feat(irv-ml1): R15 + R16 brokkr-smithy probe-harness stand-up playbooks

No-sudo phase: uv, R16 world venv + pyworld (setuptools<81 pin for pyworld's pkg_resources import). Sudo phase (run as infra-ops): R + soundgen system deps, and pin the Ollama service to the A6000 BY UUID (CUDA FASTEST_FIRST ordering != nvidia-smi PCI order on this box, so an index pin is ambiguous). R15 models share the A6000's free headroom — yt-voice-clipper + R14 keep running, card not freed.
2026-06-03 14:49:46 -07:00
parent 8c32a0540f
commit 1733cedab0
2 changed files with 105 additions and 0 deletions
@@ -0,0 +1,48 @@
+# R15 + R16 box stand-up on irv-ml1 — NO-SUDO phase.
+#
+# Provisions everything that does NOT need root: uv, the R16 `world` venv +
+# pyworld, and the engine-root dir skeleton. Two things are deliberately
+# OUT of this playbook because they need sudo (separate operator-run phase):
+#   - R + soundgen   (apt install r-base libfftw3-dev libsndfile1-dev; CRAN build)
+#   - Ollama A6000 GPU pin (systemd drop-in CUDA_VISIBLE_DEVICES=1 + restart)
+# The ollama model pulls are also run separately (long-running; ollama pull is
+# itself idempotent).
+#
+# Setup contracts (brokkr-smithy @ irv-ml1:~/r14/brokkr-smithy, pulled to main):
+#   R15 src/probes/P01/box_models.py  — ollama models @ $R15_OLLAMA_HOST (:11434)
+#   R16 src/probes/P01/box_engines.py — R16_ENGINE_ROOT default ~/r16/engines;
+#       world venv at $ROOT/world/.venv (pyworld); soundgen runner at $ROOT/soundgen
+#
+# Rerunnable: a second run should report mostly ok/skipped.
+
+vars:
+  engine_root: /home/lkraven/r16/engines
+  uv_bin: /home/lkraven/.local/bin/uv
+
+steps:
+  - name: Install uv (user-local, no sudo)
+    shell: curl -LsSf https://astral.sh/uv/install.sh | sh
+    creates: "{{ uv_bin }}"
+
+  - name: Create R16 engine-root dirs (soundgen + world)
+    shell: mkdir -p {{ engine_root }}/soundgen {{ engine_root }}/world
+    creates: "{{ engine_root }}/world"
+
+  - name: Create the R16 world venv (python3.11)
+    shell: "{{ uv_bin }} venv {{ engine_root }}/world/.venv --python 3.11"
+    creates: "{{ engine_root }}/world/.venv/bin/python"
+
+  - name: Install pyworld (+ setuptools<81 — pyworld imports pkg_resources, removed in setuptools 81)
+    shell: "{{ uv_bin }} pip install --python {{ engine_root }}/world/.venv/bin/python pyworld 'setuptools<81'"
+    when: "! {{ engine_root }}/world/.venv/bin/python -c 'import pyworld' 2>/dev/null"
+
+verify:
+  - name: uv present
+    shell: test -x {{ uv_bin }}
+    changed_when: "false"
+  - name: world venv python present
+    shell: test -x {{ engine_root }}/world/.venv/bin/python
+    changed_when: "false"
+  - name: pyworld importable in the world venv
+    shell: "{{ engine_root }}/world/.venv/bin/python -c 'import pyworld; print(\"pyworld\", getattr(pyworld, \"__version__\", \"ok\"))'"
+    changed_when: "false"
@@ -0,0 +1,57 @@
+# R15 + R16 box stand-up on irv-ml1 — SUDO phase (run AS infra-ops).
+#
+#   scripts/elway infra-ops@10.100.79.3 --playbook playbooks/irv-ml1-r15-r16-sudo.yaml
+#
+# infra-ops has NOPASSWD sudo (see playbooks/bootstrap-infra-ops-user.yaml), so
+# this runs unattended — no operator password prompt.
+#
+# Does two things:
+#   1. R16 soundgen engine: install R + the system libs soundgen's CRAN deps
+#      need (fftw/sndfile + the usual R-build trio). The soundgen package itself
+#      compiles separately (slow) — kicked outside this playbook.
+#   2. R15 GPU placement: pin the Ollama service to the A6000 so inference lands
+#      in the A6000's free headroom. yt-voice-clipper + R14 keep running on the
+#      same card — we share the ~31.5 GB free, we do NOT free the card.
+#
+# A6000 pinned BY UUID, not index. CUDA's default ordering (FASTEST_FIRST) does
+# not match nvidia-smi's PCI order on this box (R14 sees the A6000 as cuda:0
+# while nvidia-smi lists it as index 1), so an index pin is ambiguous and could
+# select the 3090. The UUID is unambiguous.
+
+vars:
+  a6000_uuid: "GPU-9672f0d5-3caf-b15c-898a-cd9d653c654b"
+
+steps:
+  - name: apt update
+    sudo: true
+    shell: apt-get update -qq
+
+  - name: Install R + soundgen system deps (fftw/sndfile + R-build libs)
+    sudo: true
+    shell: >
+      DEBIAN_FRONTEND=noninteractive apt-get install -y
+      r-base libfftw3-dev libsndfile1-dev
+      libcurl4-openssl-dev libssl-dev libxml2-dev
+    when: "! command -v Rscript >/dev/null 2>&1"
+
+  - name: Pin the Ollama service to the A6000 (UUID) via systemd drop-in
+    sudo: true
+    shell: |
+      install -d -m 755 /etc/systemd/system/ollama.service.d
+      printf '[Service]\nEnvironment=CUDA_VISIBLE_DEVICES=%s\n' '{{ a6000_uuid }}' \
+        > /etc/systemd/system/ollama.service.d/gpu-pin.conf
+      systemctl daemon-reload
+      systemctl restart ollama
+    when: "! systemctl show ollama -p Environment 2>/dev/null | grep -q '{{ a6000_uuid }}'"
+
+verify:
+  - name: Rscript present
+    shell: command -v Rscript
+    changed_when: "false"
+  - name: Ollama service carries the A6000 UUID pin
+    sudo: true
+    shell: systemctl show ollama -p Environment | grep -q '{{ a6000_uuid }}'
+    changed_when: "false"
+  - name: Ollama API healthy on :11434 after restart
+    shell: curl -sf http://localhost:11434/api/tags >/dev/null && echo ok
+    changed_when: "false"