Qwen3.8-27B CIRU ActiveFPX + PromptForge v3

V3 is Ciru's consolidated quality-and-speed build of Qwen3.8-27B for AMD ROCm systems. It combines the best compatible work from the non-IU4 development lane:

• an FFN-equalized compact language-model body;
• 50 selectively retained late-writer Q6 tensors;
• the full Q8 output projection from v2;
• full-W8 PromptForge FFN and Gated DeltaNet prompt views;
• the small-row W8 route;
• draft-only Output-K8 shortlist generation followed by authoritative Q8 reranking;
• native depth-4 MTP, with no ngram subsystem; and
• the original BF16 Qwen3.8 vision tower and 5120-wide merger as a separate projector.

The recommended model is Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-v3-Q8-FFNEQ-LateQ6.gguf. The GGUF can run by itself. The .pfs companions require the included CIRU runtime patch and trade additional memory for faster prompt and draft-output paths.

Runtime requirement: the accelerated configuration requires the CIRU ROCmFPX v2.3 base plus runtime/qwen38-v3-output-k8-runtime.patch. The performance work was developed on gfx1151 with TheRock ROCm 7.15. A stock llama.cpp binary can load the GGUF alone, but it cannot use the PromptForge companions or the Output-K8 proxy.

What improved over v2

The new tensor allocation strictly improved all three retained BF16-relative numerical quality measures:

On the matched serving screen, v3 improved 128-token chat generation throughput by 4.79% and end-to-end chat wall throughput by 6.17%. Prompt throughput was effectively even at 2,048 tokens (-0.28%) and improved at 3,524 tokens (+1.20%). The short 512-token prompt route regressed by 34.24%; that is a named optimization target rather than a hidden result.

A locked, no-ngram, deterministic 20-task HumanEval/EvalPlus panel improved in aggregate from 14/20 to 15/20 base and from 6/20 to 7/20 Plus. V3 gained on tasks 25, 83, and 132, but introduced one genuine semantic regression on task 130 for odd-length input. Accordingly, the frozen "no new base failure" sub-gate did not pass even though aggregate base and Plus scores improved.

These are local matched measurements on one Ryzen AI Max+ 395 / Radeon 8060S system. They are evidence for this release decision, not hardware-independent leaderboard claims.

262K context and vision status

The canonical serving profile defaults to 262,144 tokens. Context length is not reduced to disguise short-prompt or allocation overhead. A 245,760-token prompt plus generation passed without truncation or OOM on the preceding v2 artifact and the same model architecture. That long screen was not repeated on the v3 file at publication time so the GPU could be returned to active work.

The repository includes the matching BF16 vision projector converted unchanged from official Qwen3.8-27B revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. Its 5120-wide merger matches the language model. Vision is enabled in the release profile on that structural and provenance basis; an additional v3-specific image-plus-MTP runtime screen was not run during publication.

Release files

The v3 language-serving payload with all three .pfs companions is 38,445,917,344 bytes (35.81 GiB). The GGUF alone is 15.44 GiB. The vision projector adds 0.87 GiB when multimodal serving is enabled. Previous v1 and v2 artifacts remain in the repository for reproducibility, but v3 is the default.

The Output-K8 proxy was extracted from the retained v1 Q4 output tensor. It is safe with v3 because v3's authoritative Q8 output payload is byte-identical to the qualified public-v2 payload. The proxy proposes only a shortlist; it never replaces final Q8 token selection.

Build the required runtime

The accelerated path is a custom ROCm build. The commands below provide a copyable Ubuntu/Debian-family source-build path. The release itself was validated on NixOS with a pinned TheRock ROCm 7.15 toolchain; the Ubuntu package combination is a portable build path, not a separately benchmarked environment.

Install ordinary build dependencies and a ROCm 7.15 installation that supports your AMD target:

sudo apt update
sudo apt install -y build-essential cmake git ninja-build python3 libcurl4-openssl-dev

export ROCM_PATH=/opt/rocm
export PATH="$ROCM_PATH/bin:$ROCM_PATH/llvm/bin:$PATH"
export LD_LIBRARY_PATH="$ROCM_PATH/lib:$ROCM_PATH/lib/llvm/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
"$ROCM_PATH/bin/rocminfo" | grep -m1 gfx1151

Clone the public v2.3 base, apply this repository's v3 patch, and pin Composable Kernel:

git clone https://github.com/ciru-ai/ROCmFPX.git
cd ROCmFPX
git checkout qwen3.8-activefpx-promptforge-v2.3
git am /models/runtime/qwen38-v3-output-k8-runtime.patch

git clone https://github.com/ROCm/composable_kernel.git ../composable_kernel
git -C ../composable_kernel checkout fdf4bb7fcc984811cef48ce817d89aac064b984a

Configure and build for the validated gfx1151 target:

cmake -S . -B build-promptforge -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_C_COMPILER="$ROCM_PATH/llvm/bin/clang" \
  -DCMAKE_CXX_COMPILER="$ROCM_PATH/llvm/bin/clang++" \
  -DCMAKE_HIP_COMPILER="$ROCM_PATH/llvm/bin/clang++" \
  -DCMAKE_PREFIX_PATH="$ROCM_PATH" \
  -DGGML_HIP=ON -DGGML_CUDA=OFF -DGGML_VULKAN=OFF \
  -DGGML_HIP_FORCE_MMQ=ON -DGGML_HIP_GRAPHS=ON \
  -DGGML_HIP_MMQ_MFMA=ON -DGGML_HIP_NO_VMM=ON \
  -DGGML_HIP_ROCWMMA_FATTN=OFF -DGGML_NATIVE=ON \
  -DAMDGPU_TARGETS=gfx1151 -DGPU_BUILD_TARGETS=gfx1151 \
  -DPROMPTFORGE_CK_ROOT="$PWD/../composable_kernel" \
  -DGGML_BUILD_TESTS=OFF -DLLAMA_BUILD_TESTS=OFF \
  -DLLAMA_BUILD_SERVER=ON

cmake --build build-promptforge --target llama-server -j"$(nproc)"

Change the GPU target only if your ROCm toolchain and the required kernels support it. The advertised measurements are specific to gfx1151.

Serving: full v3 configuration

Download the v3 GGUF, three .pfs companions, and the projector into /models. Then launch with the exact non-IU4/no-ngram profile:

export ROCM_PATH=/opt/rocm
export LD_LIBRARY_PATH="$PWD/build-promptforge/bin:$ROCM_PATH/lib:$ROCM_PATH/lib/llvm/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
export HSA_OVERRIDE_GFX_VERSION=11.5.1
export HIP_VISIBLE_DEVICES=0
export GGML_CUDA_GRAPH_OPT=0

export PROMPTFORGE_SIDECAR=/models/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-FFN.pfs
export PROMPTFORGE_GDN_SIDECAR=/models/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-GDN.pfs
export PROMPTFORGE_MODE=m2048_fused_tail1476
export PROMPTFORGE_ENABLE_SMALLM_W8=1
export PROMPTFORGE_MTP_OUTPUT_K8=1
export PROMPTFORGE_MTP_OUTPUT_K8_VALIDATE=0
export PROMPTFORGE_MTP_OUTPUT_K8_PROXY=/models/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-Output-K8.pfs

unset LLAMA_MTP_CPU_ARGMAX_FASTPATH LLAMA_MTP_BACKEND_GREEDY
unset PROMPTFORGE_IU4_SIDECAR PROMPTFORGE_GDN_IU4_SIDECAR
unset PROMPTFORGE_ENABLE_NGRAM_MOD PROMPTFORGE_ENABLE_NGRAM_M65_IU4

./build-promptforge/bin/llama-server \
  -m /models/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-v3-Q8-FFNEQ-LateQ6.gguf \
  --mmproj /models/Qwen3.8-27B-mmproj-BF16.mmproj \
  --alias main --host 127.0.0.1 --port 8080 --jinja \
  -fit off -dev ROCm0 -ngl 999 \
  -c 262144 -b 2048 -ub 2048 -fa on \
  -ctk f16 -ctv f16 -t 16 -tb 16 -np 1 \
  -ctxcp 0 --cache-ram 0 --no-cache-prompt --no-cache-idle-slots \
  --timeout 3600 --metrics \
  --spec-type draft-mtp --spec-draft-device ROCm0 --spec-draft-ngl 999 \
  --spec-draft-type-k f16 --spec-draft-type-v f16 \
  --spec-draft-n-max 4 --spec-draft-n-min 0 \
  --spec-draft-p-min 0.0 --spec-draft-p-split 0.10 \
  --spec-draft-backend-sampling \
  --temp 0.7 --top-p 0.95 \
  --reasoning off --reasoning-format none --reasoning-budget -1

At 262K, request latency and memory use are workload-dependent. The long context ceiling is intentional; optimize short-prompt allocation and routing without silently lowering the configured maximum.

GGUF-only fallback

The v3 GGUF is independently usable without the three .pfs files. Remove all PROMPTFORGE_* environment variables and use a compatible GGUF runtime. This retains the v3 tensor allocation and Q8 output head but not DualView prompt acceleration or Output-K8 draft shortlisting.

Intended use and limitations

This release targets local inference, coding, agents, general assistants, long-context work, and multimodal experimentation on AMD systems. It is a hardware-sensitive custom runtime release. Users remain responsible for application-specific evaluation, safety controls, and compliance.

Known release-specific follow-ups are the 512-row prompt route, the observed HumanEval/130 odd-length failure, and a dedicated v3 vision-plus-MTP runtime screen. These are disclosed so downstream users can choose appropriate tests.

Credits

• Qwen/Qwen3.8-27B
• CIRU ROCmFPX
• ROCmFPX
• llama.cpp
• AMD Composable Kernel
• CIRU Inference Lab

License

The derived model artifact follows the Qwen3.8-27B Apache-2.0 license. Runtime source and third-party components retain their respective licenses.