Follow-up to ae259a9f58 after an independent review:
- The 136-character task name broke the 80-character name lint
(tests/lint/repository/test_name_emoji_and_length.py). Names are short
again; the device/driver/host matrix lives in the role README.
- Both stats now always run and lookup('gpu') requires both registers,
instead of reading a skipped device stat as "no device".
- A bare host that shows /dev/nvidia0 without nvidia-smi stops the
deploy with a message naming the fix. It used to die on the missing
binary anyway; ae259a9f58 would have run its GPU services on the CPU
silently. A container in that state (a swarm lab node, the dev
container without its GPU override) still runs on the CPU.
- svc-ai-s1 decided its CUDA build from its own stat of /dev/nvidia0 and
nvidia-container-runtime while its runtime came from lookup('gpu'), so
a host with the toolkit but no nvidia-smi would have built a CUDA image
that gets no GPU. S1_GPU now reads lookup('gpu'), the one detection.
- The libretranslate-engine README and the s1 TODO describe the new
detection.
test_gpu_detection walks the real tasks over device, driver and
container; against the original 00_core.yml it fails in the
driverless-container, driverless-bare and device-and-driver cells.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The local 061 Nextcloud swarm step failed on every lab node in
"Ask the driver which CUDA devices this host carries" with
"[Errno 2] No such file or directory: b'nvidia-smi'". The lab nodes are
privileged containers (compose/swarm/compose.yml), so on a workstation
with an NVIDIA card they see the host's /dev/nvidia0, but they carry no
driver. The device node alone made them GPU hosts, for the GPU include
and for lookup('gpu') alike, which would also have picked a -cuda image
and runtime: nvidia for a GPU service there.
The role now stats SYS_SVC_CONTAINER_GPU_DRIVER (/usr/bin/nvidia-smi)
first and stats the device node only where the driver exists; the
skipped register reads as "no device" in lookup('gpu'), and the include
requires a device stat that ran and found the node. The query runs the
same binary the detection probed. Hosts with driver and device (bare
metal with the driver, the compose dev container through the GPU
override) and hosts without a device (CI) behave as before.
test_gpu_detection walks the real tasks over the four driver/device
cells and fails on the old 00_core.yml in exactly the device-without-
driver cell.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
make autoformat reflowed a resource_filter call, a generator expression and five test files it touched along the way. The new architecture test also carried the two findings the first gate run stopped on: three RUF012, where a TestCase class attribute needs ClassVar, and two PERF401, where an append loop becomes list.extend over a generator.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Whether a service gets a GPU needs two facts to agree: the declaration in
applications under services.<service>.gpu, and sys-svc-container's stat of
/dev/nvidia0 on the target host. The conjunction was restated at each call
site, and web-svc-libretranslate restated only the first half: it appended the
-cuda image suffix from the declaration alone. That tag has no arm64 manifest,
so the build died with `no match for platform in manifest`, which 10 of the 79
failing jobs of run 37252453141 carry, surfacing through
sys-svc-compose/handlers/{swarm,compose}.yml.
lookup('gpu', application_id[, service]) is now the single place that answers
it. Three call sites collapse onto it: both libretranslate roles' *_GPU var
and the runtime line of sys-svc-container's resource template. A lookup runs
on the controller and cannot stat the target, so it reads the register out of
the play's variables; the applications view comes through lookup_loader
because test_no_cache_imports_in_lookups forbids a lookup the direct
utils.cache import.
The coercion is ansible's boolean(), not python's bool(). The call sites this
replaces read the flag through Jinja's `| bool`, which answers False for the
strings "false", "False", "0" and "no" where bool() answers True, and a
templated declaration renders to such a string. Measured over 13 values the
two disagree on four of them; boolean(strict=False) matches `| bool` on all
13. Two tests pin it, and with bool() restored exactly the "false" string test
fails.
Measured rather than argued: over the merged applications, 1688
(application, service) pairs resolve the predicate identically under both
coercions, 0 disagreements. The predicate turns on for 21 services, not 5,
because resource_filter falls back to services.<entity_name>.gpu and every
service of a role whose entity declares gpu inherits it.
The assert_gpu include in both roles dropped its two trailing conditions: they
re-checked the register that *_GPU now already contains.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
kevinveenbirkenbachandClaude Opus 5.5