Enable GPU if available in Makefile

Also some podman-compatibility fixes, trying to compare why GPU
access works in podman and not docker and vice versa for debugging
reasons.

Also, make "docker run" a shell script, mainly so we can run
shellcheck on it.

Signed-off-by: Eric Curtin <eric.curtin@docker.com>

Author: ericcurtin

Dockerfile

@@ -3,10 +3,11 @@
 ARG GO_VERSION=1.24.2
 ARG LLAMA_SERVER_VERSION=latest
 ARG LLAMA_SERVER_VARIANT=cpu
+ARG TARGETARCH=${BUILDARCH}
 ARG LLAMA_BINARY_PATH=/com.docker.llama-server.native.linux.${LLAMA_SERVER_VARIANT}.${TARGETARCH}
 ARG BASE_IMAGE=ubuntu:24.04
 
-FROM golang:${GO_VERSION}-bookworm AS builder
+FROM docker.io/library/golang:${GO_VERSION}-bookworm AS builder
 
 # Install git for go mod download if needed
 RUN apt-get update && apt-get install -y --no-install-recommends git && rm -rf /var/lib/apt/lists/*
@@ -33,7 +34,7 @@ RUN --mount=type=cache,target=/go/pkg/mod \
 FROM docker/docker-model-backend-llamacpp:${LLAMA_SERVER_VERSION}-${LLAMA_SERVER_VARIANT} AS llama-server
 
 # --- Final image ---
-FROM ${BASE_IMAGE} AS final
+FROM docker.io/${BASE_IMAGE} AS final
 
 ARG LLAMA_SERVER_VARIANT
 

Makefile

@@ -50,18 +50,7 @@ docker-run: docker-build
 	@echo "Service will be available at: http://localhost:$(PORT)"
 	@echo "Example usage: curl http://localhost:$(PORT)/models"
 	@echo ""
-	mkdir -p $(MODELS_PATH)
-	docker run --rm \
-		-p $(PORT):$(PORT) \
-		-v "$(MODELS_PATH):/models" \
-		-e MODEL_RUNNER_PORT=$(PORT) \
-		-e LLAMA_SERVER_PATH=/app/bin \
-		-e MODELS_PATH=/models \
-		-e LLAMA_ARGS="$(LLAMA_ARGS)" \
-		-e DMR_ORIGINS="$(DMR_ORIGINS)" \
-		-e DO_NOT_TRACK=${DO_NOT_TRACK} \
-		-e DEBUG=${DEBUG} \
-		$(DOCKER_IMAGE)
+	scripts/docker-run.sh
 
 # Show help
 help:

scripts/docker-run.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+main() {
+  set -eux -o pipefail
+
+  local gpu_device_flag=()
+  for i in /dev/dri /dev/kfd; do
+    if [ -e "$i" ]; then
+      gpu_device_flag+=("--device" "$i")
+    fi
+  done
+
+  mkdir -p "$MODELS_PATH"
+  chmod a+rx "$MODELS_PATH"
+  docker run --rm \
+    -p "$PORT:$PORT" \
+    -v "$MODELS_PATH:/models" \
+    -e MODEL_RUNNER_PORT="$PORT" \
+    -e LLAMA_SERVER_PATH=/app/bin \
+    -e MODELS_PATH=/models \
+    -e LLAMA_ARGS="$LLAMA_ARGS" \
+    -e DMR_ORIGINS="$DMR_ORIGINS" \
+    -e DO_NOT_TRACK="$DO_NOT_TRACK" \
+    -e DEBUG="$DEBUG" \
+    --gpus all \
+    "${gpu_device_flag[@]}"
+    "$DOCKER_IMAGE"
+}
+
+main "$@"
+