling only (greedy / top-k / top-p / temperature).
• The tokenizer is byte-level BPE, not identical to any specific upstream BPE implementation — the merge-ranking and pre-tokenisation are simplified. A trained model from a specific corpus may need its own tokeniser file to round-trip correctly.
• The on-disk format is not portable across endiannesses; this is fine for x86 / ARM servers but not for big-endian targets.

Where to go from here

• Add a mmap-based weight loader so very large models don't need to live in RAM twice.
• Add INT8 / INT4 weight loading (block-quantised, like GGUF Q4_K).
• Add OpenMP parallelism to matmul (per-row across M).
• Replace the attention loop with a single fused kernel.
• Add a Stream API that yields tokens through a callback (handy for chat UIs).
• Add a HuggingFace tokenizer.json reader so trained tokenizers work out of the box.

License

Do whatever you want with it. No warranty — see "Limitations" above.