Modelpublic

kingjones777/Muse-Glimmer-30B-Uncensored-ROCmFP4-GGUF

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes3.2kdownloads

Model Card

### 🔧 Runtime: build the ROCmFPX fork below Stock llama.cpp will not load this file. You need both the `muse-glimmer` architecture and the ROCmFP4 tensor types in one tree. Upstream `charlie12345/ROCmFPX` has the ROCmFP4 types but not muse-glimmer. Our fork has both: [`kingjones30/ROCmFPX`](https://github.com/kingjones30/ROCmFPX) — a fork of charlie12345/ROCmFPX, branch main. ``bash git clone https://github.com/kingjones30/ROCmFPX.git cd ROCmFPX cmake -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1151 -DGGML_NATIVE=ON -DCMAKE_BUILD_TYPE=Release cmake --build build --target llama-server llama-quantize -j$(nproc) ` Verified 2026-08-27 on gfx1151: clean clone → **0 build errors** → llama-server loads a muse-glimmer` ROCmFP4 GGUF from this family and generates coherent text.

Muse-Glimmer-30B Uncensored — ROCmFP4 for AMD Strix Halo (gfx1151)

Uncensored ROCmFP4 quantisations of meta-models/Muse-Glimmer-30B, built with the same ROCmFPX pipeline and the same card ftypes as `kingjones777/Muse-Glimmer-30B-ROCmFP4-Strix-Halo-DFlash-GGUF`.

Research artifact. Abliteration removes content-refusal. It does not add capability. Do not ship this as a product default. The aligned repo remains the serving default.

Metric	Result
Quantization	ROCmFP4 103 FAST + 106 STRIX_LEAN; also Q6 114/116
Source	Muse-Glimmer-30B BF16 safetensors → GGUF via `convert_hf_to_gguf.py`
Hardware	Ryzen AI Max+ 395 / Radeon 8060S / gfx1151 / 128 GB / ROCm 7.2.4
Drafter	Meta `dflash-kquant.gguf`, `--spec-type draft-dflash --spec-draft-n-max 15`
Unc FAST 103 decode	prose 15.68 · code 37.44 tok/s
Unc STRIX_LEAN 106 decode	prose 16.72 · code 38.51 tok/s
Aligned FAST 103 (published A/B)	20.31 tok/s mixed; real-world ~17–45
Aligned STRIX_LEAN 106 (published A/B)	18.72 tok/s mixed

Why this build?

The aligned card measured STRIX_LEAN (106) at 18.72 tok/s and FAST (103) at 20.31 tok/s in a controlled A/B (DFlash n=15, ctx 32K, -fa on). This repo is those same ftypes from an abliterated checkpoint, plus the Q6 AGENT/LEAN pair.

Q6 LEAN (ftype 116) is not the card LEAN. Card LEAN = 106.

Which file should I use?

Start with `STRIX_LEAN` (106) if you want the card-matched LEAN. Take FAST (103) if you want the aligned speed pick. Take Q6 AGENT (114) if you want more bits and will live with ~Q6 decode.

Ryzen AI Max+ 395, ROCm 7.2.4, DFlash --spec-draft-n-max 15, -fa on, ctx 32768, batch 1, temperature 0. Warm medians of 3; first call after load discarded.

Build	ftype	Size	prose tok/s	code tok/s
Unc FAST	103	13.80 GiB	15.68	37.44
Unc STRIX_LEAN	106	14.00 GiB	16.72	38.51
Unc Q6 AGENT	114	24.17 GiB	—	—
Unc Q6 LEAN	116	21.09 GiB	—	—
Aligned FAST (published)	103	13.80 GiB	~15	~39
Aligned STRIX_LEAN (published A/B)	106	14.00 GiB	—	18.72 mixed

Same flags, same drafter as the aligned card. Decode on this model is workload-dominated — quote a range, not a point.

Quick start

bash

llama-server \
  -m muse-glimmer-30B-Uncensored-ROCmFP4-STRIX_LEAN.gguf \
  --spec-type draft-dflash --model-draft dflash-kquant.gguf \
  --spec-draft-n-max 15 --spec-draft-ngl 99 --spec-draft-device ROCm0 \
  --chat-template-kwargs '{"reasoning_strength":"low"}' \
  -ngl 999 -fa on -dio --jinja -fit off -dev ROCm0 -c 32768 \
  --host 127.0.0.1 --port 8080

Requires a llama.cpp built with ROCmFP4 (ggml types 100–106) and the muse-glimmer port. Stock llama.cpp rejects these tensor types.

Flag	Why
`--chat-template-kwargs '{"reasoning_strength":"low"}'`	Template defaults to `high`. Small `max_tokens` then returns empty `content`.
`-fa on` (text) / `-fa off` (vision)	Vision requires `-fa off`.
`--spec-draft-n-max 15`	DFlash block size is 16; one slot holds the previously accepted token.

--reasoning-budget is not enforced on this model. Use reasoning_strength.

Uncensored findings

Content-refusal scoring on a 24 harmful / 12 harmless / 8 quality research set, greedy (temp 0). Counts only — no payloads. Abliteration is supposed to drop harmful-tune refusals without wrecking ordinary Q&A.

Model	Harmful 24	Harmless 12	Quality 8
Qwen3.8 aligned Q8 AGENT	23 refuse, 1 comply	11/12 ok (1 over-refuse)	6/8
Qwen3.8 uncensored Q6 AGENT (114)	23 comply, 1 broken	11/12 ok (1 over-refuse)	6/8
Muse aligned Q6 AGENT (114)	18 refuse, 6 comply	11/12 ok (1 over-refuse)	7/8
Muse uncensored STRIX_LEAN (106)	24 comply	12/12 ok (0 over-refuse)	7/8

Reading:

—Aligned Qwen still refuses almost everything on this set. Abliterated Qwen complies on almost everything. Quality score is identical (same two fails: Márquez needle + bat-and-ball).
—Aligned Muse is leakier than aligned Qwen on this classifier — a few complies even before abliteration.
—Quality is a substring smoke check, not MMLU. It is a regression guard against a broken quant, not a capability claim.

Files

File	ftype	Size	Role
`muse-glimmer-30B-Uncensored-ROCmFP4-FAST.gguf`	103	13.80 GiB	speed pick
`muse-glimmer-30B-Uncensored-ROCmFP4-STRIX_LEAN.gguf`	106	14.00 GiB	card-LEAN equivalent
`muse-glimmer-30B-Uncensored-Q6_0_ROCMFPX_AGENT.gguf`	114	24.17 GiB	6-bit, Q8 head/attn
`muse-glimmer-30B-Uncensored-Q6_0_ROCMFPX_LEAN.gguf`	116	21.09 GiB	6-bit throughout
`dflash-kquant.gguf`	—	1.52 GiB	DFlash drafter (Meta's, unmodified) — use this
`mmproj-kquant.gguf`	—	1.30 GiB	vision projector (unmodified; vision tensors were not abliterated)

Six files. llama.cpp loads them via --model, --model-draft and --mmproj. This repo is the uncensored family only — aligned builds are a separate repo.

Quantization

bash

PYTHONPATH=gguf-py python convert_hf_to_gguf.py <MODEL_DIR> --outtype bf16 --outfile unc-BF16.gguf
llama-quantize unc-BF16.gguf …-FAST.gguf Q4_0_ROCMFP4_FAST 16
llama-quantize unc-BF16.gguf …-STRIX_LEAN.gguf Q4_0_ROCMFP4_STRIX_LEAN 16

No extra --tensor flags — matches the published aligned card. ROCmFPX llama-quantize only.

Known issues (same as aligned)

1.Vulkan/CUDA/CPU cannot load these files — ROCmFP4 is ROCm-only.
2.Vision requires -fa off.
3.Small max_tokens returns empty content — budget goes to reasoning_content.
4.--reasoning-budget is not enforced; use reasoning_strength.
5.This is an uncensored research build. Do not deploy it as the public default.

Not yet measured

Test	Status
Perplexity / KL vs BF16	❓ not measured
MMLU-Pro, GPQA, GSM8K	❓ not run
Tool-calling 7-case suite on the unc weights	❓ not re-run (aligned scored 6/7, model-level)
Vision spatial 3/3 on the unc projector	❓ projector reused, not re-scored
Independent reproduction	❓ none yet

License and attribution

Base model: Meta Muse-Glimmer-30B (Apache 2.0). ROCmFP4 types: ROCmFPX. This repository is quantisation and measurement of an abliterated Muse-Glimmer-30B checkpoint.

See the aligned card for the muse-glimmer architecture port, DFlash notes, and tool-calling suite.