Run Flux LoRAs on Replicate (#1395)

This PR updates the `ReplicateTextToImageTask` method to support running
arbitrary LoRAs on Replicate.

How it works:

1. Detect LoRAs by their tags before inference.
2. Use https://replicate.com/black-forest-labs/flux-dev-lora for
inference
3. Use the requested LoRA as the `weights_url` input to the
flux-dev-lora model

## Companion PR

- replicate/huggingface-model-mappings#9


## Testing

Remove the `skip` from `describe.skip("InferenceClient" ...` in
InferenceClient.spec.ts

Then: 
```
$ pnpx vitest run --config vitest.config.mts -t "Replicate"
```

---------

Co-authored-by: Simon Brandeis <[email protected]>
Co-authored-by: célina <[email protected]>
Co-authored-by: Lucain <[email protected]>

Author: 4 people

packages/inference/src/providers/replicate.ts

@@ -19,7 +19,6 @@ import { isUrl } from "../lib/isUrl";
 import type { BodyParams, HeaderParams, UrlParams } from "../types";
 import { omit } from "../utils/omit";
 import { TaskProviderHelper, type TextToImageTaskHelper, type TextToVideoTaskHelper } from "./providerHelper";
-
 export interface ReplicateOutput {
 	output?: string | string[];
 }
@@ -63,6 +62,21 @@ abstract class ReplicateTask extends TaskProviderHelper {
 }
 
 export class ReplicateTextToImageTask extends ReplicateTask implements TextToImageTaskHelper {
+	override preparePayload(params: BodyParams): Record<string, unknown> {
+		return {
+			input: {
+				...omit(params.args, ["inputs", "parameters"]),
+				...(params.args.parameters as Record<string, unknown>),
+				prompt: params.args.inputs,
+				lora_weights:
+					params.mapping?.adapter === "lora" && params.mapping.adapterWeightsPath
+						? `https://huggingface.co/${params.mapping.hfModelId}`
+						: undefined,
+			},
+			version: params.model.includes(":") ? params.model.split(":")[1] : undefined,
+		};
+	}
+
 	override async getResponse(
 		res: ReplicateOutput | Blob,
 		url?: string,

packages/inference/test/InferenceClient.spec.ts

@@ -1160,6 +1160,17 @@ describe.skip("InferenceClient", () => {
 				expect(res).toBeInstanceOf(Blob);
 			});
 
+			// Runs black-forest-labs/flux-dev-lora under the hood
+			// with fofr/flux-80s-cyberpunk as the LoRA weights
+			it("textToImage - all Flux LoRAs", async () => {
+				const res = await client.textToImage({
+					model: "fofr/flux-80s-cyberpunk",
+					provider: "replicate",
+					inputs: "style of 80s cyberpunk, a portrait photo",
+				});
+				expect(res).toBeInstanceOf(Blob);
+			});
+
 			it("textToImage canonical - stabilityai/stable-diffusion-3.5-large-turbo", async () => {
 				const res = await client.textToImage({
 					model: "stabilityai/stable-diffusion-3.5-large-turbo",