🔧 Runtime: build the ROCmFPX fork below

Stock llama.cpp will not load this file — it needs the ROCmFP4 tensor types. Build kingjones30/ROCmFPX (fork of charlie12345/ROCmFPX, branch main). This is a qwen3_5 model — the arch, its vision tower and its MTP draft head are all supported by the fork out of the box; no extra patches are needed.

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)

Qwen3.8-27B TURBO-Fable-Cold-Fusion-Heretic — ROCmFP4 FAST imatrix GGUF (Vision + MTP)

An importance-matrix-calibrated ROCmFP4 build of DavidAU's Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU — a dense qwen3_5 27B (attention + linear-attention layers). This is the full model:

• 🖼️ Vision — the complete Qwen3.5-VL tower (27-layer deepstack encoder) as mmproj-…-BF16.gguf. Verified: a red test image returned "Red" (1,094 prompt tokens, ~1,024 of them image tokens).
• ⚡ MTP speculative decoding — 27–33 tok/s on code / reasoning / JSON, up from ~13 plain. Two 1.68 GB Q4 draft heads are included (see below).

⚠️ Research artifact / uncensored. DavidAU's merge removes refusal behaviour. Use it deliberately and own the output.

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

Plain decode is ~13 tok/s: this is a dense 27B, so every token reads all ~15 GB of weights and Strix Halo's memory bandwidth caps it (the base Qwen3.8-27B does the same 13.0–13.2 on the same harness). MTP is the lever — the draft head proposes up to 4 tokens and the main model verifies them in one pass.

Same harness for every row: 4 fixed workloads, greedy, cache_prompt:false, 256 tokens, first run discarded, median of 3, identical binary and flags (see Reproduction).

Which head?

• Reasoning / agentic work → DAU's own head. It is the model's own restored MTP block, so it predicts DAU's reasoning better (draft acceptance 0.667 vs 0.593).
• General chat, code, JSON → the base Qwen3.8-27B head. It is ~1.6% faster averaged across the four workloads and noticeably better on prose. DAU is a merge of Qwen3.8-27B, so the base head is fully compatible (same arch, vocab and block layout).

Honest comparison with the base model: plain speed is identical. With MTP, DAU trails the base Qwen3.8-27B by ~9% overall and ~12% on reasoning — the merged weights are simply harder to predict (reasoning acceptance 0.59–0.67 vs 0.75 on the base). Free-form prose resists MTP on both models (~16–19 tok/s); every other workload roughly doubles or better.

Recommended command (the exact settings the table was measured with)

ROCBLAS_USE_HIPBLASLT=1 llama-server \
  -m  Qwen3.8-27B-TURBO-Fable-Cold-Fusion-Heretic-Q4_0-ROCmFP4-FAST-imatrix.gguf \
  -md mtp-Qwen3.8-27B-TURBO-Fable-Cold-Fusion-Heretic-Q4_0.gguf \
  --spec-type draft-mtp --spec-draft-n-min 0 --spec-draft-n-max 4 \
  --spec-draft-p-min 0.0 --spec-draft-ngl 99 \
  -dev ROCm0 -fa on -ngl 999 --no-mmap -fit off -np 1 \
  -ctk q8_0 -ctv q8_0 -cram 8192 -b 2048 -ub 1024 -t 16 --poll 100 -c 32768 \
  --repeat-penalty 1.05 --repeat-last-n 256 \
  --reasoning off --jinja

• For the base head, swap -md to mtp-Qwen3.8-27B-Q4_0.gguf.
• --spec-draft-p-min 0.0 matters — the default (0.75) throttles drafting; opening it was worth +47% in my sweep. -np 1 is required for MTP on gfx1151. The table uses --spec-draft-n-max 4; in an earlier sweep with the built-in head, 8 and 12 were slower. Optimal depth depends on the (model, head, box) combination — sweep 3–5 on your own machine.
• The numbers were measured with --reasoning off; drop it to let the model think.
• The MTP block is also inside the main GGUF (blk.64), so -md <the main file> works too — but it loads all ~15 GB a second time. Use one of the 1.68 GB heads.

Vision

Add the vision tower to the same command:

  --mmproj mmproj-Qwen3.8-27B-TURBO-Fable-Cold-Fusion-Heretic-BF16.gguf --image-min-tokens 1024

--image-min-tokens 1024 matters — Qwen-VL degrades below ~1024 image tokens (the server warns if you omit it). Send images via the OpenAI image_url (data-URI or URL) chat format.

Quality (held-out WikiText-2 raw, -c 512)

Honesty note: imatrix helps this model only −0.57% — much less than the −5.9% it gave the Flash-Next MoE. This heavily-merged 27B is far less imatrix-sensitive; the number is small but real. I also built a STRIX_LEAN imatrix tier (6.6944 — worse than FAST here) and did not ship it: STRIX_LEAN's advantage is upgrading a per-layer embedding table, and this model has none. FAST is the tier to run.

Recipe

Q4_0_ROCMFP4_FAST, importance-matrix weighted. 14.93 GB main file.

64 transformer layers + 1 trailing MTP block. Calibration: bartowski calibration_datav3, imatrix computed on the BF16.

Files

Single files — no sharding needed.

Reproduction

runtime : ROCmFPX fork build 264 (e771235), HIP+Vulkan build, run on -dev ROCm0, ROCm 7.2.4, gfx1151
          llama-server sha256 e860b763d5e8f496ddb4b01236d8a8f37ae09dda456b5d076ab3abe2d9deb9fc
          libllama.so  sha256 cdfc9d484e8860f91c29f5e4a2282c9c9a09c6247339aa9ae614f603fe0aabd7
env     : ROCBLAS_USE_HIPBLASLT=1
argv    : exactly the "Recommended command" above (-c 32768), single slot, nothing else loaded on the GPU
bench   : 4 fixed prompts (code / prose / reasoning / JSON), /v1/chat/completions, temperature 0,
          max_tokens 256, cache_prompt false, 1 warm-up discarded, median of 3 (plain: 1 run each)

convert : convert_hf_to_gguf.py <src> --outtype bf16                (text + MTP block)
convert : convert_hf_to_gguf.py <src> --mmproj                      (vision tower)
imatrix : llama-imatrix -m <bf16> -f calibration_datav3.txt -o dau.imatrix --output-format dat -ngl 999 -c 512 -fa on
quantize: llama-quantize --imatrix dau.imatrix --output-tensor-type q6_K <bf16> <main> Q4_0_ROCMFP4_FAST 16
head    : llama-quantize --imatrix dau.imatrix --output-tensor-type q4_0 --token-embedding-type q4_0 \
            --tensor-type eh_proj=q4_0 <bf16> <full-q4> Q4_0 16
          then copy token_embd, output, output_norm and blk.64.* (with all metadata, no renumbering)
          into a new GGUF → 18 tensors, same layout as the base Qwen3.8-27B head
ppl     : llama-perplexity -m <main> -f wiki.test.raw -ngl 999 -fa on -dev ROCm0 -c 512

sha256 of the draft heads: DAU's own 6ce06b967545f25220a10b671ffaa88d6dc2227aa561039ce79d8a1a47ca3193, base 051a1764cff8c4f3ee6ae8b00593a0364c7539c67fa50ffc58f3f96509fca38e (identical to the file in kingjones777/Qwen3.8-27B-ROCmFP4-STRIX-MTP-GGUF).

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

Acknowledgements

DavidAU — created the TURBO-Fable-Cold-Fusion-Heretic merge and restored its MTP head; all model behaviour is theirs. I only converted and quantized it. Qwen team — the Qwen3.8-27B architecture, and the base model the mtp-Qwen3.8-27B-Q4_0.gguf draft head comes from (apache-2.0). charlie12345/ROCmFPX — the ROCmFP4 tensor formats (MIT). llama.cpp — engine, GGUF, conversion + vision/MTP tooling. AMD ROCm — ROCm 7.2.4, gfx1151.