tensorcore/CITATION.cff at master · tsotchke/tensorcore · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
cff-version: 1.2.0
message: "If you use tensorcore in academic work or technical writing, please cite it."
title: "tensorcore: CUDA-equivalent tensor-core acceleration for Apple Silicon"
abstract: >-
  tensorcore is a C-ABI kernel library that turns Apple Silicon GPUs into
  a training-grade foundation for AI workloads. It provides simdgroup_matrix
  floating-point simdgroup_matrix GEMM plus an int8 MPS fallback, fused
  FlashAttention forward and backward
  (including GQA, sliding window, and ALiBi), Q4_0/Q8_0 quantized inference
  with a GGUF reader, the full transformer training kernel set
  (RMSnorm / LayerNorm / RoPE / SwiGLU / softmax / AdamW), Conv2D, and
  single-host distributed primitives. The same library binary runs on
  every M-series chip from M1 (Apple7) through M5 (Apple10), with the M5
  TensorOps path lit up via mpp::tensor_ops when SDK 26.0+ is available.
type: software
authors:
  - name: "tsotchke"
license: MIT
repository-code: "https://github.com/tsotchke/tensorcore"
url: "https://github.com/tsotchke/tensorcore"
keywords:
  - apple-silicon
  - metal
  - gemm
  - flashattention
  - llm-inference
  - mixed-precision
  - cuda-alternative
  - simdgroup-matrix
  - tensor-ops
  - gguf
preferred-citation:
  type: software
  title: "tensorcore: CUDA-equivalent tensor-core acceleration for Apple Silicon"
  authors:
    - name: "tsotchke"
  url: "https://github.com/tsotchke/tensorcore"