pxq_llama.cpp/NOTICE at main · poisonxa16/pxq_llama.cpp · 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
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
pxq_llama
=========

Copyright (c) 2026 PXA Network (https://pxanetwork.com)

pxq_llama is developed and maintained by PXA Network. It is distributed under
the MIT License; see LICENSE for the full text and for the copyright notices of
all upstream projects, which are preserved in full.


Lineage
-------

    ggml-org/llama.cpp          MIT   Copyright (c) 2023-2024 the llama.cpp authors
      ggml                      MIT   Copyright (c) 2023-2024 the ggml authors
        ikawrakow/ik_llama.cpp  MIT   Copyright (c) 2024-2025 the ik_llama.cpp authors
          pxq_llama             MIT   Copyright (c) 2026 PXA Network

This project forked from ikawrakow/ik_llama.cpp at commit 1520eda98056
("prompt cache: Fix assertion ... (#1913)", 2026-06-04) and has been developed
independently since.

The repository history is flattened, so there is no git merge-base with
upstream. To diff or cherry-pick against upstream, compare against that exact
commit. Upstream's README is preserved at docs/README-upstream-ik_llama.md.

Individual changes ported from upstream after the fork point are marked inline
in the source with the originating upstream pull request number, so their
provenance can be traced without relying on git history.


What PXA Network contributes
----------------------------

The following are original to this project and are the work of PXA Network:

  PXQ quantization family
    A set of PXA-native low-bit quantization tiers, their CPU and CUDA codecs,
    the quantizer tier-selection policy, and the GGUF metadata that describes
    them honestly (including deriving the reported tier from the tensor-type
    histogram rather than trusting an inherited file_type field).

  Pascal / Volta MoE acceleration
    An sm_60 fp16 GEMM gate fix, a flash-attention regime fix, MoE-path fixes,
    and correct np>1 hybrid concurrency. These apply to any MoE on these cards
    at any size, and scale from a single salvaged card through multi-card
    -sm layer splits to CPU/RAM offload.

  Correctness guards
    A structural graph-split guard for gated-DeltaNet hybrids, which produce
    degenerate output under -sm graph; the guard triggers on tensor names
    rather than architecture names, so it protects architectures that did not
    exist when it was written.

    An imatrix provenance gate, so a quantized artifact describes the
    importance matrix that was actually consumed rather than one merely
    supplied or inherited from its source file.

  Codebase ownership
    Removal of upstream kernel paths this project does not use, and the
    reowning of the retained CPU kernel surface under pxq-cpu.

Everything not listed above should be assumed to originate upstream and remains
the copyright of its respective authors.


Third-party components
----------------------

This project bundles third-party code retained from upstream, including
yhirose/cpp-httplib (MIT), used by llama-server. Their licenses and copyright
notices are retained in their respective source files.


Trademarks
----------

"PXA", "PXA Network" and "PXQ" identify this project and its maintainer. The
MIT License grants rights to the software; it does not grant permission to use
these names to endorse or promote derived works without permission. Nothing
here restricts factual reference to the project, including stating that a work
is derived from it.