Remove old diffusers distribution (#713)

* fix: remove old diffusers image build steps from release workflow

* fix: implement installation for diffusers backend in model runner

Author: ilopezluna

.github/workflows/release.yml

@@ -261,12 +261,6 @@ jobs:
             echo "docker/model-runner:latest-sglang-cuda" >> "$GITHUB_OUTPUT"
           fi
           echo 'EOF' >> "$GITHUB_OUTPUT"
-          echo "diffusers<<EOF" >> "$GITHUB_OUTPUT"
-          echo "docker/model-runner:$RELEASE_TAG-diffusers" >> "$GITHUB_OUTPUT"
-          if [ "$PUSH_LATEST" == "true" ]; then
-            echo "docker/model-runner:latest-diffusers" >> "$GITHUB_OUTPUT"
-          fi
-          echo 'EOF' >> "$GITHUB_OUTPUT"
           echo "rocm<<EOF" >> "$GITHUB_OUTPUT"
           echo "docker/model-runner:$RELEASE_TAG-rocm" >> "$GITHUB_OUTPUT"
           if [ "$PUSH_LATEST" == "true" ]; then
@@ -366,20 +360,6 @@ jobs:
           provenance: mode=max
           tags: ${{ steps.tags.outputs.sglang-cuda }}
 
-      - name: Build Diffusers image
-        uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8
-        with:
-          file: Dockerfile
-          target: final-diffusers
-          platforms: linux/amd64, linux/arm64
-          build-args: |
-            "LLAMA_SERVER_VERSION=${{ env.LLAMA_SERVER_VERSION }}"
-            "VERSION=${{ env.RELEASE_TAG }}"
-          push: true
-          sbom: true
-          provenance: mode=max
-          tags: ${{ steps.tags.outputs.diffusers }}
-
       - name: Build ROCm image
         uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8
         with:

cmd/cli/commands/install-runner.go

@@ -253,6 +253,28 @@ func runInstallOrStart(cmd *cobra.Command, opts runnerOptions, debug bool) error
 		return nil
 	}
 
+	// The diffusers backend uses deferred installation: it pulls a Docker
+	// image, extracts a self-contained Python environment, and installs it
+	// to a well-known local folder. Trigger installation via the running
+	// model runner's API, the same way vllm-metal is handled above.
+	if opts.backend == diffusers.Name && platform.SupportsDiffusers() {
+		// For standalone contexts (Moby/Cloud), ensure a base runner is
+		// available first so we have an API endpoint to call.
+		engineKind := modelRunner.EngineKind()
+		if engineKind == types.ModelRunnerEngineKindMoby || engineKind == types.ModelRunnerEngineKindCloud {
+			if _, err := ensureStandaloneRunnerAvailable(cmd.Context(), asPrinter(cmd), debug); err != nil {
+				return fmt.Errorf("unable to initialize standalone model runner: %w", err)
+			}
+		}
+
+		cmd.Println("Installing diffusers backend...")
+		if err := desktopClient.InstallBackend(diffusers.Name); err != nil {
+			return fmt.Errorf("failed to install diffusers backend: %w", err)
+		}
+		cmd.Println("diffusers backend installed successfully")
+		return nil
+	}
+
 	var vllmOnWSL bool
 	// Ensure that we're running in a supported model runner context.
 	engineKind := modelRunner.EngineKind()