ArtomYuan/Qwen3.8-27B-abliterated-ROCmFPX

Qwen3.8-27B-abliterated-Q4_ROCmFPX_FAST

ROCmFPX quantized version of huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF (locally requantized).

• Base model: Qwen3.8-27B (dense) abliterated (refusal behavior removed, uncensored)
• Quantization: BF16 → ROCmFPX (Q4_0_ROCMFP4_FAST, 4.25 bpw), --allow-requantize
• MTP: ✅ model has MTP head — generation speed +~50% when enabled (measured 12-13 → 17-19 t/s)
• Vision: text-only version (vision belongs to Ornith-abliterated — the ROCmFPX engine cannot run MTP and vision simultaneously; see halofpx docs)

⚠ Important: Format Notice

ROCmFPX is a proprietary quantization format of the halofpx (ROCmFPX) engine — upstream llama.cpp CANNOT load this file!

Load it with halofpx (register this GGUF in the halofpx registry, then POST /api/v1/load).

Files

File	Size	sha256
Qwen3.8-27B-abliterated-Q4_ROCmFPX_FAST.gguf	~13.6 GiB	see sha256.txt
mmproj-model-bf16.gguf	~889 MiB	see sha256.txt

Quantization Benchmarks

Measured on AMD Strix Halo (gfx1151), llama-bench, -p 256 -n 256 -t 16 -fa on -ngl 99, same prompt per row. (Dense 27B — all parameters active, so tg is inherently lower than MoE models.)

Variant	tg256 (t/s)	pp256 (t/s)	Size	bpw
Qwen3.8-27B-abliterated-Q4_0_ROCMFP4_FAST (this repo)	13.81	376.0	13.6 GiB	4.25
Qwen3.8-27B-abliterated-Q4_0_ROCMFP4_STRIX_LEAN	13.79	374.8	13.8 GiB	4.27
Qwen3.8-27B-abliterated-Q4_0_ROCMFP4 (base)	11.78	331.6	16.5 GiB	4.50
Qwen3.8-27B-abliterated-Q8_0_ROCMFPX	7.62	271.0	26.3 GiB	8.0

Benchmark environment: AMD Ryzen AI MAX+ 395 (Strix Halo, gfx1151, Radeon 8060S, 120 GB unified memory), ROCm/Vulkan backend, Mesa RADV.

Why only Q4_ROCmFPX_FAST is published (other variants not uploaded):

• Qwen3.8-27B-abliterated-Q4_0_ROCMFP4_STRIX_LEAN: measured identical to FAST (13.79 vs 13.81 t/s) — no advantage, not published
• Qwen3.8-27B-abliterated-Q4_0_ROCMFP4 (base, dual-scale): 15% slower than FAST with marginal precision gain — rejected
• Qwen3.8-27B-abliterated-Q8_0_ROCMFPX: 45% slower and 2× size on this dense model — to be published later as the quality tier

Usage

# After registering in halofpx registry (example):
curl -X POST http://127.0.0.1:8010/api/v1/load \
  -H "Authorization: Bearer ***" \
  -d '{"model_id":"qwen38-27b-abliterated","reasoning_mode":"off"}'

• Reasoning model: reasoning_mode: off outputs directly
• MTP enabled in run_config (mtp_enabled: true) — but MTP + vision simultaneously crashes the engine (known ROCmFPX limitation): use MTP for text-only, use Ornith for vision

License & Attribution

• Base model: Qwen3.8-27B (Alibaba) — Apache-2.0
• Abliterated version: huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF — Apache-2.0
• This quantized version: Apache-2.0 (ArtomYuan)

Disclaimer

This is an abliterated (uncensored) version and may produce outputs that do not conform to safety policies. Use at your own risk. No warranty is provided.

---

Qwen3.8-27B-abliterated-Q4_ROCmFPX_FAST（ROCmFPX）

基于 huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF 的 ROCmFPX 量化版（本地 requantize 产物）。

• 原模型：Qwen3.8-27B（密集架构）的 abliterated 消融版（去拒绝行为，无审查）
• 量化：BF16 → ROCmFPX（Q4_0_ROCMFP4_FAST，4.25 bpw），--allow-requantize
• MTP：✅ 模型带 MTP head，开启后生成速度提升 ~50%（实测 12-13 → 17-19 t/s）
• 视觉：纯文本版（视觉能力归 Ornith-abliterated——ROCmFPX 引擎 MTP 与视觉不可兼得，详见 halofpx 文档）

⚠ 重要：格式说明

ROCmFPX 是 halofpx（ROCmFPX）引擎专属量化格式——上游 llama.cpp 无法加载此文件！

请使用 halofpx 加载（halofpx registry 添加此 GGUF 后 POST /api/v1/load）。

文件

文件	大小	sha256
Qwen3.8-27B-abliterated-Q4_ROCmFPX_FAST.gguf	~13.6 GiB	见仓库 sha256.txt
mmproj-model-bf16.gguf	~889 MiB	见仓库 sha256.txt

量化基准测试

测试环境：AMD Strix Halo（gfx1151），llama-bench，-p 256 -n 256 -t 16 -fa on -ngl 99，每行同一 prompt。（密集 27B——全部参数激活，tg 天然低于 MoE 模型。）

变体	tg256 (t/s)	pp256 (t/s)	大小	bpw
Qwen3.8-27B-abliterated-Q4_0_ROCMFP4_FAST（本仓库）	13.81	376.0	13.6 GiB	4.25
Qwen3.8-27B-abliterated-Q4_0_ROCMFP4_STRIX_LEAN	13.79	374.8	13.8 GiB	4.27
Qwen3.8-27B-abliterated-Q4_0_ROCMFP4（基础版）	11.78	331.6	16.5 GiB	4.50
Qwen3.8-27B-abliterated-Q8_0_ROCMFPX	7.62	271.0	26.3 GiB	8.0

测试环境： AMD Ryzen AI MAX+ 395（Strix Halo 架构，gfx1151，Radeon 8060S，120GB 统一内存），ROCm/Vulkan 后端，Mesa RADV。

仅发布 Q4_ROCmFPX_FAST 的原因（其他量化不上传）：

• Qwen3.8-27B-abliterated-Q4_0_ROCMFP4_STRIX_LEAN：实测与 FAST 持平（13.79 vs 13.81 t/s）——无优势，不发布
• Qwen3.8-27B-abliterated-Q4_0_ROCMFP4（基础版，双 scale）：比 FAST 慢 15%，精度提升有限——弃用
• Qwen3.8-27B-abliterated-Q8_0_ROCMFPX：慢 45% 且体积 2 倍——待后续上传（高质量档）

使用

# halofpx registry 注册后（例）：
curl -X POST http://127.0.0.1:8010/api/v1/load \
  -H "Authorization: Bearer ***" \
  -d '{"model_id":"qwen38-27b-abliterated","reasoning_mode":"off"}'

• 思考模型：reasoning_mode: off 直接输出
• MTP 开启（run_config mtp_enabled: true）——但 MTP + 视觉同时启用会导致引擎崩溃（ROCmFPX 已知限制），纯文本场景开 MTP，视觉场景用 Ornith

许可证与归属

• 原模型：Qwen3.8-27B（阿里巴巴）— Apache-2.0
• 无审查版：huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF — Apache-2.0
• 本量化版：Apache-2.0（ArtomYuan）

免责声明

本模型为 abliterated（消融去审查）版本，可能生成不符合安全政策的输出。使用者自行承担全部责任。本仓库不提供任何保证。