---

license: apache-2.0

base_model:

• Jackrong/Qwopus3.6-27B-v2-MTP-GGUF

tags:

• gguf
• rocm
• amd
• strix-halo
• gfx1151
• rocmfpx
• magicquant
• hybrid-quantization
• mtp
• speculative-decoding
• qwen3_6

library_name: llama.cpp

pipeline_tag: text-generation

quantized_by: ROCmFPX

language:

• en

base_model_relation: quantized

---

Qwopus3.6-27B-v2-MTP — ROCmFPX × MagicQuant hybrid GGUFs (AMD-native, fork-only)

> ## ⚠️ These files do NOT load on standard llama.cpp

> They use AMD-native *_ROCMFPX tensor types from the experimental

> ciru-ai/ROCmFPX llama.cpp fork (build from source).

> If you want files that work with stock llama.cpp / LM Studio / Ollama, use the sibling repo:

> lmcoleman/Qwopus3.6-27B-v2-MTP-MagicQuant-GGUF.

ROCm-optimized versions of MagicQuant-optimized quants: each file reproduces a

MagicQuant evolutionary-search winner's per-tensor-group precision layout (selected by

measured perplexity + KL divergence against the BF16 base), re-expressed in ROCmFPX's

AMD-native tensor types via llama-quantize --tensor-type-file. Tuned for and benchmarked

on AMD Strix Halo (Radeon 8060S iGPU, gfx1151, unified memory).

Files

| File | Size | Layout source (measured PPL of the K-quant twin) | tg (t/s) | pp (t/s) |

|---|---|---|---|---|

| ...ROCMFPX-MQ-Q4.gguf | 14.63 GiB | MagicQuant Q4 tier (PPL 6.6596) | 11.63 | 249.9 |

| ...ROCMFPX-MQ-Q5.gguf | 20.69 GiB | MagicQuant Q5 tier (PPL 6.6314) | — | — |

| ...ROCMFPX-MQ-Q6.gguf | 23.73 GiB | MagicQuant Q6 tier (PPL 6.6235) | — | — |

Reference, same bench conditions: stock Q4_K_M 15.65 GiB → tg 10.85, pp 238.4;

the K-quant MagicQuant Q4 twin → tg 11.52, pp 236.2. **MQ-Q4 was the fastest artifact

of everything measured** on this hardware, on both generation and prompt processing.

\* llama-bench -p 128 -n 128 -r 3 -ngl 99 -fa 1, ROCmFPX fork build (gfx1151).

PPL figures are the measured values of the K-quant layout each file reproduces

(BF16 baseline 6.571, wikitext-2, 100 chunks, ctx 512); the ROCmFPX re-expression

was not separately PPL-measured.

MTP speculative decoding

The embedded MTP head is preserved. With the fork's llama-server:

llama-server -m Qwopus3.6-27B-v2-MTP-GGUF-ROCMFPX-MQ-Q4.gguf \
  -md Qwopus3.6-27B-v2-MTP-GGUF-ROCMFPX-MQ-Q4.gguf --spec-type draft-mtp \
  -c 8192 -ngl 99 -fa on -ctk q8_0 -ctv q8_0

1.46× generation speedup measured on the K-quant twin (81% draft acceptance);

comparable behavior expected here.

Notes

• Chat template embedded; no patching needed.
• Experimental upstream research build. Known-good commit these files were built and validated with:

git clone https://github.com/ciru-ai/ROCmFPX && cd ROCmFPX
git checkout 221402af8574faf652b101b6afe225a3f329561f

• Built with Foundry + MagicQuant: evolutionary per-group hybrid search (measured,

imatrix-weighted, KL-guarded), layouts exported to ROCmFPX types per group.

Vision (image input)

mmproj-F32.gguf is copied unmodified from

Jackrong/Qwopus3.6-27B-v2-MTP-GGUF — pair it with any text

quant here for image input:

llama-server -m Qwopus3.6-27B-v2-MTP-GGUF-ROCMFPX-MQ-Q5.gguf --mmproj mmproj-F32.gguf -c 8192 --port 8080