kingjones777/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-FAST-imatrix-GGUF

🔧 Runtime: build the ROCmFPX fork below

Stock llama.cpp will not load this file. You need both the qwen4exp architecture and the ROCmFP4 tensor types in one tree. Upstream charlie12345/ROCmFPX has the ROCmFP4 types but not qwen4exp. Our fork has both:

kingjones30/ROCmFPX — a fork of charlie12345/ROCmFPX, branch main.

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)

⚠️ Apply the bundled fix patches before cmake: qwen4exp-qsa-checkpoint-fix.patch always, plus qwen4exp-mtp-graph-fork.patch if you want --spec-type draft-mtp on this clone. Full steps further down.

Qwen3.8-Flash-Next-Uncensored — ROCmFP4 FAST imatrix GGUF — AMD Ryzen AI Max+ 395 / gfx1151

The importance-matrix-calibrated FAST build. Same recipe, same 4-bit size, same speed as the plain FAST build — measurably lower perplexity because quantization was weighted by a calibrated importance matrix.

⚠️ Research artifact. Refusal behaviour has been removed. This does not add capability — it removes guardrails. Use it deliberately, in a context where that is appropriate, and own the output.

Measured quality — held-out WikiText-2 raw, -c 512

build	PPL	Δ
plain FAST (no imatrix) — sibling repo	5.3465 ± 0.034	—
this — FAST imatrix	5.0337 ± 0.031	−5.9%

Both files are the same 4-bit recipe at 87.9 GiB / 4.27 bpw — the only difference is that this one's quantization was importance-weighted. imatrix moves quality, not speed: decode t/s is identical (same bits per weight), so this is graded on perplexity, not tok/s.

• Calibration corpus: bartowski calibration_datav3.
• ⚠️ The imatrix was computed on the 4-bit model, not BF16 — the 51.2B PLE table plus the 128 GB GTT ceiling blocks a BF16 forward pass on Strix Halo. A higher-precision imatrix source would likely gain a little more. It is still a real, measured −5.9%.

For reference: dropping the protected Q6_K head to a fully 4-bit head costs only +1.5% PPL (measured 5.1075 with imatrix) — small, now that imatrix calibration absorbs most of the head damage. This build keeps the protected Q6_K head.

Speculative decoding (MTP) — now working on this arch

The qwen4exp MTP graph shipped with a broken combiner (it mean-pooled the hyper-connection streams); draft-mtp acceptance sat near 0.36. This repo ships the fix as qwen4exp-mtp-graph.patch — apply it to the fork above and rebuild. With it, plus the stock Flash-Next MTP head from kingjones777/Qwen3.8-Flash-Next-MTP-Heads-GGUF:

• acceptance 0.94, 31.80 tok/s with --spec-type draft-mtp vs 24.9 tok/s no-draft (+27.7%) — warm 160-token completion at short context (-c 2048), neutral prompt, cache_prompt:false, same box/binary, drafting with the Q8_0 head (mtp-Qwen3.8-Flash-Next-Q8_0.gguf, in the heads repo). The Q6_K / Q4 heads were not benchmarked here and may land somewhat differently.

The bundled MTP head is stock Flash-Next (not uncensored). It only proposes draft tokens; the main model verifies every one, so it never alters this model's output — it just makes generation faster on content it can predict. Without the patch, plain decode is unaffected; only draft-mtp needs it.

llama-server -m Qwen3.8-Flash-Next-Uncensored-Q4_0-ROCmFP4-FAST-imatrix-00001-of-00003.gguf \
  -md mtp-Qwen3.8-Flash-Next-Q8_0.gguf --spec-type draft-mtp \
  --spec-draft-n-min 0 --spec-draft-n-max 1 --n-gpu-layers-draft 99 \
  -ngl 999 -fa on -np 1 -c 32768 --jinja

With the bundled checkpoint fix applied this is verified to 128K (see below); raise -c to suit your context. Without that patch, keep speculative decoding at ≤32K.

⚠️ Updated 2026-09-17 — re-download if you pulled it earlier. qwen4exp-mtp-graph.patch now carries the models.h and llama-model.cpp hunks it needs. The previous version applied cleanly but failed to compile ('graph_mtp' was not declared in this scope). The bundled patch matches the build steps on this card; for the other build path use qwen4exp-mtp-graph-fork.patch (if you build from a kingjones30/ROCmFPX clone), also bundled here.

Measured plain vs draft-mtp — median of 3 per cell, one binary, greedy, cache_prompt:false, 256 generated tokens, -c 2048, Q8_0 head, Uncensored STRIX_LEAN-imatrix weights, gfx1151 / ROCm 7.2.4 (2026-09-17):

workload	plain	--spec-draft-n-max 4	--spec-draft-n-max 1
reasoning	23.91	30.94 (+29%, acc 0.680)	31.94 (+34%, acc 0.945)
JSON output	23.99	28.31 (+18%, acc 0.597)	27.24 (+14%, acc 0.758)
code	24.09	21.56 (−10%, acc 0.422)	26.80 (+11%, acc 0.711)
long-document summary	23.80	20.36 (−14%, acc 0.352)	24.14 (+1%, acc 0.641)

⭐ Use --spec-draft-n-max 1. It did not lose a single workload here, and it wins most where the next token is predictable. n-max 4 pays for four draft forward passes per step, so it only wins when acceptance is high (reasoning, JSON) and is a genuine loss on code and long-document work. MTP also costs prefill speed, because the draft head processes the prompt too. The older +27.7% figure came from one reasoning-shaped prompt — it holds for that shape, not universally, so measure your own.

✅ Depth: draft-mtp is fixed and measured (2026-09-17)

The ≥64K wedge came from context-checkpoint restores leaving the QSA indexer cache (mem_idx) out of the checkpoint. The fix ships here as qwen4exp-qsa-checkpoint-fix.patch — it overrides state_write / state_read on llama_memory_hybrid_idx. Apply it with the build steps on this card even if you never use speculative decoding.

With it applied, --spec-type draft-mtp ran clean from 2K to 128K on gfx1151: 8 depth rungs, 864 context-checkpoint restores (2 of them prompt-cache rollbacks at 64K), 0 GPU faults, coherent output at every depth. Measured 2026-09-17 on Ryzen AI MAX+ 395 / ROCm 7.2.4 with the Uncensored STRIX_LEAN-imatrix weights + mtp-Qwen3.8-Flash-Next-Q8_0.gguf, -c 262144, --spec-draft-n-max 4, default context checkpoints. That 128K run used my own fork tree; the exact build steps on this card were verified to 16K.

⚠️ Still open: --spec-type ngram-mod at ≥64K has not been retested with the patch — the original field report (…-STRIX-GGUF#6, thanks @liusecret) was ngram-mod, so keep -ctxcp 0 -cpent -1 when you use it. And do not use speculative decoding of any kind on Vulkan/gfx1151 — acceptance collapses to 0.

A speculative replay stalled warning on ~2% of restores is expected and harmless: that is the server's livelock guard dropping one draft and decoding that token normally.

Recipe

Quantized from the BF16 weights published by orcarouter/Qwen3.8-Flash-Next-Uncensored — the abliteration work is theirs, not mine. Go star their repo.

tensor group	type
MoE expert weights (ffn_*_exps)	TYPE_101 (ROCmFP4, 4.251 bpw)
shared expert (ffn_*_shexp)	TYPE_101
attention (attn_*)	all TYPE_101
per_layer_token_embd.weight (PLE, 51.2B params)	TYPE_101
token_embd.weight	TYPE_101
output.weight (lm head)	Q6_K (protected)

The Q6_K head: output.weight is Q6_K, never 4-bit. Every sampled token passes through the lm head, so its quantization error lands directly in the argmax. Verified by exact tensor name after both quantize and split. With imatrix, a fully 4-bit head costs only +1.5% PPL (measured); the Q6_K head here keeps that last 1.5%.

Building a runtime that loads these files

Two patches, both in this repo: qwen4exp-on-rocmfpx-d3ca537.patch (arch enablement, 156 KB) and qwen4exp-mtp-graph.patch (draft-mtp fix).

git clone https://github.com/charlie12345/ROCmFPX.git
cd ROCmFPX && git checkout d3ca537
curl -LO https://huggingface.co/kingjones777/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-FAST-imatrix-GGUF/resolve/main/qwen4exp-on-rocmfpx-d3ca537.patch
git apply qwen4exp-on-rocmfpx-d3ca537.patch
curl -LO https://huggingface.co/kingjones777/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-FAST-imatrix-GGUF/resolve/main/qwen4exp-mtp-graph.patch   # optional, for draft-mtp
git apply qwen4exp-mtp-graph.patch
# the checkpoint fix also ships in this repo — apply it before configuring:
curl -LO https://huggingface.co/kingjones777/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-FAST-imatrix-GGUF/resolve/main/qwen4exp-qsa-checkpoint-fix.patch
git apply qwen4exp-qsa-checkpoint-fix.patch      # checkpoint safety at >=64K: apply this always
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)

Speed — Ryzen AI MAX+ 395, gfx1151, ROCm 7.2.4, full offload

Decode speed is identical to the plain FAST build (same bits, same layout) — measured there at 22.75 tok/s gen / 387.3 tok/s prompt / 63.3 GiB GTT (one fixed 6,963-token prompt, cache_prompt:false, median of 4 settled samples). This model's native max context is 262,144 and it runs there on a 128 GB box; the window is cheap (QSA caps KV), depth is what costs.

Files

Sharded to stay under HF's 50 GB limit. Point --model at the first shard.

file	size
Qwen3.8-Flash-Next-Uncensored-Q4_0-ROCmFP4-FAST-imatrix-00001-of-00003.gguf	44.70 GB
Qwen3.8-Flash-Next-Uncensored-Q4_0-ROCmFP4-FAST-imatrix-00002-of-00003.gguf	44.67 GB
Qwen3.8-Flash-Next-Uncensored-Q4_0-ROCmFP4-FAST-imatrix-00003-of-00003.gguf	5.06 GB
mmproj-Qwen3.8-Flash-Next-Uncensored-BF16.gguf	0.91 GB (vision tower)
qwen4exp-on-rocmfpx-d3ca537.patch	arch enablement
qwen4exp-mtp-graph.patch	draft-mtp graph fix

Usage

llama-server \
  --model Qwen3.8-Flash-Next-Uncensored-Q4_0-ROCmFP4-FAST-imatrix-00001-of-00003.gguf \
  --mmproj mmproj-Qwen3.8-Flash-Next-Uncensored-BF16.gguf \
  --host 127.0.0.1 --port 8080 \
  --n-gpu-layers 999 --flash-attn on --fit off \
  --ctx-size 131072 --threads 16 --jinja

Do not use --no-mmap (and do not use -dio). The PLE table is streamed from the file through the page cache; forcing it into anonymous memory gets the process OOM-killed with nothing in the server log.

Reproduction

imatrix : llama-imatrix -f calibration_datav3.txt -o unc.imatrix --output-format dat -ngl 999 -c 512 -b 512 -fa on -dev ROCm0
quantize: llama-quantize --imatrix unc.imatrix --output-tensor-type q6_K <BF16> <out> Q4_0_ROCMFP4_FAST 16
ppl     : llama-perplexity -m <this> -f wiki.test.raw -ngl 999 -fa on -dev ROCm0 -c 512   (NO -dio)

A number without its binary is a rumour — every figure above is measured on the runtime above.

Acknowledgements

charlie12345/ROCmFPX — defines the ROCmFP4 tensor formats; every file here was produced with its llama-quantize and runs on its runtime. MIT. The qwen4exp architecture is applied on top via the bundled patches. llama.cpp — engine, GGUF format, conversion tooling. AMD ROCm — the compute platform (ROCm 7.2.4, gfx1151). orcarouter — published the uncensored BF16 checkpoint this is built from; the abliteration is their engineering. Qwen team — the original base model. License qwen-community-1.0.