docker-images: glm53-jovian-r12 — in-docker build of lab fork (vllm base 54f6e9826c + PR heads 550-557 merges; b12x master + PR heads 260-270 merges; day-0 arm64-cu130 base; nvrtc/headers cmake fixes); teb c1 on vavrusa lane: 88/100 (default 69-scenario set)

docker-images/glm53-jovian-r12/Dockerfile

@@ -0,0 +1,89 @@
+# Two-stage, in-docker build of the local-inference-lab r12 stack for GB10
+# (arm64 / SM121a), per the rtx6kpro source contract (models/glm-5.3-flash.md):
+#
+#   vLLM  = local-inference-lab/vllm composition commit
+#           7d8a09c42c7ba743b9e936562aa9205f9d0fda9d
+#           (dev/jovian-judgement @54f6e9826c20... + PR heads 550/552/553/554/555/557)
+#   B12X  = local-inference-lab/b12x composition commit
+#           d064ca4cc3aa25292f1a8756fa51b38134d1df84
+#           (master @139e04048bc3b... + PR heads 260/267/268/269/270)
+#   LMCache optional — NOT vendored (opt-in).
+#
+# The compile happens INSIDE docker (operator directive). Torch/flashinfer/
+# CUDA come from the working day-0 arm64 cu130 base that already serves GLM
+# 5.3 Flash; the fork sources install with --no-deps so the base toolchain
+# stays pinned. sm_121f applies to block-scaled FP4 kernels (kk3 gencode
+# rule); TORCH_CUDA_ARCH_LIST pins the extension build.
+#
+# Verification markers: pip lines for vllm==0.1.dev...; python -c
+# "import vllm" in a `--gpus` container; quantprobe in modelopt.py.
+
+ARG BASE_IMAGE
+FROM ${BASE_IMAGE} AS build
+
+ARG VLLM_REF
+ARG B12X_REF
+ARG VLLM_PRS="550 552 553 554 555 557"
+ARG B12X_PRS="260 267 268 269 270"
+ARG VLLM_REPO=https://github.com/local-inference-lab/vllm.git
+ARG B12X_REPO=https://github.com/local-inference-lab/b12x.git
+
+RUN set -eux; apt-get update && apt-get install -y --no-install-recommends git ca-certificates && rm -rf /var/lib/apt/lists/*
+
+# The rtx6kpro "composition commit" (7d8a09c… / d064ca4c…) is NOT a public
+# ref — reproduce it: base branch checkout + sequential merge of the open
+# PR heads (refs/pull/*/head), ascending order.
+RUN set -eux; \
+    git config --global user.email build@sm121.lan; \
+    git config --global user.name sm121-builder; \
+    git clone ${VLLM_REPO} /tmp/vllm-src; \
+    git -C /tmp/vllm-src checkout ${VLLM_REF}; \
+    for pr in ${VLLM_PRS}; do \
+        git -C /tmp/vllm-src fetch origin "refs/pull/$pr/head:prs/$pr"; \
+        git -C /tmp/vllm-src merge --no-ff -m "PR $pr" "prs/$pr"; \
+    done; \
+    git clone ${B12X_REPO} /tmp/b12x-src; \
+    git -C /tmp/b12x-src checkout ${B12X_REF}; \
+    for pr in ${B12X_PRS}; do \
+        git -C /tmp/b12x-src fetch origin "refs/pull/$pr/head:prs/$pr"; \
+        git -C /tmp/b12x-src merge --no-ff -m "PR $pr" "prs/$pr"; \
+    done
+
+RUN set -eux; \
+    cd /tmp/vllm-src; \
+    pip install setuptools-rust; \
+    apt-get update && apt-get install -y --no-install-recommends ninja-build && rm -rf /var/lib/apt/lists/*; \
+    pip install cmake; \
+    # cu13 wheels carry missing dev headers (cusparse.h etc.) outside the
+    # default include path — symlink only what /usr/local/cuda/include lacks.
+    INC=/usr/local/lib/python3.12/dist-packages/nvidia/cu13/include; \
+    for h in "$INC"/*.h; do b=$(basename "$h"); [ -e "/usr/local/cuda/include/$b" ] || ln -s "$h" /usr/local/cuda/include/$b; done; \
+    # The spinloop target needs an explicit nvrtc: the cu13 toolkit ships
+    # under targets/sbsa-linux (ldconfig knows it, FindCUDA does not).
+    NVRTC=$(ls /usr/local/cuda*/targets/sbsa-linux/lib/libnvrtc.so* | head -1); \
+    CMAKE_ARGS="-DCUDAToolkit_ROOT=/usr/local/cuda -DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda -DCUDA_nvrtc_LIBRARY=${NVRTC} -DCUDAToolkit_LIBRARY_DIR=/usr/local/cuda/targets/sbsa-linux/lib" \
+    VLLM_TARGET_DEVICE=cuda TORCH_CUDA_ARCH_LIST="12.1a" \
+        pip install --no-build-isolation --no-deps . 2>&1; \
+    python3 -c "import vllm; print('vllm import ok', vllm.__version__)"
+
+# B12X builds its own CUDA extension set; compile gate fails loudly on
+# anchor drift. Keep base-provided deps untouched (--no-deps).
+RUN set -eux; \
+    cd /tmp/b12x-src; \
+    TORCH_CUDA_ARCH_LIST="12.1f;12.1a" \
+        pip install --no-build-isolation --no-deps . 2>&1; \
+    python3 -c "import b12x; print('b12x import ok', b12x.__version__)" || \
+    python3 -c "import importlib.util; \
+        assert importlib.util.find_spec('b12x'), 'b12x module not found'; \
+        print('b12x module present')"
+
+RUN set -eux; python3 -c "import vllm; print('vllm', vllm.__version__)"
+
+FROM ${BASE_IMAGE}
+
+# Compiled artifacts only (site-packages), dropped onto a pristine base so
+# the container keeps the day-0 entrypoint/env layout.
+COPY --from=build /usr/local/lib/python3.12/dist-packages/ /usr/local/lib/python3.12/dist-packages/
+
+RUN set -eux; \
+    python3 -c "import vllm; print('final vllm', vllm.__version__)"

docker-images/glm53-jovian-r12/build.sh

@@ -0,0 +1,24 @@
+#!/usr/bin/env bash
+# Build the SM121 GLM-5.3-Flash NVFP4 (lab quant) image: day-0 base +
+# NoPE-MLA mod + FujitsuPolycom lab-checkpoint patches.
+#
+# Generic driver: forwards every KEY=VAL pin in versions.env as a --build-arg
+# to the Dockerfile. TAG is set by the sm121 launcher to the recipe's
+# docker_image (manual build: TAG=glm53-flash-nvfp4-lab:local bash build.sh).
+set -euo pipefail
+cd "$(dirname "$0")"
+TAG="${TAG:-glm53-flash-nvfp4-lab:local}"
+
+ARGS=()
+while IFS= read -r line; do
+    case "$line" in
+        ''|\#*) continue ;;
+    esac
+    key="${line%%=*}"; val="${line#*=}"
+    val="${val%\"}"; val="${val#\"}"   # versions.env values may be shell-quoted
+    [ -n "$key" ] && ARGS+=(--build-arg "${key}=${val}")
+done < versions.env
+
+echo "=== Building ${TAG} ==="
+DOCKER_BUILDKIT=1 docker build "${ARGS[@]}" --progress=plain -t "${TAG}" -f Dockerfile . 2>&1
+echo "=== Done: ${TAG} ==="

docker-images/glm53-jovian-r12/versions.env

@@ -0,0 +1,8 @@
+# glm53-jovian-r12 build inputs — rtx6kpro models/glm-5.3-flash.md pins.
+# digest-pinned day-0 arm64 cu130 base (same as glm53-flash-nvfp4-lab).
+# NOTE: the doc's "composition commits" (7d8a09c… / d064ca4c…) are not
+# public refs; we reproduce them as base + sequential PR-head merges —
+# so here the REFs are the upstream BASES (branch tip / master).
+BASE_IMAGE="vllm/vllm-openai:glm53-flash-arm64-cu130@sha256:905c02933be6021301db2dc284e24e3727467aa3a0f63b41d609885778a07bce"
+VLLM_REF="54f6e9826c20ef06ed65d838c0ad497ad0abdecf"
+B12X_REF="139e04048bc3bb4f7210c99e7184d8d2f0e345e7"