vh
5cd4678d9f
Default gcc is gcc-11 (CUDA/torch R14 pin) but the only gfortran is 12, whose libgfortran-12-dev puts the unversioned libgfortran.so only in the gcc-12 dir. R links shared objects with gcc-11 + -lgfortran, so the whole Fortran dependency chain (quadprog/tseries/RcppArmadillo -> soundgen) failed 'cannot find -lgfortran'. Adds an idempotent step installing libgfortran-11-dev (guarded on 'gcc -print-file-name=libgfortran.so' not resolving). Additive; default toolchain untouched.
70 lines
3.1 KiB
YAML
70 lines
3.1 KiB
YAML
# 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"
|
|
|
|
# gcc/gfortran version mismatch fix. irv-ml1's default `gcc` is gcc-11
|
|
# (deliberate — the CUDA/torch R14 stack pins it), but the only gfortran is
|
|
# gfortran-12, whose libgfortran-12-dev puts the unversioned `libgfortran.so`
|
|
# ONLY in the gcc-12 dir. R links shared objects with SHLIB_LD=$(CC)=gcc-11 +
|
|
# `-lgfortran`, so every Fortran package (quadprog→tseries→soundgen, etc.)
|
|
# fails "cannot find -lgfortran". libgfortran-11-dev adds the symlink to the
|
|
# gcc-11 dir. Additive — does NOT change the default toolchain.
|
|
- name: Ensure the default gcc's Fortran link lib resolves (gcc-11 here)
|
|
sudo: true
|
|
shell: DEBIAN_FRONTEND=noninteractive apt-get install -y libgfortran-11-dev
|
|
when: "! gcc -print-file-name=libgfortran.so | grep -q /"
|
|
|
|
- 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"
|