fix: fix dual-gpu port conflict + move GPU config to overlay files

- Remove ollama-gpu service (was colliding with ollama on port 11434)
- Strip inline deploy.resources GPU blocks from vision and vllm
- Add compose.gpu.yml: Docker NVIDIA overlay for ollama (GPU 0),
  vision (GPU 0), vllm (GPU 1), finetune (GPU 0)
- Fix compose.podman-gpu.yml: rename ollama-gpu → ollama to match
  service name after removal of ollama-gpu
- Update Makefile: apply compose.gpu.yml for Docker + GPU profiles
  (was only applying podman-gpu.yml for Podman + GPU profiles)

Makefile

@@ -15,13 +15,18 @@ COMPOSE ?= $(shell \
       && echo "podman compose" \
       || echo "podman-compose"))
 
-# GPU profiles on Podman require a CDI override (rootless Podman can't use driver: nvidia)
+# GPU profiles require an overlay for NVIDIA device reservations.
-# Generate CDI spec first: sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml
+# Docker uses deploy.resources (compose.gpu.yml); Podman uses CDI device specs (compose.podman-gpu.yml).
+# Generate CDI spec for Podman first: sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml
 COMPOSE_FILES := -f compose.yml
 ifneq (,$(findstring podman,$(COMPOSE)))
   ifneq (,$(findstring gpu,$(PROFILE)))
     COMPOSE_FILES := -f compose.yml -f compose.podman-gpu.yml
   endif
+else
+  ifneq (,$(findstring gpu,$(PROFILE)))
+    COMPOSE_FILES := -f compose.yml -f compose.gpu.yml
+  endif
 endif
 
 setup:          ## Install dependencies (Docker or Podman + NVIDIA toolkit)

compose.gpu.yml

@@ -0,0 +1,46 @@
+# compose.gpu.yml — Docker NVIDIA GPU overlay
+#
+# Adds NVIDIA GPU reservations to Peregrine services.
+# Applied automatically by `make start PROFILE=single-gpu|dual-gpu` when Docker is detected.
+# Manual: docker compose -f compose.yml -f compose.gpu.yml --profile single-gpu up -d
+#
+# Prerequisites:
+#   sudo nvidia-ctk runtime configure --runtime=docker
+#   sudo systemctl restart docker
+#
+services:
+  ollama:
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ["0"]
+              capabilities: [gpu]
+
+  vision:
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ["0"]
+              capabilities: [gpu]
+
+  vllm:
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ["1"]
+              capabilities: [gpu]
+
+  finetune:
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ["0"]
+              capabilities: [gpu]

compose.podman-gpu.yml

@@ -1,7 +1,7 @@
 # compose.podman-gpu.yml — Podman GPU override
 #
 # Replaces Docker-specific `driver: nvidia` reservations with CDI device specs
-# for rootless Podman. Apply automatically via `make start PROFILE=single-gpu`
+# for rootless Podman. Applied automatically via `make start PROFILE=single-gpu|dual-gpu`
 # when podman/podman-compose is detected, or manually:
 #   podman-compose -f compose.yml -f compose.podman-gpu.yml --profile single-gpu up -d
 #
@@ -10,7 +10,7 @@
 #   (requires nvidia-container-toolkit >= 1.14)
 #
 services:
-  ollama-gpu:
+  ollama:
     devices:
       - nvidia.com/gpu=0
     deploy:

compose.yml

@@ -48,18 +48,6 @@ services:
     profiles: [cpu, single-gpu, dual-gpu]
     restart: unless-stopped
 
-  ollama-gpu:
-    extends:
-      service: ollama
-    deploy:
-      resources:
-        reservations:
-          devices:
-            - driver: nvidia
-              device_ids: ["0"]
-              capabilities: [gpu]
-    profiles: [single-gpu, dual-gpu]
-
   vision:
     build:
       context: .
@@ -69,13 +57,6 @@ services:
     environment:
       - VISION_MODEL=${VISION_MODEL:-vikhyatk/moondream2}
       - VISION_REVISION=${VISION_REVISION:-2025-01-09}
-    deploy:
-      resources:
-        reservations:
-          devices:
-            - driver: nvidia
-              device_ids: ["0"]
-              capabilities: [gpu]
     profiles: [single-gpu, dual-gpu]
     restart: unless-stopped
 
@@ -93,13 +74,6 @@ services:
       --enforce-eager
       --max-num-seqs 8
       --cpu-offload-gb ${CPU_OFFLOAD_GB:-0}
-    deploy:
-      resources:
-        reservations:
-          devices:
-            - driver: nvidia
-              device_ids: ["1"]
-              capabilities: [gpu]
     profiles: [dual-gpu]
     restart: unless-stopped