Add stable diffusion backend

Dockerfile

@@ -75,7 +75,7 @@ ENV MODEL_RUNNER_PORT=12434
 ENV LLAMA_SERVER_PATH=/app/bin
 ENV HOME=/home/modelrunner
 ENV MODELS_PATH=/models
-ENV LD_LIBRARY_PATH=/app/lib:$LD_LIBRARY_PATH
+ENV LD_LIBRARY_PATH=/app/lib
 
 # Label the image so that it's hidden on cloud engines.
 LABEL com.docker.desktop.service="model-runner"
@@ -144,6 +144,60 @@ RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
     && ~/.local/bin/uv pip install --python /opt/sglang-env/bin/python "sglang==${SGLANG_VERSION}"
 
 RUN /opt/sglang-env/bin/python -c "import sglang; print(sglang.__version__)" > /opt/sglang-env/version
+
+# --- Diffusers variant ---
+FROM llamacpp AS diffusers
+
+# Python package versions for reproducible builds
+ARG DIFFUSERS_VERSION=0.36.0
+ARG TORCH_VERSION=2.9.1
+ARG TRANSFORMERS_VERSION=4.57.5
+ARG ACCELERATE_VERSION=1.3.0
+ARG SAFETENSORS_VERSION=0.5.2
+ARG HUGGINGFACE_HUB_VERSION=0.34.0
+ARG BITSANDBYTES_VERSION=0.49.1
+ARG FASTAPI_VERSION=0.115.12
+ARG UVICORN_VERSION=0.34.1
+ARG PILLOW_VERSION=11.2.1
+
+USER root
+
+RUN apt update && apt install -y \
+    python3 python3-venv python3-dev \
+    curl ca-certificates build-essential \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN mkdir -p /opt/diffusers-env && chown -R modelrunner:modelrunner /opt/diffusers-env
+
+USER modelrunner
+
+# Install uv and diffusers as modelrunner user
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && ~/.local/bin/uv venv --python /usr/bin/python3 /opt/diffusers-env \
+    && ~/.local/bin/uv pip install --python /opt/diffusers-env/bin/python \
+    "diffusers==${DIFFUSERS_VERSION}" \
+    "torch==${TORCH_VERSION}" \
+    "transformers==${TRANSFORMERS_VERSION}" \
+    "accelerate==${ACCELERATE_VERSION}" \
+    "safetensors==${SAFETENSORS_VERSION}" \
+    "huggingface_hub==${HUGGINGFACE_HUB_VERSION}" \
+    "bitsandbytes==${BITSANDBYTES_VERSION}" \
+    "fastapi==${FASTAPI_VERSION}" \
+    "uvicorn[standard]==${UVICORN_VERSION}" \
+    "pillow==${PILLOW_VERSION}"
+
+# Copy Python server code
+USER root
+COPY python/diffusers_server /tmp/diffusers_server/
+RUN PYTHON_SITE_PACKAGES=$(/opt/diffusers-env/bin/python -c "import site; print(site.getsitepackages()[0])") && \
+    mkdir -p "$PYTHON_SITE_PACKAGES/diffusers_server" && \
+    cp -r /tmp/diffusers_server/* "$PYTHON_SITE_PACKAGES/diffusers_server/" && \
+    chown -R modelrunner:modelrunner "$PYTHON_SITE_PACKAGES/diffusers_server/" && \
+    rm -rf /tmp/diffusers_server
+USER modelrunner
+
+RUN /opt/diffusers-env/bin/python -c "import diffusers; print(diffusers.__version__)" > /opt/diffusers-env/version
+
 FROM llamacpp AS final-llamacpp
 # Copy the built binary from builder
 COPY --from=builder /app/model-runner /app/model-runner
@@ -155,3 +209,7 @@ COPY --from=builder /app/model-runner /app/model-runner
 FROM sglang AS final-sglang
 # Copy the built binary from builder-sglang (without vLLM)
 COPY --from=builder-sglang /app/model-runner /app/model-runner
+
+FROM diffusers AS final-diffusers
+# Copy the built binary from builder (with diffusers support)
+COPY --from=builder /app/model-runner /app/model-runner

Makefile

@@ -8,6 +8,7 @@ VLLM_BASE_IMAGE := nvidia/cuda:13.0.2-runtime-ubuntu24.04
 DOCKER_IMAGE := docker/model-runner:latest
 DOCKER_IMAGE_VLLM := docker/model-runner:latest-vllm-cuda
 DOCKER_IMAGE_SGLANG := docker/model-runner:latest-sglang
+DOCKER_IMAGE_DIFFUSERS := docker/model-runner:latest-diffusers
 DOCKER_TARGET ?= final-llamacpp
 PORT := 8080
 MODELS_PATH := $(shell pwd)/models-store
@@ -25,7 +26,7 @@ DOCKER_BUILD_ARGS := \
 BUILD_DMR ?= 1
 
 # Main targets
-.PHONY: build run clean test integration-tests test-docker-ce-installation docker-build docker-build-multiplatform docker-run docker-build-vllm docker-run-vllm docker-build-sglang docker-run-sglang docker-run-impl help validate lint
+.PHONY: build run clean test integration-tests test-docker-ce-installation docker-build docker-build-multiplatform docker-run docker-build-vllm docker-run-vllm docker-build-sglang docker-run-sglang docker-run-impl help validate lint docker-build-diffusers docker-run-diffusers
 # Default target
 .DEFAULT_GOAL := build
 
@@ -117,6 +118,16 @@ docker-build-sglang:
 docker-run-sglang: docker-build-sglang
 	@$(MAKE) -s docker-run-impl DOCKER_IMAGE=$(DOCKER_IMAGE_SGLANG)
 
+# Build Diffusers Docker image
+docker-build-diffusers:
+	@$(MAKE) docker-build \
+		DOCKER_TARGET=final-diffusers \
+		DOCKER_IMAGE=$(DOCKER_IMAGE_DIFFUSERS)
+
+# Run Diffusers Docker container with TCP port access and mounted model storage
+docker-run-diffusers: docker-build-diffusers
+	@$(MAKE) -s docker-run-impl DOCKER_IMAGE=$(DOCKER_IMAGE_DIFFUSERS)
+
 # Common implementation for running Docker container
 docker-run-impl:
 	@echo ""
@@ -151,6 +162,8 @@ help:
 	@echo "  docker-run-vllm		- Run vLLM Docker container"
 	@echo "  docker-build-sglang		- Build SGLang Docker image"
 	@echo "  docker-run-sglang		- Run SGLang Docker container"
+	@echo "  docker-build-diffusers	- Build Diffusers Docker image"
+	@echo "  docker-run-diffusers		- Run Diffusers Docker container"
 	@echo "  help				- Show this help message"
 	@echo ""
 	@echo "Backend configuration options:"

cmd/cli/commands/compose_test.go

@@ -45,6 +45,12 @@ func TestParseBackendMode(t *testing.T) {
 			expected:    inference.BackendModeReranking,
 			expectError: false,
 		},
+		{
+			name:        "image-generation mode",
+			input:       "image-generation",
+			expected:    inference.BackendModeImageGeneration,
+			expectError: false,
+		},
 		{
 			name:        "invalid mode",
 			input:       "invalid",

cmd/cli/commands/configure_flags.go

@@ -146,7 +146,7 @@ func (f *ConfigureFlags) RegisterFlags(cmd *cobra.Command) {
 	cmd.Flags().StringVar(&f.HFOverrides, "hf_overrides", "", "HuggingFace model config overrides (JSON) - vLLM only")
 	cmd.Flags().Var(NewFloat64PtrValue(&f.GPUMemoryUtilization), "gpu-memory-utilization", "fraction of GPU memory to use for the model executor (0.0-1.0) - vLLM only")
 	cmd.Flags().Var(NewBoolPtrValue(&f.Think), "think", "enable reasoning mode for thinking models")
-	cmd.Flags().StringVar(&f.Mode, "mode", "", "backend operation mode (completion, embedding, reranking)")
+	cmd.Flags().StringVar(&f.Mode, "mode", "", "backend operation mode (completion, embedding, reranking, image-generation)")
 }
 
 // BuildConfigureRequest builds a scheduling.ConfigureRequest from the flags.
@@ -243,7 +243,9 @@ func parseBackendMode(mode string) (inference.BackendMode, error) {
 		return inference.BackendModeEmbedding, nil
 	case "reranking":
 		return inference.BackendModeReranking, nil
+	case "image-generation":
+		return inference.BackendModeImageGeneration, nil
 	default:
-		return inference.BackendModeCompletion, fmt.Errorf("invalid mode %q: must be one of completion, embedding, reranking", mode)
+		return inference.BackendModeCompletion, fmt.Errorf("invalid mode %q: must be one of completion, embedding, reranking, image-generation", mode)
 	}
 }

cmd/cli/commands/package.go

@@ -38,39 +38,43 @@ func newPackagedCmd() *cobra.Command {
 	var opts packageOptions
 
 	c := &cobra.Command{
-		Use:   "package (--gguf <path> | --safetensors-dir <path> | --from <model>) [--license <path>...] [--mmproj <path>] [--context-size <tokens>] [--push] MODEL",
-		Short: "Package a GGUF file, Safetensors directory, or existing model into a Docker model OCI artifact.",
-		Long: "Package a GGUF file, Safetensors directory, or existing model into a Docker model OCI artifact, with optional licenses and multimodal projector. The package is sent to the model-runner, unless --push is specified.\n" +
+		Use:   "package (--gguf <path> | --safetensors-dir <path> | --dduf <path> | --from <model>) [--license <path>...] [--mmproj <path>] [--context-size <tokens>] [--push] MODEL",
+		Short: "Package a GGUF file, Safetensors directory, DDUF file, or existing model into a Docker model OCI artifact.",
+		Long: "Package a GGUF file, Safetensors directory, DDUF file, or existing model into a Docker model OCI artifact, with optional licenses and multimodal projector. The package is sent to the model-runner, unless --push is specified.\n" +
 			"When packaging a sharded GGUF model, --gguf should point to the first shard. All shard files should be siblings and should include the index in the file name (e.g. model-00001-of-00015.gguf).\n" +
 			"When packaging a Safetensors model, --safetensors-dir should point to a directory containing .safetensors files and config files (*.json, merges.txt). All files will be auto-discovered and config files will be packaged into a tar archive.\n" +
+			"When packaging a DDUF file (Diffusers Unified Format), --dduf should point to a .dduf archive file.\n" +
 			"When packaging from an existing model using --from, you can modify properties like context size to create a variant of the original model.\n" +
 			"For multimodal models, use --mmproj to include a multimodal projector file.",
 		Args: func(cmd *cobra.Command, args []string) error {
 			if err := requireExactArgs(1, "package", "MODEL")(cmd, args); err != nil {
 				return err
 			}
 
-			// Validate that exactly one of --gguf, --safetensors-dir, or --from is provided (mutually exclusive)
+			// Validate that exactly one of --gguf, --safetensors-dir, --dduf, or --from is provided (mutually exclusive)
 			sourcesProvided := 0
 			if opts.ggufPath != "" {
 				sourcesProvided++
 			}
 			if opts.safetensorsDir != "" {
 				sourcesProvided++
 			}
+			if opts.ddufPath != "" {
+				sourcesProvided++
+			}
 			if opts.fromModel != "" {
 				sourcesProvided++
 			}
 
 			if sourcesProvided == 0 {
 				return fmt.Errorf(
-					"One of --gguf, --safetensors-dir, or --from is required.\n\n" +
+					"One of --gguf, --safetensors-dir, --dduf, or --from is required.\n\n" +
 						"See 'docker model package --help' for more information",
 				)
 			}
 			if sourcesProvided > 1 {
 				return fmt.Errorf(
-					"Cannot specify more than one of --gguf, --safetensors-dir, or --from. Please use only one source.\n\n" +
+					"Cannot specify more than one of --gguf, --safetensors-dir, --dduf, or --from. Please use only one source.\n\n" +
 						"See 'docker model package --help' for more information",
 				)
 			}
@@ -141,6 +145,15 @@ func newPackagedCmd() *cobra.Command {
 				}
 			}
 
+			// Validate DDUF path if provided
+			if opts.ddufPath != "" {
+				var err error
+				opts.ddufPath, err = validateAbsolutePath(opts.ddufPath, "DDUF")
+				if err != nil {
+					return err
+				}
+			}
+
 			// Validate dir-tar paths are relative (not absolute)
 			for _, dirPath := range opts.dirTarPaths {
 				if filepath.IsAbs(dirPath) {
@@ -167,6 +180,7 @@ func newPackagedCmd() *cobra.Command {
 
 	c.Flags().StringVar(&opts.ggufPath, "gguf", "", "absolute path to gguf file")
 	c.Flags().StringVar(&opts.safetensorsDir, "safetensors-dir", "", "absolute path to directory containing safetensors files and config")
+	c.Flags().StringVar(&opts.ddufPath, "dduf", "", "absolute path to DDUF archive file (Diffusers Unified Format)")
 	c.Flags().StringVar(&opts.fromModel, "from", "", "reference to an existing model to repackage")
 	c.Flags().StringVar(&opts.chatTemplatePath, "chat-template", "", "absolute path to chat template file (must be Jinja format)")
 	c.Flags().StringArrayVarP(&opts.licensePaths, "license", "l", nil, "absolute path to a license file")
@@ -182,6 +196,7 @@ type packageOptions struct {
 	contextSize      uint64
 	ggufPath         string
 	safetensorsDir   string
+	ddufPath         string
 	fromModel        string
 	licensePaths     []string
 	dirTarPaths      []string
@@ -197,7 +212,7 @@ type builderInitResult struct {
 	cleanupFunc func()               // Optional cleanup function for temporary files
 }
 
-// initializeBuilder creates a package builder from GGUF, Safetensors, or existing model
+// initializeBuilder creates a package builder from GGUF, Safetensors, DDUF, or existing model
 func initializeBuilder(cmd *cobra.Command, opts packageOptions) (*builderInitResult, error) {
 	result := &builderInitResult{}
 
@@ -246,7 +261,14 @@ func initializeBuilder(cmd *cobra.Command, opts packageOptions) (*builderInitRes
 			return nil, fmt.Errorf("add gguf file: %w", err)
 		}
 		result.builder = pkg
-	} else {
+	} else if opts.ddufPath != "" {
+		cmd.PrintErrf("Adding DDUF file from %q\n", opts.ddufPath)
+		pkg, err := builder.FromPath(opts.ddufPath)
+		if err != nil {
+			return nil, fmt.Errorf("add dduf file: %w", err)
+		}
+		result.builder = pkg
+	} else if opts.safetensorsDir != "" {
 		// Safetensors model from directory
 		cmd.PrintErrf("Scanning directory %q for safetensors model...\n", opts.safetensorsDir)
 		safetensorsPaths, tempConfigArchive, err := packaging.PackageFromDirectory(opts.safetensorsDir)
@@ -276,6 +298,8 @@ func initializeBuilder(cmd *cobra.Command, opts packageOptions) (*builderInitRes
 			}
 		}
 		result.builder = pkg
+	} else {
+		return nil, fmt.Errorf("no model source specified")
 	}
 
 	return result, nil

cmd/cli/docs/reference/docker_model_configure.yaml

@@ -40,7 +40,8 @@ options:
       swarm: false
     - option: mode
       value_type: string
-      description: backend operation mode (completion, embedding, reranking)
+      description: |
+        backend operation mode (completion, embedding, reranking, image-generation)
       deprecated: false
       hidden: false
       experimental: false

cmd/cli/docs/reference/docker_model_package.yaml

@@ -1,13 +1,14 @@
 command: docker model package
 short: |
-    Package a GGUF file, Safetensors directory, or existing model into a Docker model OCI artifact.
+    Package a GGUF file, Safetensors directory, DDUF file, or existing model into a Docker model OCI artifact.
 long: |-
-    Package a GGUF file, Safetensors directory, or existing model into a Docker model OCI artifact, with optional licenses and multimodal projector. The package is sent to the model-runner, unless --push is specified.
+    Package a GGUF file, Safetensors directory, DDUF file, or existing model into a Docker model OCI artifact, with optional licenses and multimodal projector. The package is sent to the model-runner, unless --push is specified.
     When packaging a sharded GGUF model, --gguf should point to the first shard. All shard files should be siblings and should include the index in the file name (e.g. model-00001-of-00015.gguf).
     When packaging a Safetensors model, --safetensors-dir should point to a directory containing .safetensors files and config files (*.json, merges.txt). All files will be auto-discovered and config files will be packaged into a tar archive.
+    When packaging a DDUF file (Diffusers Unified Format), --dduf should point to a .dduf archive file.
     When packaging from an existing model using --from, you can modify properties like context size to create a variant of the original model.
     For multimodal models, use --mmproj to include a multimodal projector file.
-usage: docker model package (--gguf <path> | --safetensors-dir <path> | --from <model>) [--license <path>...] [--mmproj <path>] [--context-size <tokens>] [--push] MODEL
+usage: docker model package (--gguf <path> | --safetensors-dir <path> | --dduf <path> | --from <model>) [--license <path>...] [--mmproj <path>] [--context-size <tokens>] [--push] MODEL
 pname: docker model
 plink: docker_model.yaml
 options:
@@ -30,6 +31,15 @@ options:
       experimentalcli: false
       kubernetes: false
       swarm: false
+    - option: dduf
+      value_type: string
+      description: absolute path to DDUF archive file (Diffusers Unified Format)
+      deprecated: false
+      hidden: false
+      experimental: false
+      experimentalcli: false
+      kubernetes: false
+      swarm: false
     - option: dir-tar
       value_type: stringArray
       default_value: '[]'