feat(strixhalo): full source build of vLLM for gfx1151 (v1.0.20)

- Build vLLM v0.15.1 from source against vendor torch 2.9.1
- Preserve AMD's vendor PyTorch from rocm/pytorch:rocm7.0.2 base
- use_existing_torch.py --prefix to strip torch from build-requires
- Compile C++/HIP extensions for gfx1100 (mapped from gfx1151)
- Install triton/flash-attn from wheels.vllm.ai/rocm with --no-deps
- Add torch vendor verification step to catch accidental overwrites
- Fix GPU_RESOURCE default to match cluster (gpu_strixhalo)
- Remove unsupported expandable_segments from PYTORCH_ALLOC_CONF
- AITER is gfx9-only; gfx11 uses TRITON_ATTN backend by default

dockerfiles/Dockerfile.ray-worker-strixhalo

@@ -3,56 +3,70 @@
 # Used for: vLLM (Llama 3.1 70B)
 #
 # Build:
-#   docker build -t git.daviestechlabs.io/daviestechlabs/ray-worker-strixhalo:latest \
+#   docker build -t registry.lab.daviestechlabs.io/daviestechlabs/ray-worker-strixhalo:v1.0.20 \
 #     -f dockerfiles/Dockerfile.ray-worker-strixhalo .
 #
-# Uses ROCm 7.0.2 vendor image (Ubuntu 24.04 / glibc 2.38) so system ROCm
+# STRATEGY: Full source build of vLLM on AMD's vendor PyTorch image.
-# libraries (libhsa-runtime64, libhipblas, libMIOpen, etc.) are ABI-compatible
+#
-# with torch + vLLM wheels from wheels.vllm.ai/rocm/ (compiled for ROCm 7.0).
+# The vendor image (rocm/pytorch ROCm 7.0.2 / Ubuntu 24.04 / Python 3.12)
-# ROCm 7.1 base causes segfault in libhsa-runtime64.so due to ABI mismatch.
+# ships torch 2.9.1 compiled by AMD CI against the exact ROCm libraries in
-# Note: Python 3.12 required — vLLM ROCm wheel (wheels.vllm.ai/rocm) is cp312 only
+# the image.  Pre-built vLLM torch wheels (wheels.vllm.ai) carry a custom
+# torch 2.9.1+git8907517 that segfaults in libhsa-runtime64.so on gfx1151
+# during HSA queue creation.  By keeping the vendor torch and compiling vLLM
+# from source we guarantee ABI compatibility across the entire stack.
+#
+# gfx1151 is mapped to gfx1100 at runtime via HSA_OVERRIDE_GFX_VERSION=11.0.0,
+# so all HIP kernels are compiled for the gfx1100 target.
+#
+# Note: AITER is gfx9-only.  On gfx11, vLLM defaults to TRITON_ATTN backend.
 
 FROM docker.io/rocm/pytorch:rocm7.0.2_ubuntu24.04_py3.12_pytorch_release_2.9.1
 
-# OCI Image Spec labels
+# ── Build arguments ─────────────────────────────────────────────────────
+ARG VLLM_VERSION=v0.15.1
+ARG PYTORCH_ROCM_ARCH="gfx1100"
+ARG MAX_JOBS=16
+
+# ── OCI labels ──────────────────────────────────────────────────────────
 LABEL org.opencontainers.image.title="Ray Worker - AMD Strix Halo"
-LABEL org.opencontainers.image.description="Ray Serve worker for AMD Strix Halo (vLLM LLM inference)"
+LABEL org.opencontainers.image.description="Ray Serve worker for AMD Strix Halo (vLLM source-built)"
 LABEL org.opencontainers.image.vendor="DaviesTechLabs"
 LABEL org.opencontainers.image.source="https://git.daviestechlabs.io/daviestechlabs/kuberay-images"
 LABEL org.opencontainers.image.licenses="MIT"
-LABEL gpu.target="amd-rocm-7.0-gfx1151"
+LABEL gpu.target="amd-rocm-7.0.2-gfx1151"
 LABEL ray.version="2.53.0"
+LABEL vllm.build="source"
 
 WORKDIR /app
 
+# ── Persistent environment ──────────────────────────────────────────────
 # The vendor image ships a venv at /opt/venv with Python 3.12 + torch 2.9.1.
-# We keep using that venv for all pip installs.
+# All pip installs go into this venv.
 ENV ROCM_HOME=/opt/rocm \
     VIRTUAL_ENV=/opt/venv
 ENV PATH="/opt/venv/bin:/opt/rocm/bin:/opt/rocm/llvm/bin:/home/ray/.local/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin" \
     LD_LIBRARY_PATH="/opt/rocm/lib:/opt/rocm/lib64:/opt/venv/lib" \
     HSA_PATH="/opt/rocm/hsa" \
     HIP_PATH="/opt/rocm/hip" \
-    # Strix Halo (gfx1151) specific settings
+    # Strix Halo (gfx1151 / RDNA 3.5) runtime settings
     HIP_VISIBLE_DEVICES=0 \
     HSA_ENABLE_SDMA=0 \
-    PYTORCH_HIP_ALLOC_CONF="expandable_segments:True,max_split_size_mb:512" \
+    PYTORCH_ALLOC_CONF="max_split_size_mb:512" \
     HSA_OVERRIDE_GFX_VERSION="11.0.0" \
     ROCM_TARGET_LST="gfx1151,gfx1100"
 
-# System dependencies + create ray user (uid 1000 / gid 100) for KubeRay
+# ── System dependencies + build tools ───────────────────────────────────
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt,sharing=locked \
     apt-get update && apt-get install -y --no-install-recommends \
-        libelf1 \
+        # Runtime
-        libnuma1 \
+        libelf1 libnuma1 libdrm2 libdrm-amdgpu1 kmod libopenmpi3 \
-        libdrm2 \
+        # Build (for vLLM C++/HIP compilation)
-        libdrm-amdgpu1 \
+        git cmake ninja-build ccache \
-        kmod \
-        libopenmpi3 \
     && rm -rf /var/lib/apt/lists/* \
+    # Create ray user (uid 1000 / gid 100) for KubeRay.
+    # Vendor image may already have UID 1000 — rename or create.
     && (groupadd -g 100 -o users 2>/dev/null || true) \
-    # Vendor image may already have UID 1000 — rename it to ray, or create ray
     && existing=$(getent passwd 1000 | cut -d: -f1) \
     && if [ -n "$existing" ] && [ "$existing" != "ray" ]; then \
            usermod -l ray -d /home/ray -m -s /bin/bash "$existing"; \
@@ -64,51 +78,117 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
 # Install uv for fast Python package management (ADR-0014)
 COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
 
-# Remove vendor torch — replace with the exact torch from wheels.vllm.ai/rocm/
+# ── Python build dependencies ──────────────────────────────────────────
-# (2.9.1+git8907517) that vLLM 0.15.1+rocm700 was compiled against.
-# The vendor torch is close but may differ in C++ ABI details.
-RUN uv pip uninstall --python /opt/venv/bin/python3 \
-        torch torchaudio torchvision 2>/dev/null || true
-
-# Install Ray, vLLM ROCm, torch ROCm 7.0, and inference dependencies.
-# --index-strategy unsafe-best-match: let uv pull each package from whichever
-# index has the best-matching version (vllm from rocm/, torch from rocm7.0,
-# setuptools/others from PyPI).
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --python /opt/venv/bin/python3 \
-        --index-strategy unsafe-best-match \
+        'cmake>=3.26.1' \
+        ninja \
+        'packaging>=24.2' \
+        'setuptools>=77.0.3,<80.0.0' \
+        'setuptools-scm>=8' \
+        wheel \
+        'jinja2>=3.1.6' \
+        regex
+
+# ── Build vLLM from source ─────────────────────────────────────────────
+# Clone at the specified version, then strip torch from build-requires so
+# the build system uses the vendor torch already in /opt/venv.
+WORKDIR /tmp/vllm-build
+RUN git clone --depth 1 --branch ${VLLM_VERSION} \
+        https://github.com/vllm-project/vllm.git .
+
+# Remove torch from build-requires.  use_existing_torch.py is provided by
+# newer vLLM branches; fall back to sed for older releases.
+# --prefix strips only torch=/torchvision=/torchaudio= lines (not
+# unrelated packages whose name happens to contain "torch").
+RUN if [ -f use_existing_torch.py ]; then \
+        python3 use_existing_torch.py --prefix; \
+    else \
+        sed -i '/"torch[= ]/d; /"torchvision[= ]/d; /"torchaudio[= ]/d' pyproject.toml 2>/dev/null || true; \
+    fi
+
+# Compile C++/HIP extensions and install the vLLM Python package.
+#   --no-build-isolation : use vendor torch + our build deps directly
+#   --no-deps            : we install runtime deps ourselves (below)
+ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} \
+    VLLM_TARGET_DEVICE=rocm \
+    MAX_JOBS=${MAX_JOBS} \
+    CMAKE_BUILD_TYPE=Release \
+    CCACHE_DIR=/root/.cache/ccache
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=cache,target=/root/.cache/ccache \
+    uv pip install --python /opt/venv/bin/python3 \
+        --no-build-isolation --no-deps .
+
+# ── ROCm-specific Python wheels ────────────────────────────────────────
+# triton (ROCm HIP backend) and flash-attn (Triton AMD kernels for gfx11)
+# from vLLM's ROCm wheels index.  --no-deps prevents torch replacement.
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --python /opt/venv/bin/python3 \
+        --no-deps \
         --prerelease=allow \
         --extra-index-url https://wheels.vllm.ai/rocm/ \
-        --extra-index-url https://download.pytorch.org/whl/rocm7.0 \
+        triton triton-kernels flash_attn
+
+# ── Runtime Python dependencies ─────────────────────────────────────────
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --python /opt/venv/bin/python3 \
         'ray[default]==2.53.0' \
-        'vllm==0.15.1+rocm700' \
-        torch torchaudio torchvision \
         'transformers>=4.35.0,<5.0' \
         'accelerate>=0.25.0,<1.0' \
         'sentence-transformers>=2.3.0,<3.0' \
         'httpx>=0.27.0,<1.0' \
         'scipy>=1.11.0,<2.0' \
         'pandas>=2.0.0,<3.0' \
-        'numpy>=2.1.0,<2.3'
+        'numpy>=2.1.0,<2.3' \
+        'numba>=0.60.0,<0.62' \
+        'tokenizers>=0.20.0' \
+        'safetensors>=0.4.0' \
+        'uvloop>=0.21.0' \
+        'pyyaml>=6.0' \
+        'requests>=2.31.0' \
+        'aiohttp>=3.9.0' \
+        'pillow>=10.0' \
+        'prometheus-client>=0.20.0' \
+        'py-cpuinfo>=9.0.0' \
+        'filelock>=3.13.0' \
+        'psutil>=5.9.0' \
+        'msgpack>=1.0.0' \
+        'gguf>=0.6.0' \
+        'compressed-tensors>=0.8.0' \
+        'outlines>=0.1.0' \
+        'lm-format-enforcer>=0.10.0' \
+        'partial-json-parser>=0.2.0' \
+        'mistral-common>=1.5.0'
 
-# Install amdsmi sysfs shim (required for vLLM ROCm platform detection).
+# ── Verify vendor torch survived ───────────────────────────────────────
-# Even though the native amdsmi works on Ubuntu 24.04, the sysfs shim is
+# Fail early if any install step accidentally replaced the vendor torch.
-# still needed because the native library reports GTT instead of VRAM on
+RUN python3 -c "\
-# unified-memory APUs. Must be installed after vLLM/torch so PyPI amdsmi
+import torch; \
-# doesn't overwrite it.
+v = torch.__version__; \
+assert '+git' not in v, f'vLLM torch detected ({v}) — vendor torch was overwritten!'; \
+print(f'torch {v} (vendor) OK')"
+
+# ── amdsmi sysfs shim ──────────────────────────────────────────────────
+# Required for vLLM ROCm platform detection.  The native amdsmi reports
+# GTT instead of VRAM on unified-memory APUs.
 COPY amdsmi-shim /tmp/amdsmi-shim
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --python /opt/venv/bin/python3 /tmp/amdsmi-shim \
     && rm -rf /tmp/amdsmi-shim
 
-# Patch torch.cuda.mem_get_info for unified memory APUs.
+# ── VRAM fix for unified memory APU ────────────────────────────────────
-# On Strix Halo, PyTorch reports GTT (128 GiB) instead of real VRAM (96 GiB).
+# Monkey-patches torch.cuda.mem_get_info to report actual VRAM (96 GiB)
-# The .pth file auto-patches mem_get_info on Python startup.
+# rather than GTT (128 GiB).
 COPY amdsmi-shim/strixhalo_vram_fix.py \
     /opt/venv/lib/python3.12/site-packages/strixhalo_vram_fix.py
 RUN echo "import strixhalo_vram_fix" > \
     /opt/venv/lib/python3.12/site-packages/strixhalo_vram_fix.pth
 
+# ── Clean up build artifacts ────────────────────────────────────────────
+WORKDIR /app
+RUN rm -rf /tmp/vllm-build
+
 # Copy entrypoint script (ray-serve-apps is installed from PyPI at runtime)
 COPY --chmod=755 dockerfiles/ray-entrypoint.sh /app/ray-entrypoint.sh
 
@@ -123,7 +203,7 @@ ENV PYTHONPATH=/app \
     PYTHONUNBUFFERED=1 \
     PYTHONDONTWRITEBYTECODE=1 \
     RAY_HEAD_SVC="ai-inference-raycluster-head-svc" \
-    GPU_RESOURCE="gpu_amd_strixhalo" \
+    GPU_RESOURCE="gpu_strixhalo" \
     NUM_GPUS="1"
 
 # Health check