jaywyawhare/C-ML has 119 GitHub stars and 19 forks. GitHot measured 0 stars in 24 hours and +1 in 7 days. Last checked 2026-11-15 15:49 UTC. Last push 2026-11-13.
What you can do
- Layers - 28: Linear, Conv1d/2d/3d, ConvTranspose, RNN/LSTM/GRU, Transformer, BatchNorm, LayerNorm, RMSNorm, Pooling, Dropout, PixelShuffle, ...
- LLM - LoRA/QLoRA, Flash Attention, GQA, ALiBi, RoPE scaling, paged + sliding-window KV cache, MoE, ring attention, logits-processor sampling, speculative decoding (draft-model + prompt-lookup), continuous-batch serving, LLaMA 7B-70B
- Training - 11 optimizers, 13 losses, 8 LR schedulers, gradient checkpointing, DDP, pipeline & tensor parallel
- Compiler - IR fusion (5 patterns), linearization, codegen to C/PTX/SPIR-V/WGSL/MSL, AOT, JIT, kernel cache
Measured star growth
- in 24h
- 0
- in 7 days
- +1
- in 15 days
- Update delayed
- in 1 month
- Update delayed
Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.
About
Machine learning framework written in C.
Machine LearningREADME.md
Original
Machine learning in pure C. No frameworks. No runtime. Just code.
C-ML is a from-scratch C11 ML library. Tensors, autograd, 28 layers, optimizers, a compiler that fuses ops and emits GPU kernels (PTX/SPIR-V/WGSL/MSL), distributed training, LLM inference, and model I/O for GGUF/ONNX/SafeTensors. Zero required dependencies.
Build
git clone https://github.com/jaywyawhare/C-ML.git
cd C-ML && mkdir -p build && cd build
cmake -DBUILD_EXAMPLES=ON -DBUILD_TESTS=ON .. && make -j$(nproc)
Usage
#include "cml.h"
int main(void) {
cml_init();
Dataset* ds = cml_dataset_load("iris");
dataset_normalize(ds, "minmax");
Dataset *train, *test;
dataset_split(ds, 0.8f, &train, &test);
Sequential* model = cml_nn_sequential();
DeviceType dev = cml_get_default_device();
DType dt = cml_get_default_dtype();
model = cml_nn_sequential_add(model, (Module*)cml_nn_linear(4, 16, dt, dev, true));
model = cml_nn_sequential_add(model, (Module*)cml_nn_relu(false));
model = cml_nn_sequential_add(model, (Module*)cml_nn_linear(16, 3, dt, dev, true));
Optimizer* opt = cml_optim_adam_for_model((Module*)model, 0.01f, 0.0f, 0.9f, 0.999f, 1e-8f);
cml_nn_module_set_training((Module*)model, true);
for (int epoch = 0; epoch < 100; epoch++) {
cml_optim_zero_grad(opt);
Tensor* out = cml_nn_module_forward((Module*)model, train->X);
Tensor* loss = cml_nn_mse_loss(out, train->y);
cml_backward(loss, NULL, false, false);
cml_optim_step(opt);
tensor_free(loss);
tensor_free(out);
}
cml_cleanup();
}
Features
| Layers | 28: Linear, Conv1d/2d/3d, ConvTranspose, RNN/LSTM/GRU, Transformer, BatchNorm, LayerNorm, RMSNorm, Pooling, Dropout, PixelShuffle, ... |
| LLM | LoRA/QLoRA, Flash Attention, GQA, ALiBi, RoPE scaling, paged + sliding-window KV cache, MoE, ring attention, logits-processor sampling, speculative decoding (draft-model + prompt-lookup), continuous-batch serving, LLaMA 7B-70B |
| Training | 11 optimizers, 13 losses, 8 LR schedulers, gradient checkpointing, DDP, pipeline & tensor parallel |
| Compiler | IR fusion (5 patterns), linearization, codegen to C/PTX/SPIR-V/WGSL/MSL, AOT, JIT, kernel cache |
| GPU | CUDA, ROCm, Vulkan, Metal, WebGPU, OpenCL - userspace drivers for NV (RM ioctl) and AMD (KFD) (see maturity below) |
| Runtime | LLVM JIT by default (shape-specialized SIMD; interpreter fallback), multi-dtype compute (f32/f64/f16/bf16/int), BLAS, TLSF allocator, memory pools, thread pool |
| I/O | GGUF, SafeTensors, ONNX, PyTorch .pth, int8/int4/NF4 quantization |
| Python | CFFI bindings, NumPy integration, operator overloading |
Backend maturity
GPU paths are exercised by CI only through the CPU reference backend, mocks, and emulators - driver-call correctness needs the real hardware. Status per backend as of 2026-08:
| Backend | Maturity | Notes |
|---|---|---|
| CPU (SIMD/BLAS/LLVM JIT) | stable | default; full test suite runs on it |
| HCQ queue/signal/pipeline | stable | uniform ops dispatch incl. CPU reference (test_hcq_cpu, test_hcq_registry) |
| CUDA / ROCm / OpenCL / Vulkan / WebGPU adapters | build-verified | compile everywhere, dlopen at runtime, fail cleanly without hardware; driver bodies unvalidated on-GPU |
| NV (RM ioctl) / AM (KFD) drivers | experimental | full mock-based tests (nv_mock, am_mock); real ioctl paths need hardware |
| Adreno / Hexagon / Thunder / USB | placeholder | identity/generic kernels or stubs; not for production |
| Multi-GPU (peer copy, device placement) | simulated only | DEVICE_SIM_GPU device simulator |
Architecture
user code C or Python
|
cml.h 28 layers, optimizers, losses, LLM ops, serving
|
autograd dynamic graphs, checkpointing, distributed
|
tensor ops broadcasting, SIMD, BLAS
|
compiler IR -> schedule -> linearize -> codegen -> cache
|
drivers HCQ, NV (RM ioctl), AM (KFD), TLSF, pools
|
hardware CUDA, ROCm, Vulkan, Metal, WebGPU, OpenCL, Adreno, Hexagon, CPU
Testing
180 C test programs plus Python pytest run in CI: cross-dtype conformance
(77 ops x 8 dtypes vs f32 references), a VJP sweep against finite differences,
an API-contract sweep over every public surface, and HCQ/quantization/matmul
exactness tests. Coverage baseline is ~58% line / ~38% branch; the gaps are the
hardware-gated gpu/* backends and the f32 SIMD paths. See the coverage build
in tools/coverage_report.sh.