File size: 3,148 Bytes
fe202c6 | 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 | NOTICE
Swift-1.5-Qwen3.8-27B huihui-style abliterated NVFP4 NInfer artifact
====================================================================
This repository contains a single-file NInfer engine artifact (NVFP4/FP8
mixed-precision weights, container v3, with vision tower, MTP head and a
DFlash2 drafter) plus the build records that make it reproducible.
Licence
-------
The Swift 1.5 checkpoint is distributed under the Swift Open License v1.0 (see
LICENSE in this repository); this artifact is a derivative of it, so the Swift
Open License v1.0 continues to apply to the Swift Contribution it contains.
The Qwen3.8-27B base-model portion is Apache-2.0 (Copyright 2026 Alibaba Cloud;
see LICENSE-APACHE-2.0). The abliteration delta is derived from
huihui-ai/Huihui-Qwen3.8-27B-abliterated (Apache-2.0). The DFlash2 drafter is
z-lab/Qwen3.8-27B-DFlash2 (see its model card for its licence). Free commercial
use is subject to the $1,000,000-per-fiscal-year gross-revenue threshold of the
Swift Open License v1.0. This is a description, not legal advice.
Statement of changes (Swift Open License v1.0, sections 4-5)
------------------------------------------------------------
The weights in this repository are MODIFIED relative to the Swift 1.5 checkpoint:
1. Abliteration (weights changed). 70 tensors of the language model were changed
by a per-tensor weight-difference transfer:
W_ablated = W_swift15 + (W_huihui - W_qwen) (fp32, rounded to bf16)
where W_qwen is Qwen/Qwen3.8-27B and W_huihui is
huihui-ai/Huihui-Qwen3.8-27B-abliterated. The changed tensors are the
self_attn.o_proj / linear_attn.out_proj / mlp.down_proj matrices of language
layers 17..51 (measured, not assumed; see ablation-report.json). Every other
tensor - including all vision (model.visual.*), MTP (mtp.*), lm_head and
embedding tensors - is bit-identical to Swift 1.5 (source-audit.json).
2. Quantization (numerical format changed). llm-compressor oneshot with the
NVFP4+FP8 allocation of unsloth/Qwen3.8-27B-NVFP4 (recipe/unsloth_qconfig.json),
32 calibration samples from ultrachat-200k, sequence length 2048.
3. Conversion (container format changed). Neroued/ninfer, official
tools.convert, recipe qwen3_8_27b_nvfp4, components text,vision,mtp,dflash2,
with the maintained tools/chat_templates/qwen3_8.jinja chat template.
Pinned inputs
-------------
ukisai/Swift-1.5-Qwen3.8-27b bc7a1e10b689648585a3ef41494c8d84cf77271a
Qwen/Qwen3.8-27B 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
huihui-ai/Huihui-Qwen3.8-27B-abliterated 739e3c5b89849f6c238ce1e5b70008612ae42cdd
z-lab/Qwen3.8-27B-DFlash2 50307d4c4cde6860d4eee73e2547cd786fe8e8a4
Neroued/ninfer bace20dc70249eed6402b66d4852c6c3f9612905
The refusal-rate evaluation code in recipe/ reproduces the keyword scorer of
p-e-w/heretic @ 3521f8648a0dccf6e12a92666862632235fac7e6
(heretic/scorers/keyword_rate.py, AGPL-3.0-or-later); the marker list is used
for comparability of numbers only.
The engine is built exclusively for NVIDIA Blackwell compute capability
12.0 (sm_120a).
|