feat: mutiple backends for recipes

PACKAGING-GUIDE.md

@@ -33,56 +33,61 @@ The catalog also lists the models that may be associated to recipes. A model is
 citizen in AI Lab as they will be listed in the Models page and can be tested through the playground.
 
 A model has the following attributes:
-- ```id```: a unique identifier for the model
-- ```name```: the model name
-- ```description```: a detailed description about the model
-- ```registry```: the model registry where the model is stored
-- ```popularity```: an integer field giving the rating of the model. Can be thought as the number of stars
-- ```license```: the license under which the model is available
-- ```url```: the URL used to download the model
-- ```memory```: the memory footprint of the model in bytes, as computed by the workflow `.github/workflows/compute-model-sizes.yaml`
-- ```sha256```: the SHA-256 checksum to be used to verify the downloaded model is identical to the original. It is optional and it must be HEX encoded
+
+- `id`: a unique identifier for the model
+- `name`: the model name
+- `description`: a detailed description about the model
+- `registry`: the model registry where the model is stored
+- `popularity`: an integer field giving the rating of the model. Can be thought as the number of stars
+- `license`: the license under which the model is available
+- `url`: the URL used to download the model
+- `memory`: the memory footprint of the model in bytes, as computed by the workflow `.github/workflows/compute-model-sizes.yaml`
+- `sha256`: the SHA-256 checksum to be used to verify the downloaded model is identical to the original. It is optional and it must be HEX encoded
 
 #### Recipes
 
 A recipe is a sample AI application that is packaged as one or several containers. It is built by AI Lab when the user chooses to download and run it on their workstation. It is provided as
 source code and AI Lab will make sure the container images are built prior to launching the containers.
 
 A recipe has the following attributes:
-- ```id```: a unique identifier to the recipe
-- ```name```: the recipe name
-- ```description```: a detailed description about the recipe
-- ```repository```: the URL where the recipe code can be retrieved
-- ```ref```: an optional ref in the repository to checkout (a branch name, tag name, or commit full id - short commit id won't be recognized). If not defined, the default branch will be used
-- ```categories```: an array of category id to be associated by this recipe
-- ```basedir```: an optional path within the repository where the ai-lab.yaml file is located. If not provided, the ai-lab.yaml is assumed to be located at the root the repository
-- ```readme```: a markdown description of the recipe
-- ```models```: an array of model id to be associated with this recipe
+
+- `id`: a unique identifier to the recipe
+- `name`: the recipe name
+- `description`: a detailed description about the recipe
+- `repository`: the URL where the recipe code can be retrieved
+- `ref`: an optional ref in the repository to checkout (a branch name, tag name, or commit full id - short commit id won't be recognized). If not defined, the default branch will be used
+- `categories`: an array of category id to be associated by this recipe
+- `basedir`: an optional path within the repository where the ai-lab.yaml file is located. If not provided, the ai-lab.yaml is assumed to be located at the root the repository
+- `readme`: a markdown description of the recipe
+- `models`: an array of model id to be associated with this recipe
+- `backends`: an array of backends from which models may be associated with this recipe. The backends are used to filter models when associating them to a recipe.
 
 #### Recipe configuration file
 
-The configuration file is called ```ai-lab.yaml``` and follows the following syntax.
+The configuration file is called `ai-lab.yaml` and follows the following syntax.
 
-The root elements are called ```version``` and ```application```.
+The root elements are called `version` and `application`.
 
-```version``` represents the version of the specifications that ai-lab adheres to (so far, the only accepted value here is `v1.0`).
+`version` represents the version of the specifications that ai-lab adheres to (so far, the only accepted value here is `v1.0`).
 
-```application``` contains an attribute called ```containers``` whose syntax is an array of objects containing the following attributes:
-- ```name```: the name of the container
-- ```contextdir```: the context directory used to build the container.
-- ```containerfile```: the containerfile used to build the image
-- ```model-service```: a boolean flag used to indicate if the container is running the model or not
-- ```arch```: an optional array of architecture for which this image is compatible with. The values follow the
-[GOARCH specification](https://go.dev/src/go/build/syslist.go)
-- ```gpu-env```: an optional array of GPU environment for which this image is compatible with. The only accepted value here is cuda.
-- ```ports```: an optional array of ports for which the application listens to.
+`application` contains an attribute called `containers` whose syntax is an array of objects containing the following attributes:
+
+- `name`: the name of the container
+- `contextdir`: the context directory used to build the container.
+- `containerfile`: the containerfile used to build the image
+- `model-service`: a boolean flag used to indicate if the container is running the model or not
+- `arch`: an optional array of architecture for which this image is compatible with. The values follow the
+  [GOARCH specification](https://go.dev/src/go/build/syslist.go)
+- `gpu-env`: an optional array of GPU environment for which this image is compatible with. The only accepted value here is cuda.
+- `ports`: an optional array of ports for which the application listens to.
 - `image`: an optional image name to be used when building the container image.
 
-The container that is running the service (having the ```model-service``` flag equal to ```true```) can use at runtime
-the model managed by AI Lab through an environment variable ```MODEL_PATH``` whose value is the full path name of the
+The container that is running the service (having the `model-service` flag equal to `true`) can use at runtime
+the model managed by AI Lab through an environment variable `MODEL_PATH` whose value is the full path name of the
 model file.
 
 Below is given an example of such a configuration file:
+
 ```yaml
 application:
   containers:

packages/backend/src/managers/catalogManager.ts

@@ -289,9 +289,14 @@ export class CatalogManager extends Publisher<ApplicationCatalog> implements Dis
           result = res;
           break;
         }
-        case 'tools':
-          result = result.filter(r => values.includes(r.backend ?? ''));
+        case 'tools': {
+          let res: Recipe[] = [];
+          for (const value of values) {
+            res = [...res, ...result.filter(r => r.backends?.includes(value))];
+          }
+          result = res;
           break;
+        }
         case 'frameworks': {
           let res: Recipe[] = [];
           for (const value of values) {
@@ -325,13 +330,13 @@ export class CatalogManager extends Publisher<ApplicationCatalog> implements Dis
       choices.tools = this.filterRecipes(subfilters).choices.tools;
     } else {
       choices.tools = result
-        .map(r => r.backend)
+        .flatMap(r => r.backends)
         .filter(b => b !== undefined)
         .filter((value, index, array) => array.indexOf(value) === index)
         .sort((a, b) => a.localeCompare(b))
         .map(t => ({
           name: t,
-          count: result.filter(r => r.backend === t).length,
+          count: result.filter(r => r.backends?.includes(t)).length,
         }));
     }
 

packages/backend/src/managers/recipes/RecipeManager.ts

@@ -106,7 +106,7 @@ export class RecipeManager implements Disposable {
 
     let inferenceServer: InferenceServer | undefined;
     // if the recipe has a defined backend, we gives priority to using an inference server
-    if (recipe.backend && recipe.backend === model.backend) {
+    if (model.backend && recipe.backends?.includes(model.backend)) {
       let task: Task | undefined;
       try {
         inferenceServer = this.inferenceManager.findServerByModel(model);

packages/backend/src/tests/ai-test.json

@@ -27,7 +27,7 @@
       "description": "",
       "repository": "",
       "readme": "",
-      "backend": "tool1",
+      "backends": ["tool1"],
       "languages": ["lang1", "lang10"],
       "frameworks": ["fw1", "fw10"]
     },
@@ -38,7 +38,7 @@
       "description": "",
       "repository": "",
       "readme": "",
-      "backend": "tool2",
+      "backends": ["tool2"],
       "languages": ["lang2", "lang10"],
       "frameworks": ["fw2", "fw10"]
     },
@@ -49,7 +49,7 @@
       "description": "",
       "repository": "",
       "readme": "",
-      "backend": "tool3",
+      "backends": ["tool3"],
       "languages": ["lang3", "lang11"],
       "frameworks": ["fw2", "fw10", "fw11"]
     }

packages/backend/src/utils/catalogUtils.spec.ts

@@ -173,6 +173,98 @@ describe('sanitize', () => {
     expect(catalog.recipes[0].languages).toStrictEqual(['lang1']);
     expect(catalog.recipes[0].frameworks).toStrictEqual(['fw1']);
   });
+
+  test('should return backend recipe as string', () => {
+    const raw = {
+      version: '1.0',
+      recipes: [
+        {
+          id: 'chatbot',
+          description: 'This is a Streamlit chat demo application.',
+          name: 'ChatBot',
+          repository: 'https://github.com/containers/ai-lab-recipes',
+          ref: 'v1.1.3',
+          icon: 'natural-language-processing',
+          categories: ['natural-language-processing'],
+          basedir: 'recipes/natural_language_processing/chatbot',
+          readme: '',
+          recommended: ['hf.instructlab.granite-7b-lab-GGUF', 'hf.instructlab.merlinite-7b-lab-GGUF'],
+          backends: ['llama-cpp'],
+          languages: ['lang1'],
+          frameworks: ['fw1'],
+        },
+      ],
+      models: [
+        {
+          id: 'Mistral-7B-Instruct-v0.3-Q4_K_M.gguf',
+          name: 'Mistral-7B-Instruct-v0.3-Q4_K_M',
+          description: 'Model imported from path\\Mistral-7B-Instruct-v0.3-Q4_K_M.gguf',
+          hw: 'CPU',
+          file: {
+            path: 'path',
+            file: 'Mistral-7B-Instruct-v0.3-Q4_K_M.gguf',
+            size: 4372812000,
+            creation: '2024-06-19T12:14:12.489Z',
+          },
+          memory: 4372812000,
+        },
+      ],
+    };
+    expect(hasCatalogWrongFormat(raw)).toBeFalsy();
+    const catalog = sanitize(raw);
+    expect(catalog.version).equals(CatalogFormat.CURRENT);
+    expect(catalog.models[0].backend).toBeUndefined();
+    expect(catalog.models[0].name).equals('Mistral-7B-Instruct-v0.3-Q4_K_M');
+    expect(catalog.recipes[0].languages).toStrictEqual(['lang1']);
+    expect(catalog.recipes[0].frameworks).toStrictEqual(['fw1']);
+    expect(catalog.recipes[0].backends).toStrictEqual(['llama-cpp']);
+  });
+
+  test('should return multiple backend recipe as array', () => {
+    const raw = {
+      version: '1.0',
+      recipes: [
+        {
+          id: 'chatbot',
+          description: 'This is a Streamlit chat demo application.',
+          name: 'ChatBot',
+          repository: 'https://github.com/containers/ai-lab-recipes',
+          ref: 'v1.1.3',
+          icon: 'natural-language-processing',
+          categories: ['natural-language-processing'],
+          basedir: 'recipes/natural_language_processing/chatbot',
+          readme: '',
+          recommended: ['hf.instructlab.granite-7b-lab-GGUF', 'hf.instructlab.merlinite-7b-lab-GGUF'],
+          backends: ['llama-cpp', 'openvino'],
+          languages: ['lang1'],
+          frameworks: ['fw1'],
+        },
+      ],
+      models: [
+        {
+          id: 'Mistral-7B-Instruct-v0.3-Q4_K_M.gguf',
+          name: 'Mistral-7B-Instruct-v0.3-Q4_K_M',
+          description: 'Model imported from path\\Mistral-7B-Instruct-v0.3-Q4_K_M.gguf',
+          hw: 'CPU',
+          file: {
+            path: 'path',
+            file: 'Mistral-7B-Instruct-v0.3-Q4_K_M.gguf',
+            size: 4372812000,
+            creation: '2024-06-19T12:14:12.489Z',
+          },
+          memory: 4372812000,
+        },
+      ],
+    };
+    expect(hasCatalogWrongFormat(raw)).toBeFalsy();
+    const catalog = sanitize(raw);
+    expect(catalog.version).equals(CatalogFormat.CURRENT);
+    expect(catalog.models[0].backend).toBeUndefined();
+    expect(catalog.models[0].name).equals('Mistral-7B-Instruct-v0.3-Q4_K_M');
+    expect(catalog.recipes[0].languages).toStrictEqual(['lang1']);
+    expect(catalog.recipes[0].frameworks).toStrictEqual(['fw1']);
+    expect(catalog.recipes[0].backends).toStrictEqual(['llama-cpp', 'openvino']);
+  });
 });
 
 describe('merge', () => {

packages/backend/src/utils/catalogUtils.ts

@@ -114,6 +114,15 @@ export function isStringArray(obj: unknown): obj is Array<string> {
   return Array.isArray(obj) && obj.every(item => typeof item === 'string');
 }
 
+function sanitizeBackends(recipe: object): string[] | undefined {
+  if ('backend' in recipe && typeof recipe.backend === 'string') {
+    return [recipe.backend];
+  } else if ('backends' in recipe && Array.isArray(recipe.backends)) {
+    return recipe.backends;
+  }
+  return undefined;
+}
+
 export function sanitizeRecipe(recipe: unknown): Recipe {
   if (
     isNonNullObject(recipe) &&
@@ -143,7 +152,7 @@ export function sanitizeRecipe(recipe: unknown): Recipe {
       icon: 'icon' in recipe && typeof recipe.icon === 'string' ? recipe.icon : undefined,
       basedir: 'basedir' in recipe && typeof recipe.basedir === 'string' ? recipe.basedir : undefined,
       recommended: 'recommended' in recipe && isStringArray(recipe.recommended) ? recipe.recommended : undefined,
-      backend: 'backend' in recipe && typeof recipe.backend === 'string' ? recipe.backend : undefined,
+      backends: sanitizeBackends(recipe),
       languages: 'languages' in recipe && isStringArray(recipe.languages) ? recipe.languages : undefined,
       frameworks: 'frameworks' in recipe && isStringArray(recipe.frameworks) ? recipe.frameworks : undefined,
     };

packages/frontend/src/lib/RecipeCardTags.spec.ts

@@ -31,7 +31,7 @@ const recipe = {
   categories: ['natural-language-processing', 'audio'],
   languages: ['java', 'python'],
   frameworks: ['langchain', 'vectordb'],
-  backend: 'whisper-cpp',
+  backends: ['whisper-cpp'],
 };
 
 class ResizeObserver {

packages/frontend/src/lib/RecipeCardTags.svelte

@@ -12,7 +12,7 @@ let { recipe }: Props = $props();
 
 const TAGS: string[] = [
   ...recipe.categories,
-  ...(recipe.backend !== undefined ? [recipe.backend] : []),
+  ...(recipe.backends ?? []),
   ...(recipe.frameworks ?? []),
   ...(recipe.languages ?? []),
 ];

packages/frontend/src/pages/StartRecipe.spec.ts

@@ -62,7 +62,7 @@ vi.mock('../utils/client', async () => ({
 
 const fakeRecipe: Recipe = {
   id: 'dummy-recipe-id',
-  backend: InferenceType.LLAMA_CPP,
+  backends: [InferenceType.LLAMA_CPP],
   name: 'Dummy Recipe',
   description: 'Dummy description',
   recommended: ['dummy-model-1'],

packages/frontend/src/pages/StartRecipe.svelte

@@ -42,7 +42,7 @@ let startedContainerProviderConnectionInfo: ContainerProviderConnectionInfo[] =
 let localPath: LocalRepository | undefined = $derived(findLocalRepositoryByRecipeId($localRepositories, recipe?.id));
 // Filter all models based on backend property
 let models: ModelInfo[] = $derived(
-  $modelsInfo.filter(model => (model.backend ?? InferenceType.NONE) === (recipe?.backend ?? InferenceType.NONE)),
+  $modelsInfo.filter(model => recipe?.backends?.includes(model.backend ?? InferenceType.NONE)),
 );
 // Hold the selected model
 let model: ModelInfo | undefined = $state(undefined);

packages/shared/src/models/IRecipe.ts

@@ -53,10 +53,10 @@ export interface Recipe {
   basedir?: string;
   recommended?: string[];
   /**
-   * The backend field aims to target which inference
+   * The backends field aims to target which inference
    * server the recipe requires
    */
-  backend?: string;
+  backends?: string[];
   languages?: string[];
   frameworks?: string[];
 }