SPDP/NOTICE at main · AIDASLab/SPDP · 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
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
SPDP — Unified Static–Dynamic Pruning for Efficient LLM Inference
Copyright (c) 2026 Jinhyeok Kim, Yejoon Lee, Jaeyoung Do (Seoul National University).

This artifact contains code derived from, or vendored from, the third-party
projects listed below. Their original copyright notices and licenses apply to
the corresponding files and are retained in place.

--------------------------------------------------------------------------------
1. FastGEMV — https://github.com/wangsiping97/FastGEMV
   License: MIT
--------------------------------------------------------------------------------
SPDP's CUDA-core spMspV (GEMV) decode kernels are derived from FastGEMV. The
half-precision GEMV structure, the warp-level reduction scheme and the
thread/tile decomposition originate there; SPDP extends them with the Tiled-CBC
sparse format, bitmap-based dynamic pruning (HAD-SMBD) and Split-K reduction.

Affected files:
  extension-pytorch/csrc/spdp-kernel/csrc/spdp_kernel.cuh   (was fast_gemv_v3.cuh)
  extension-pytorch/csrc/spdp-kernel/csrc/SpMM_API.cu
  extension-pytorch/csrc/spdp-kernel/csrc/utility.cuh

--------------------------------------------------------------------------------
2. SpInfer — https://github.com/HPMLL/SpInfer_EuroSys25
   Copyright 2025 The SpInfer Authors. License: Apache-2.0
--------------------------------------------------------------------------------
Used both as an evaluation baseline and as the basis for SPDP's Tensor-Core
SpMM data-movement primitives. SpInfer itself derives from Flash-LLM (below),
and the vendored headers carry both copyright lines accordingly.

Affected files:
  extension-pytorch/csrc/spdp-kernel/csrc/Spinfer_utils.cuh
  extension-pytorch/csrc/spdp-kernel/csrc/Spinfer_API.cu
  extension-pytorch/csrc/spdp-kernel/csrc/AsyncCopy_PTX.cuh
  extension-pytorch/csrc/spdp-kernel/csrc/MMA_PTX.cuh

--------------------------------------------------------------------------------
3. Flash-LLM — https://github.com/AlibabaResearch/flash-llm
   Copyright 2023 The Flash-LLM Authors. License: Apache-2.0
--------------------------------------------------------------------------------
The ancestor of the SpInfer primitives above (asynchronous copy and MMA PTX
wrappers, themselves extended from NVIDIA CUTLASS), and an evaluation baseline.

Affected files:
  extension-pytorch/csrc/spdp-kernel/csrc/Flashllm_utils.cuh
  extension-pytorch/csrc/spdp-kernel/csrc/Flashllm_API.cu
  extension-pytorch/csrc/spdp-kernel/csrc/AsyncCopy_PTX.cuh
  extension-pytorch/csrc/spdp-kernel/csrc/MMA_PTX.cuh

--------------------------------------------------------------------------------
4. Other vendored components
--------------------------------------------------------------------------------
  TEAL              https://github.com/FasterDecoding/TEAL          (see TEAL/LICENSE)
  Wanda             https://github.com/locuslab/wanda               (see wanda/LICENSE)
  HF Transformers   https://github.com/huggingface/transformers     (Apache-2.0; see
                    transformers-spdp/LICENSE) — `transformers-spdp` is a fork adding
                    the SPDP sparse execution path.
  Sputnik           https://github.com/google-research/sputnik      (git submodule)
  glog              https://github.com/google/glog                  (git submodule)
  cuSPARSELt, cuBLAS, cuSPARSE — NVIDIA, subject to the NVIDIA Software License.
  SparTA / Flash-LLM / SpInfer / Sputnik / cuSPARSE kernels are included solely as
  evaluation baselines.