@bendyline/gezel-inference 1.0.6 → 1.0.7

package/dist/providers/image/sd-cpp.d.ts

@@ -15,10 +15,11 @@
  * catalog manifest (so local state is self-describing even if the
  * catalog entry later disappears).
  */
+import { type ImageLoraPrompt, type ImageLoraPurpose, type ImageLoraRender } from '@bendyline/gezel';
 import { type ModelStorageRoots } from '../../models/storage-roots.js';
 import type { GpuArbiter } from '../gpu-arbiter.js';
 import type { NativeEngineSupervisor } from '../native/supervisor.js';
-import { type ImageEngineHealth, type ImageGenerationInput, type ImageGenerationOutput, type ImageGenerationProgress, type ImageModelPullEvent, type ImageModelPullSpec, type ImageProvider, type InstalledImageModelInfo } from './types.js';
+import { type ImageEngineHealth, type ImageGenerationInput, type ImageGenerationOutput, type ImageGenerationProgress, type ImageLoraPullSpec, type ImageModelPullEvent, type ImageModelPullSpec, type ImageProvider, type InstalledImageLoraInfo, type InstalledImageModelInfo } from './types.js';
 export interface StableDiffusionCppProviderOptions {
     /** Base URL of the running sd-server (no trailing slash). */
     baseUrl: string;
@@ -26,6 +27,13 @@ export interface StableDiffusionCppProviderOptions {
     modelsRoot: string;
     /** Optional read-only machine-store overlay. */
     storageRoots?: ModelStorageRoots;
+    /**
+     * Where LoRAs install (`modelStorageRoots({ engine: 'sd-cpp-loras' })`).
+     * Only the writable root reaches the engine, which takes one
+     * `--lora-model-dir`. Absent when the engine is someone else's server
+     * (`GEZEL_SD_SERVER_URL`): renders then go without LoRAs, with a note.
+     */
+    loraRoots?: ModelStorageRoots;
     /** Awake-time budget for one render request, in ms. Defaults to 30 minutes. */
     timeoutMs?: number;
     /** Injected for tests. Defaults to global fetch. */
@@ -130,6 +138,8 @@ export interface SdServerLoadProfile {
      * measured not to change.
      */
     stepCache?: 'easycache';
+    /** `--model-args` key=value list. See MODEL_ARGS. */
+    modelArgs?: string;
 }
 export declare function sdServerLoadProfile(modelId: string): SdServerLoadProfile;
 export declare class StableDiffusionCppProvider implements ImageProvider {
@@ -137,6 +147,7 @@ export declare class StableDiffusionCppProvider implements ImageProvider {
     private readonly baseUrl;
     private readonly modelsRoot;
     private readonly storageRoots;
+    private readonly loraRoots?;
     private readonly timeoutMs;
     private readonly fetchImpl;
     private readonly supervisor?;
@@ -150,6 +161,28 @@ export declare class StableDiffusionCppProvider implements ImageProvider {
     generate(input: ImageGenerationInput): Promise<ImageGenerationOutput>;
     private generateInner;
     listInstalledModels(): Promise<InstalledImageModelInfo[]>;
+    /** Installed LoRAs, from the root the engine reads; none when this engine takes none. */
+    listLoras(): Promise<InstalledImageLoraInfo[]>;
+    /**
+     * Install one LoRA: a single safetensors file, sha256-checked like a model,
+     * beside a `manifest.json` that keeps the catalog's recipe so the engine can
+     * apply it without the catalog.
+     */
+    pullLora(id: string, spec: ImageLoraPullSpec, signal?: AbortSignal): AsyncIterable<ImageModelPullEvent>;
+    deleteLora(id: string): Promise<void>;
+    /**
+     * The LoRAs a render asked for that this engine can apply to `modelId`:
+     * installed in the root the engine reads and checked on that model. The
+     * rest are dropped with a note, never an error.
+     */
+    private resolveLoras;
+    /**
+     * Keep the LoRAs the running engine can see. sd-server fills its LoRA list
+     * only when it is asked for it, and answers `invalid lora path` for any LoRA
+     * named before then, so the list is read before every LoRA render.
+     */
+    private lorasTheEngineHas;
+    private readInstalledLora;
     pullModel(id: string, spec: ImageModelPullSpec, signal?: AbortSignal): AsyncIterable<ImageModelPullEvent>;
     /**
      * Stream a single URL to disk with sha256 verification, reporting cumulative
@@ -167,11 +200,13 @@ export declare class StableDiffusionCppProvider implements ImageProvider {
     health(): Promise<ImageEngineHealth>;
     shutdown(): Promise<void>;
     /**
-     * The configured default when it is installed, else the first installed
-     * model id, else undefined. Stable ordering matches
-     * {@link listInstalledModels} so the default is deterministic.
+     * The model a fresh picture renders with: the configured default when it
+     * is installed and draws, else the first installed model that draws. Stable
+     * ordering matches {@link listInstalledModels} so the default is
+     * deterministic.
      */
     resolveDefaultModelId(): Promise<string | undefined>;
+    private chooseModel;
     private defaultSampleSteps;
     /**
      * Read the installed `manifest.json` fields that steer per-model
@@ -182,6 +217,30 @@ export declare class StableDiffusionCppProvider implements ImageProvider {
      */
     private readInstalledManifestFields;
 }
+/**
+ * sd-server reads `<sd_cpp_extra_args>{json}</sd_cpp_extra_args>` out of a
+ * prompt and applies the JSON as generation parameters, including LoRA and
+ * ADetailer model paths it never checks. Prompts come from gezels reading
+ * untrusted content and from other apps, so the tag never reaches the engine.
+ * Repeated until stable: removing one tag can join the pieces around it into
+ * another.
+ */
+export declare function withoutEngineDirectives(text: string): string;
+/** An installed LoRA with the recipe its catalog entry gave it. */
+export interface InstalledLora {
+    id: string;
+    name: string;
+    filename: string;
+    purpose: ImageLoraPurpose;
+    strength: number;
+    baseModels: string[];
+    prompt?: ImageLoraPrompt;
+    render?: ImageLoraRender;
+    approxSizeBytes: number;
+    installedAt: string;
+}
+/** Every LoRA installed under `loraRoot`, skipping malformed ones. */
+export declare function readInstalledLoras(loraRoot: string): Promise<InstalledLora[]>;
 /**
  * Parse one line of sd-server's weight-load bar, which it prints inside the
  * first request after a launch because it loads weights lazily: