CHADROCK3.6-40B-Opus-Deckard-Uncensored-Thinking-NEO-CODE-Di-IMatrix-ROCmFP4

CHADROCK3.6 40B Opus Deckard is a ROCmFP4/MTP GGUF release of the Deckard 40B lineage for AMD Ryzen AI Max+ 395 / Strix Halo systems.

It keeps the public release identity, file name, and downloadable artifact under the Chadrock name:

CHADROCK3.6-40B-Opus-Deckard-Uncensored-Thinking-NEO-CODE-Di-IMatrix-ROCmFP4.gguf

The behavior lineage comes from DavidAU's Qwen3.6 40B Opus Deckard release, with the MTP graft and GGUF recipe from PiehSoft's Deckard MTP work. The Chadrock release converts that line into a Strix-focused ROCmFP4 runtime format and uses a tuned draft-MTP serving profile.

This GGUF will not run correctly with stock llama.cpp. It needs the pinned ciru-ai/ROCmFPX runner because the file uses ROCmFP4 tensor types that upstream llama.cpp does not currently understand. For image input, use the Vulkan-only ROCmFP4 build path shown below.

The model file is already provided here. You do not need to rebuild or quantize the model. Build the custom llama server once, download this GGUF, and run the profile below.

Why This Build Exists

This release is for people who want the 40B Deckard/Opus behavior line with native MTP speculative decoding and lower AMD unified-memory pressure than the available K-quants.

The goal is not a generic GGUF quant. The goal is a model/runtime pairing for Strix Halo:

• Deckard 40B Opus / Heretic / Uncensored / Thinking lineage
• MTP speculative decoding from the PiehSoft Deckard MTP recipe
• ROCmFP4 Strix Lean tensor recipe
• Vulkan target+draft serving on AMD Strix Halo
• 262144 token context target
• draft-mtp depth 2 profile
• lower GPU pressure than the Q4_K_M baseline in the local sweep

Model Lineage

Qwen3.6 27B / 40B compatible hidden-size family
  -> DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking
       behavior lineage:
         - Claude 4.6 Opus / Deckard style
         - Heretic / uncensored / thinking / NEO-CODE / IMatrix line
  -> PiehSoft/Qwen3.6-40B-Deckard-MTP
       adds:
         - GGUF MTP graft recipe
         - Deckard MTP quants
         - MTP head from the compatible Qwen3.6 27B MTP donor line
  -> CHADROCK3.6-40B-Opus-Deckard-Uncensored-Thinking-NEO-CODE-Di-IMatrix-ROCmFP4
       adds:
         - ROCmFP4 Strix Lean conversion
         - Chadrock AMD serving profile

Direct source links:

• DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking
• PiehSoft/Qwen3.6-40B-Deckard-MTP
• unsloth/Qwen3.6-27B-MTP-GGUF, used locally as the MTP donor reference for the compatible 27B MTP head line.

Technical Metadata

Vision Status

Vision is now locally validated for this exact ROCmFP4 GGUF with the included Qwen3.6 27B projector:

mmproj-Qwen3.6-27B-f16.mmproj

Validated command path:

• runner repo: ciru-ai/ROCmFPX
• pinned runner commit: 7aa484a2f0a504dc612a3d74a068024f3e6d6353
• legacy vision smoke-test commit: 4795079b0
• build mode: GGML_VULKAN=ON, GGML_HIP=OFF, GGML_CUDA=OFF
• runtime device: Vulkan0
• required flags: --mmproj mmproj-Qwen3.6-27B-f16.mmproj --image-min-tokens 1024 -fa off

Smoke test image result on June 15, 2026 through the live OpenAI-compatible endpoint:

- Red box
- Blue circle
- Green triangle
- Text: "RED BOX BLUE CIRCLE GREEN TRIANGLE"

The earlier failure mode was runner-side: older HIP-enabled Strix ROCmFP4 builds loaded the projector but produced wrong image descriptions. The model file and projector are usable when served by the Vulkan-only ROCmFP4 build above.

The projector is uploaded with a .mmproj extension so Hugging Face's GGUF parser keeps the primary model file as the main artifact. Pass it to llama.cpp like a normal mmproj file.

Speed: Baseline vs Chadrock

These are clean local server/API measurements on AMD Ryzen AI Max+ 395 / Strix Halo. The comparison uses the same custom llamaCHADROCK runner, the same prompt files, no prompt cache, one slot, 512 generated tokens, draft-mtp, n_max=2, f16 target/draft KV, -b 2048, -ub 512, -sm none, -ngl 999, and Vulkan0.

Baseline:

PiehSoft/Qwen3.6-40B-Deckard-MTP
Qwen3.6-40B-Deckard-MTP-Q4_K_M.gguf

Chadrock:

CHADROCK3.6-40B-Opus-Deckard-Uncensored-Thinking-NEO-CODE-Di-IMatrix-ROCmFP4.gguf

Uplift

On Strix Halo, "GPU pressure" here is peak_vram_used_bytes + peak_gtt_used_bytes. The iGPU has a small dedicated VRAM aperture plus a large GTT/unified-memory pool, so GTT is part of the real pressure picture.

The local benchmark artifact is:

/home/crown/bench-results/llama/context-sweeps/20260615T074356Z-deckard40-clean-baseline-vs-chadrock/api/results.sqlite3

These are local server measurements, not universal llama.cpp claims. MTP throughput depends on prompt shape, driver version, clocks, KV cache settings, draft depth, and draft acceptance.

Best Settings / Advanced Setup

For the pinned score-build runner notes and copy-paste ROCmFP4 setup commands, use the advanced Ciru setup page:

https://llm.ciru.ai/chadrock-rocmfpx/

The shared pinned runner build is:

ciru-ai/ROCmFPX commit: 7aa484a2f0a504dc612a3d74a068024f3e6d6353
historical score tag: chadrock-rocmfp4-mtp-scores-20260621

For this Deckard 40B release, use the vision-validated Vulkan-only profile below rather than an older HIP-enabled Strix runner:

backend: Vulkan0 target + Vulkan0 draft
context: 262144
batch / ubatch: 2048 / 512
target KV: f16 / f16
draft KV: f16 / f16
MTP: draft-mtp, n_max=2, n_min=0, p_min=0.0, p_split=0.10
serving: one slot, metrics on, --mmproj mmproj-Qwen3.6-27B-f16.mmproj
vision: --image-min-tokens 1024, flash attention off
sampler: temperature=0, top_p=0.95, top_k=20
reasoning: off, reasoning_format=none, reasoning_budget=0

The projector path was validated with a real image-answer gate. Do not use the older HIP-enabled runner for multimodal validation; it could load the projector but produced wrong image descriptions in local testing.

Run With llama-server

Build Charlie's custom llama.cpp once, download this GGUF, then run:

/path/to/rocmfp4-llama/build-vulkan-rocmfp4/bin/llama-server \
  -m CHADROCK3.6-40B-Opus-Deckard-Uncensored-Thinking-NEO-CODE-Di-IMatrix-ROCmFP4.gguf \
  --alias main \
  --host 0.0.0.0 \
  --port 8080 \
  --jinja \
  -c 262144 \
  -n 16384 \
  --reasoning off \
  --reasoning-format none \
  --reasoning-budget 0 \
  -ngl 999 \
  -fa off \
  -sm none \
  -dev Vulkan0 \
  --mmproj mmproj-Qwen3.6-27B-f16.mmproj \
  --image-min-tokens 1024 \
  -b 2048 \
  -ub 512 \
  -t 16 \
  -tb 32 \
  -ctk f16 \
  -ctv f16 \
  --ctx-checkpoints 0 \

  --checkpoint-every-n-tokens -1 \

  --spec-type draft-mtp \
  --spec-draft-device Vulkan0 \
  --spec-draft-ngl all \
  --spec-draft-type-k f16 \
  --spec-draft-type-v f16 \
  --spec-draft-threads 16 \
  --spec-draft-threads-batch 32 \
  --spec-draft-n-max 2 \
  --spec-draft-n-min 0 \
  --spec-draft-p-min 0.0 \
  --spec-draft-p-split 0.10 \
  --poll 100 \
  --poll-batch 1 \
  --spec-draft-poll 1 \
  --spec-draft-poll-batch 1 \
  --temp 0 \
  --min-p 0.0 \
  --top-p 0.95 \
  --top-k 20 \
  --repeat-penalty 1.0 \
  --seed 123 \
  --cache-ram 8192 \
  --parallel 1 \
  --metrics

Use --parallel 1 for this MTP profile. Multi-slot serving changes draft-MTP behavior and is not the intended configuration.

The local profile uses Vulkan0 for target and draft on Strix Halo. The required runner is still the ROCmFP4 llama.cpp fork because its Vulkan path understands the custom ROCmFP4 GGUF tensor types.

UI Notes

The reproducible public build above uses -DLLAMA_BUILD_WEBUI=OFF, so it is an API server build. Use it with OpenAI-compatible clients pointed at /v1/chat/completions.

If you need the bundled llama.cpp WebUI, keep the same Vulkan-only backend settings but build from a revision where the WebUI assets build or provision successfully. Do not switch to the older HIP-enabled Strix runner path for multimodal validation.

Build The Required ROCmFP4 Runner

The GGUF is already provided. You only need to build the custom server once:

git clone https://github.com/ciru-ai/ROCmFPX.git
cd ROCmFPX
git checkout 7aa484a2f0a504dc612a3d74a068024f3e6d6353

cmake -S . -B build-vulkan-rocmfp4 -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DGGML_VULKAN=ON \
  -DGGML_HIP=OFF \
  -DGGML_CUDA=OFF \
  -DLLAMA_BUILD_WEBUI=OFF

cmake --build build-vulkan-rocmfp4 -j 16 --target llama-server llama-quantize

The server binary will be here:

build-vulkan-rocmfp4/bin/llama-server

Recommended Strix Halo prerequisites:

sudo apt-get update
sudo apt-get install -y git cmake ninja-build build-essential clang pkg-config \
  glslc vulkan-tools

vulkaninfo --summary

This Vulkan-only build is the recommended path for vision on this release. The older HIP-enabled Strix runner path can load the projector but was observed to give wrong image descriptions, so do not use that path for multimodal validation.

About ROCmFP4 / Chadrock

ROCmFP4 is not stock Q4, MXFP4, NVFP4, or a renamed K-quant. It is a custom GGUF tensor format with AMD-focused backend support.

Charlie's ROCmFP4 fork adds:

• Q4_0_ROCMFP4 and Q4_0_ROCMFP4_FAST tensor formats
• tensor-aware Strix presets such as Q4_0_ROCMFP4_STRIX_LEAN
• ROCm/HIP kernels for the new layouts
• Vulkan shader support for ROCmFP4 decode paths
• MTP regression guards for target/draft serving

The practical point: this model is a Chadrock runtime artifact. Use the ROCmFP4 runner named above.

File

Credits

• DavidAU: Qwen3.6 40B Opus Deckard / Heretic / Uncensored / Thinking source lineage.
• PiehSoft: Deckard 40B MTP GGUF recipe, MTP validation notes, and Q4_K_M baseline used for this comparison.
• froggeric: Qwen3.6 27B mmproj used for upstream multimodal validation and included here as an experimental companion.
• Unsloth: Qwen3.6 27B MTP GGUF donor reference used in the compatible MTP-head lineage.
• Qwen: Qwen3.6 base model family.
• charlie12345 / @Italianclownz: ROCmFP4 llama.cpp fork, Strix Halo build path, and AMD-focused MTP runtime work.

Notes

• This is an experimental AMD ROCmFP4/MTP runtime release.
• Stock llama.cpp, Ollama, LM Studio, and other tools will not understand the ROCmFP4 GGUF tensor types unless they incorporate the required runner support.
• Vision is validated with mmproj-Qwen3.6-27B-f16.mmproj using the Vulkan-only ROCmFP4 build path above.
• The benchmark comparison is intentionally server/API based because standalone llama-bench TG is only diagnostic for MTP models and can understate served draft-MTP throughput.
• The model can be served under any alias. The local service profile uses main; the public release identity and actual file name are the full Chadrock name above.