tekBlueprint

# Prebuilt (recommended) -- self-contained, bundles the Vulkan driver
curl -LO https://github.com/Nathanw1014/strix-halo-llamacpp/releases/download/v0.7.5/strix-halo-llamacpp-vulkan-portable.tar.gz
tar xzf strix-halo-llamacpp-vulkan-portable.tar.gz
# -> vulkan/bin/{llama-server,llama-cli,llama-bench}, vulkan/driver/, vulkan/_run

HERE=/path/to/strix-halo-llamacpp/vulkan
export VK_ICD_FILENAMES="$HERE/driver/radeon_icd.x86_64.json"
export VK_DRIVER_FILES="$VK_ICD_FILENAMES"
export LD_LIBRARY_PATH="$HERE/driver:$HERE/bin"
export GGML_VK_MMID_ROWLISTS=1 GGML_VK_MMID_SMALLN=1 GGML_VK_MMID_BM64=1
export GGML_VK_MMID_WAVE32=1 GGML_VK_MMID_F16B=1 GGML_VK_MMID_M128=1
export GGML_VK_FA_WAVE32=0

#!/usr/bin/env bash
HERE=/path/to/strix-halo-llamacpp/vulkan
MODEL=/path/to/Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf   # all 4 shards auto-loaded
TEMPLATE=/path/to/chat_template.jinja

# (driver env from "Getting the build" above, or just call $HERE/_run)

$HERE/bin/llama-server \
    --host 0.0.0.0 \
    --port 1234 \
    -c 0 \
    --n-gpu-layers 999 \
    --batch-size 2048 \
    --ubatch-size 1024 \
    --flash-attn on \
    --temp 1.0 \
    --top-p 0.95 \
    --top-k 20 \
    --min-p 0.0 \
    --prio 2 \
    --cpu-strict 1 \
    --threads 16 \
    --threads-batch 32 \
    --no-warmup \
    --jinja \
    --chat-template-file ${TEMPLATE} \
    -ctk q8_0 \
    -ctv q8_0 \
    --parallel 1 \
    --metrics \
    --alias "qwen3.8-flash-next-q4_k_xl" \
    -m ${MODEL}

MM=/path/to/mmproj-BF16.gguf
# add to the launch command:
    --mmproj ${MM} \

MTP=/path/to/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf

    --spec-type draft-mtp \
    --spec-draft-model ${MTP} \
    --spec-draft-n-max 3 \

ARG BASE_IMAGE=rocm/dev-ubuntu-24.04:10.0.0-full
FROM ${BASE_IMAGE}
# build deps — note the non-obvious ones: xxd (ROCk HSACO->ASCII),
# libgl-dev (clr OpenGL), libelf-dev (ROCk), CppHeaderParser (clr hipamd)
RUN apt-get update && apt-get install -y --no-install-recommends \
        build-essential cmake ninja-build git pkg-config ccache python3 python3-pip \
        xxd perl libdrm-dev libnuma-dev libelf-dev libdwarf-dev \
        libgl-dev libegl-dev libglu1-mesa-dev libopenmpi-dev libssl-dev \
      && rm -rf /var/lib/apt/lists/*
RUN pip3 install --break-system-packages --no-cache-dir CppHeaderParser
WORKDIR /work
COPY fetch-sources.sh build.sh ./
RUN ./fetch-sources.sh   # clone the two repos at pinned commits (see .env)
RUN ./build.sh           # ~1h: ROCr -> HIP(clr) -> llama.cpp (GGML_HIP, gfx1151)
# custom ROCk/HIP (retained-PM4) FIRST so they shadow stock /opt/rocm at load
ENV LD_LIBRARY_PATH=/work/runtime/hip/lib:/work/runtime/rocr/lib:/opt/rocm/lib:/opt/rocm/lib64:/opt/rocm/lib/llvm/lib:/work/build/llama.cpp/bin \
    HSA_OVERRIDE_GFX_VERSION=11.5.1 ENABLE_RETAINED_PM4=1 HSA_ENABLE_SDMA=0
ENV PATH=/work/build/llama.cpp/bin:${PATH}

# 1/3 ROCk (retained-PM4 HSA runtime)
cmake -S src/rocm-systems/projects/rocr-runtime -B build/rocr -G Ninja \
  -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -DCMAKE_INSTALL_PREFIX=$RUNTIME/rocr
cmake --build build/rocr -j$(nproc) && cmake --install build/rocr

# 2/3 HIP/clr (retained-PM4). HIP_COMMON_DIR must be the HIP *source* tree
# (has include/hip + cmake + *.cmake.in); stock /opt/rocm lacks the .in templates.
HIP_COMMON_DIR=$SRC/rocm-systems/projects/hip
cmake -S src/rocm-systems/projects/clr -B build/hip -G Ninja \
  -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -DCLR_BUILD_HIP=ON -DCLR_BUILD_OCL=OFF \
  -DROCM_PATH=$RUNTIME/rocr -DCMAKE_PREFIX_PATH=$RUNTIME/rocr;/opt/rocm \
  -DHIP_COMMON_DIR=$HIP_COMMON_DIR -DHIP_PLATFORM=amd \
  -DLLVM_BIN=/opt/rocm/lib/llvm/bin -DClang_BIN=/opt/rocm/lib/llvm/bin \
  -DCMAKE_INSTALL_PREFIX=$RUNTIME/hip
cmake --build build/hip -j$(nproc) && cmake --install build/hip

# 3/3 llama.cpp (GGML_HIP, gfx1151) — against stock /opt/rocm
cmake -S src/llama.cpp -B build/llama.cpp -G Ninja \
  -DCMAKE_BUILD_TYPE=Release -DGGML_HIP=ON -DGGML_HIP_NO_VMM=ON -DGGML_HIP_GRAPHS=ON \
  -DGGML_HIP_MMQ_MFMA=ON -DGGML_CUDA_FA=ON -DGGML_NATIVE=ON -DGGML_OPENMP=ON \
  -DGPU_TARGETS=gfx1151 -DCMAKE_HIP_COMPILER=/opt/rocm/lib/llvm/bin/clang++
cmake --build build/llama.cpp -j$(nproc)

x-rocm-gpu: &rocm-gpu
  devices: ["/dev/kfd", "/dev/dri"]          # ROCm = raw KFD + DRM nodes
  security_opt: ["seccomp:unconfined"]
  environment:
    - HSA_OVERRIDE_GFX_VERSION=11.5.1
    - ENABLE_RETAINED_PM4=1
    - HSA_ENABLE_SDMA=0
  volumes: ["/mnt/local/models:/mnt/local/models:ro"]

services:
  server:
    build: .
    image: strix-pwilkin:gfx1151
    <<: *rocm-gpu
    ports: ["${SERVER_PORT:-1234}:1234"]
    command: >
      llama-server
      -m $MODEL -md $MTP_DRAFT --mmproj $MM_PROJ
      --spec-type draft-mtp --spec-draft-n-max 3
      -ngl 999 -fa on -ctk f16 -ctv f16
      -t 16 -tb 32 -b 2048 -ub 1024
      -c ${CTX:-262144}
      --cache-ram 24576 --cache-reuse 1024
      --host 0.0.0.0 --port 1234 --jinja
    restart: unless-stopped
  bench:
    image: strix-pwilkin:gfx1151
    <<: *rocm-gpu
    command: >
      llama-bench -m $MODEL -ngl 999 -fa on -p 512 -n 8 -d 32768
      -b 2048 -ub 1024 -ctk f16 -ctv f16 -t 16 -r 3
    profiles: ["bench"]

LLAMA_GIT_SHA=b0f31f5876ef3856b55f5bb88072cc96e5effafe        # pwilkin/llama.cpp@strix-halo
ROCM_SYSTEMS_GIT_SHA=7dda3ac6cfe6bbe0b7f08c23a67cfa118d8641a1 # pwilkin/rocm-systems@ilintar-experiments
GPU_TARGETS=gfx1151
CTX=262144
SERVER_PORT=1234
MM_PROJ=/mnt/local/models/.../Qwen3.8-Flash-Next-GGUF/mmproj-BF16.gguf   # vision projector

cd strix-pwilkin/
docker compose build          # ~1h; bakes /work/build/llama.cpp/bin + /work/runtime/{rocr,hip}
docker compose up -d server   # serve: ctx 262144, MTP on, f16 KV, prompt cache, port 1234
docker compose run --rm bench # plain-decode A/B (llama-bench can't see draft flags)

--temp 0.3 --top-p 0.5 --top-k 5

--temp 1.0 --top-p 0.95 --top-k 40