v1.0.26: dynamic VRAM via GTT for 32GB carve-out

strixhalo_vram_fix.py: compute effective VRAM as
min(GTT_pool, physical_RAM) - 4GB OS reserve instead of
raw sysfs VRAM. Prevents OOM when carve-out < model size
and prevents kernel OOM when GTT > physical RAM.

dockerfiles/Dockerfile.ray-worker-strixhalo

@@ -2,28 +2,29 @@
 # Used for: vLLM (Llama 3.1 70B)
 #
 # Build:
-#   docker build -t registry.lab.daviestechlabs.io/daviestechlabs/ray-worker-strixhalo:v1.0.21 \
+#   docker build -t registry.lab.daviestechlabs.io/daviestechlabs/ray-worker-strixhalo:v2.0.0 \
 #     -f dockerfiles/Dockerfile.ray-worker-strixhalo .
 #
 # STRATEGY: Full source build of vLLM on AMD's vendor PyTorch image.
 #
-# The vendor image (rocm/pytorch ROCm 7.0.2 / Ubuntu 24.04 / Python 3.12)
-# ships torch 2.9.1 compiled by AMD CI against the exact ROCm libraries in
-# 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.
+# The vendor image (rocm/pytorch ROCm 7.2 / Ubuntu 24.04 / Python 3.12)
+# ships torch 2.9.1+rocm7.2.0 compiled by AMD CI against the exact ROCm
+# libraries in the image.  ROCm 7.2 includes the hsakmt VGPR count fix
+# for gfx1151 (TheRock #2991) — ROCm 7.0.x/7.1.x segfault during HSA
+# queue creation due to incorrect VGPR sizing.  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.
+# ROCm 7.2 supports gfx1151 natively — no HSA_OVERRIDE_GFX_VERSION needed.
+# HIP kernels are compiled directly for the gfx1151 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
+FROM docker.io/rocm/pytorch:rocm7.2_ubuntu24.04_py3.12_pytorch_release_2.9.1
 
 # ── Build arguments ─────────────────────────────────────────────────────
 ARG VLLM_VERSION=v0.15.1
-ARG PYTORCH_ROCM_ARCH="gfx1100"
+ARG PYTORCH_ROCM_ARCH="gfx1151"
 ARG MAX_JOBS=16
 
 # ── OCI labels ──────────────────────────────────────────────────────────
@@ -32,7 +33,7 @@ LABEL org.opencontainers.image.description="Ray Serve worker for AMD Strix Halo
 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-7.0.2-gfx1151"
+LABEL gpu.target="amd-rocm-7.2-gfx1151"
 LABEL ray.version="2.53.0"
 LABEL vllm.build="source"
 
@@ -52,17 +53,16 @@ ENV PATH="/opt/venv/bin:/opt/rocm/bin:/opt/rocm/llvm/bin:/home/ray/.local/bin:/u
     HIP_VISIBLE_DEVICES=0 \
     HSA_ENABLE_SDMA=0 \
     PYTORCH_ALLOC_CONF="max_split_size_mb:512" \
-    HSA_OVERRIDE_GFX_VERSION="11.0.0" \
-    ROCM_TARGET_LST="gfx1151,gfx1100"
+    ROCM_TARGET_LST="gfx1151"
 
 # ── System setup ─────────────────────────────────────────────────────────
-# The vendor image already ships ALL needed packages:
-#   cmake 4.0, hipcc 7.0.2, clang++ 20.0 (AMD ROCm LLVM), git,
-#   libelf, libnuma, libdrm, libopenmpi3, and HIP dev headers/cmake configs.
+# The vendor image ships hipcc 7.2, clang++ (AMD ROCm LLVM), git,
+# libelf, libnuma, libdrm, libopenmpi3, and HIP dev headers/cmake configs.
+# cmake is NOT in the 7.2 image — installed via pip below.
 #
 # CRITICAL: Do NOT run apt-get upgrade or install ANY packages from apt.
 # Even installing ccache triggers a dependency cascade that pulls in
-# Ubuntu's hipcc 5.7.1 (which overwrites the vendor hipcc 7.0.2) and
+# Ubuntu's hipcc 5.7.1 (which overwrites the vendor hipcc 7.2) and
 # a broken /usr/bin/hipconfig.pl that makes cmake find_package(hip)
 # report version 0.0.0 → "Can't find CUDA or HIP installation."
 #
@@ -81,9 +81,9 @@ RUN (groupadd -g 100 -o users 2>/dev/null || true) \
 COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
 
 # ── Python build dependencies ──────────────────────────────────────────
-# CRITICAL: vLLM requires cmake<4.  The vendor image ships cmake 4.0.0
-# which changed find_package(MODULE) behaviour and breaks FindHIP.cmake
-# (reports HIP version 0.0.0).  Downgrade to 3.x per vLLM's rocm-build.txt.
+# CRITICAL: vLLM requires cmake<4.  cmake 4.0+ changed find_package(MODULE)
+# behaviour and breaks FindHIP.cmake (reports HIP version 0.0.0).
+# The ROCm 7.2 image does not ship cmake, so we install 3.x here.
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --python /opt/venv/bin/python3 \
         'cmake>=3.26.1,<4' \
@@ -234,10 +234,11 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 # ── Verify vendor torch survived ───────────────────────────────────────
 # Fail early if any install step accidentally replaced the vendor torch.
+# ROCm 7.2 vendor torch version: 2.9.1+rocm7.2.0.git7e1940d4
 RUN python3 -c "\
 import torch; \
 v = torch.__version__; \
-assert '+git' not in v, f'vLLM torch detected ({v}) — vendor torch was overwritten!'; \
+assert 'rocm7.2' in v, f'Expected ROCm 7.2 vendor torch, got {v}'; \
 print(f'torch {v} (vendor) OK')"
 
 # ── amdsmi sysfs shim ──────────────────────────────────────────────────