Accelerate PyTorch

Kwker makes PyTorch faster on the CPU at four levels. Each is one change to your code; pick the one that fits how much you want to change.

Level	What you write	What runs on Kwker	Measured speed-up
Drop-in kernels	`kwker.torch_ops.install()` once	torch's own `sort`, `topk`, `unique`, `quantile`, optimizer steps, ... for every caller	each call about 9x torch's own (median of 97 real ML inputs, one thread)
torch.compile	`torch.compile(model, backend="kwker")`	linear layers, attention and convolutions, plus the sorting ops	language-model decoding 1.5-2.0x Inductor; float32 CNNs 1.3-2.3x eager (batch 1)
Whole-model runners	`KwkCNN(model)`, `KwkEncoder(model)`, `KwkDecoder(model, ...)`	the entire forward pass as one native call	KwkCNN 2.8x eager in float32; KwkDecoder bf16 4.2x float32 eager, the same tokens
Operators	`torch.ops.kwker.sort(x)` and friends	that call	the same 9x per call

All of it runs on the CPU, on x86-64 Linux for now (Installation has the wheel for your torch release). Tensors on other devices stay with PyTorch. The figures are from 4-core Intel Xeon servers; the full tables are at kwker.io/benchmarks.

Examples

Speed up a model you already have

Speed up inference with one line: install(), the same results, measured with kwker audit.
Compile with torch.compile: backend="kwker", never slower than Inductor on the graphs it tunes.
Lower precision with an accuracy check: bfloat16 and int8 on AMX and VNNI CPUs.
A faster model, step by step: drop-in kernels, then torch.compile, timed on one model.

Train on the CPU

Train faster on the CPU: fused Adam, AdamW and SGD steps, the same parameters bit for bit.

Run whole models natively

Generate text with a Hugging Face model: KwkDecoder, ahead of llama.cpp.
Sentence embeddings: KwkEncoder for BERT-family models.
Classify images: KwkCNN in float32 or calibrated int8.

Build with Kwker's operators

Sorting operators and layers: torch.ops.kwker, median pooling, k-winners-take-all, torch.export.
Accelerate JAX: the same calls under jit and vmap.

Next steps

Quickstart: PyTorch: five minutes from install to a checked speed-up.
Runtime controls: the switches for single features.