fix(compose): reserve a GPU only where the daemon has one

A device reservation is not conditional in compose, and a count of 0 does not disable it: the daemon still selects the nvidia driver and fails where there is none. Every GPU-less host therefore lost compose up with 'could not select device driver', which is what broke test-dns on all five CI distros.

Move the reservation out of compose.yml and i18n/compose.override.yml into compose/gpu.override.yml, layered by compose_file_args() only when Profile.gpu_reservable() sees a non-zero INFINITO_GPU_COUNT. A GPU host renders exactly what it rendered before, so the passthrough sys-svc-container needs to find /dev/nvidia0 and emit runtime: nvidia for a gpu:true service stays intact.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

cli/administration/deploy/development/env.py

@@ -25,6 +25,8 @@ def compose_file_args() -> list[str]:
         out += ["-f", "compose/cache.override.yml"]
         if profile.shared_cache_network():
             out += ["-f", "compose/cache.shared.override.yml"]
+    if profile.gpu_reservable():
+        out += ["-f", "compose/gpu.override.yml"]
     if (os.environ.get("INFINITO_PUBLISH_PORTS") or "").strip().lower() == "false":
         out += ["-f", "compose/noports.override.yml"]
     out += ["-f", "i18n/compose.override.yml"]

cli/administration/deploy/development/profile.py

@@ -35,6 +35,10 @@ class Profile:
         declared = _declared("INFINITO_CACHE_STACK")
         return (not self.is_ci()) if declared is None else declared
 
+    def gpu_reservable(self) -> bool:
+        """True iff the docker daemon registers an NVIDIA runtime."""
+        return (os.environ.get("INFINITO_GPU_COUNT") or "0").strip() not in ("", "0")
+
     def image_mirror_enabled(self) -> bool:
         """True iff image references should be rewritten to the GHCR mirror."""
         return self.runs_on_github()

compose.yml

@@ -61,13 +61,6 @@ services:
     mem_limit: ${INFINITO_MEM_LIMIT:?Run 'make dotenv' to generate the .env single source of truth}
     memswap_limit: ${INFINITO_MEMSWAP_LIMIT:?Run 'make dotenv' to generate the .env single source of truth}
     cpus: ${INFINITO_CPUS:?Run 'make dotenv' to generate the .env single source of truth}
-    deploy:
-      resources:
-        reservations:
-          devices:
-            - driver: nvidia
-              count: ${INFINITO_GPU_COUNT:?Run 'make dotenv' to generate the .env single source of truth}
-              capabilities: [gpu]
     <<: *ci_defaults
     volumes:
       - /sys/fs/cgroup:/sys/fs/cgroup:rw

compose/gpu.override.yml

@@ -0,0 +1,24 @@
+---
+# Layered on compose.yml by the dev tooling when the docker daemon registers an
+# NVIDIA runtime. A count of 0 does not disable the reservation, the daemon still
+# selects the driver and fails where there is none, so a GPU-less host must not
+# carry this file at all.
+
+services:
+  infinito:
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: ${INFINITO_GPU_COUNT:?Run 'make dotenv' to generate the .env single source of truth}
+              capabilities: [gpu]
+
+  i18n:
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: ${INFINITO_GPU_COUNT:?Run 'make dotenv' to generate the .env single source of truth}
+              capabilities: [gpu]

i18n/compose.override.yml

@@ -44,13 +44,6 @@ services:
       - "${INFINITO_BIND_IP:?Run 'make dotenv' to generate the .env single source of truth}:${INFINITO_I18N_LIBRETRANSLATE_PORT:?Run 'make dotenv' to generate the
         .env single source of truth}:${INFINITO_I18N_LIBRETRANSLATE_PORT:?Run 'make dotenv' to generate the .env single source of truth}"
 
-    deploy:
-      resources:
-        reservations:
-          devices:
-            - driver: nvidia
-              count: ${INFINITO_GPU_COUNT:?Run 'make dotenv' to generate the .env single source of truth}
-              capabilities: [gpu]
     networks:
       default:
 volumes: