Gemma 4 E2B CPU inference in 700 lines of pure C.
An educational project made to understand how LLM inference works. The full inference path is implemented in one file without external libraries.
Measured on an AMD Ryzen 7 7700 using the default native build.
| Implementation | Prefill (pp512) | Decode (tg128) |
|---|---|---|
| gemma4.c (int8) | 638.86 ± 2.93 tok/s | 25.90 ± 0.01 tok/s |
| llama.cpp (Q8_0) | 276.32 ± 1.85 tok/s | 23.83 ± 0.01 tok/s |
Results are the mean ± sample standard deviation over 12 timed runs after one discarded warmup, with each runtime using its fastest tested thread count. Both were built natively for CPU and dynamically quantize matrix inputs to int8.
./run -m ./gemma4-E2B-int8.bin --bench 512 128
You need a CPU with AVX2, an OpenMP-capable C compiler, and make
. The model takes about 5.0 GB of disk space, and 8 GB of RAM is recommended.
Clone the repository and download the ready-to-run model:
git clone https://github.com/ryanssenn/gemma4.c
cd gemma4.c
python3 -m pip install -U huggingface_hub
hf download QmogAI/gemma4-e2b-int8 gemma4-E2B-int8.bin --local-dir .
On Linux:
make
./run -t 0 -n 256 "Why is the sky blue?"
On Windows, use a MinGW-w64 environment that provides gcc
, OpenMP, and make
:
make win64 WINCC=gcc
.\run.exe -t 0 -n 256 "Why is the sky blue?"
-m
sets the model path. The default isgemma4-E2B-int8.bin
.-t
sets the temperature. The default is1.0
. Use0
for greedy decoding.-n
sets the maximum number of tokens to generate. The default is1,024
.--bench
measures prefill and decode throughput.--dump-logits
writes prompt logits as float32 binary data.
The C runtime cannot read the original checkpoint directly. exporter.py
takes the tokenizer and language-model weights from the Hugging Face checkpoint and writes them in the exact layout used by gemma4.c
.
Matrix weights are stored as int8 with FP16 scales. Inputs to linear layers are dynamically quantized to int8 while the rest of the activations remain float32. The resulting file is about 5.0 GB (4.7 GiB).
To create it yourself instead:
python3 -m pip install -r requirements.txt
python3 exporter.py /path/to/gemma-4-E2B-it-qat-q4_0-unquantized -o ./gemma4-E2B-int8.bin
Python is only needed to export the model or run numerical validation. Once the .bin
file exists, inference runs entirely through the C program.
The implementation is validated against the Hugging Face Transformers reference implementation running Google's unquantized Gemma 4 E2B QAT checkpoint in BF16. Both implementations process the complete Éva Gauthier article from the WikiText-103 validation split under teacher forcing. The passage contains 2,387 text tokens, and the added BOS token brings the comparison to 2,388 model positions.
| Metric | Result |
|---|---|
| Top-1 agreement | 2,305 / 2,388 (96.5%) |
| Mean KL divergence | 0.005207 |
An exact match is not expected because gemma4.c uses int8 matrix weights and linear inputs while the reference runs the unquantized checkpoint in BF16. The output distributions nevertheless remain closely aligned.
Build the C runtime first (make
or make win64
). The complete validation peaks at about 10 GiB of RAM:
python3 validation.py
Pass a smaller token count for a quicker check:
python3 validation.py 64
validation.py
uses the runtime's --dump-logits
flag to collect float32 logits after each prompt position. The flag can also be used directly when comparing gemma4.c with another implementation:
./run -m ./gemma4-E2B-int8.bin --dump-logits "Why is the sky blue?" > logits.bin
gemma4.c
contains the tokenizer, model definitions, kernels, transformer, KV cache, and generation loop.exporter.py
converts the original checkpoint into the binary layout read by the C runtime.validation.py
compares the runtime's logits with Hugging Face Transformers.validation.txt
contains the WikiText-103 passage used for numerical validation.win.c
andwin.h
provide the small Windows memory-mapping compatibility layer.Makefile
builds the Linux or Windows executable.