684 add UI for llm olena

discojs/src/index.ts

@@ -38,6 +38,7 @@ export type {
   ModelCardInfo,
   BatchLogs,
   HellaSwagDataset,
+  GenerationConfig,
 } from "./models/index.js";
 
 export {
@@ -50,6 +51,7 @@ export {
   ONNXModel,
   HELLASWAG_URL,
   evaluate_hellaswag,
+  DefaultGenerationConfig,
 } from "./models/index.js";
 export type { GPTConfig, HellaSwagExample } from "./models/index.js";
 
@@ -69,7 +71,7 @@ export {
 export type { TaskProvider } from "./task/index.js";
 
 export type { DataType, Network, DataFormat } from "./types/index.js";
-export { dataTypeValues } from "./types/index.js";
+export { dataTypeValues, isDataType } from "./types/index.js";
 
 export { extractColumn } from "./processing/index.js";
 

discojs/src/models/generation.ts

@@ -0,0 +1,19 @@
+export interface GenerationConfig {
+  // take random token weighted by its probability
+  // If false, predict the token with the highest probability.
+  doSample: boolean;
+  // the generation temperature (higher means more randomness).
+  // Set to 0 for greedy decoding.
+  temperature: number;
+  // only consider the topk most likely tokens for sampling.
+  // used if doSample is true.
+  topk: number;
+  // optional random seed for sampling.
+  seed?: number;
+}
+
+export const DefaultGenerationConfig: GenerationConfig = {
+  temperature: 1.0,
+  doSample: true,
+  topk: 50,
+};

discojs/src/models/implementations/gpt/config.ts

@@ -46,7 +46,7 @@ export const DefaultGPTConfig: Required<GPTConfig> = {
   nLayer: 3,
   nHead: 3,
   nEmbd: 48,
-  seed: Math.random(),
+  seed: Math.floor(Math.random() * Number.MAX_SAFE_INTEGER),
 };
 
 export type ModelSize = {
@@ -73,24 +73,3 @@ export function getModelSizes(modelType: GPTModelType): Required<ModelSize> {
       return { nLayer: 3, nHead: 3, nEmbd: 48 };
   }
 }
-
-export interface GenerationConfig {
-  // take random token weighted by its probability
-  // If false, predict the token with the highest probability.
-  doSample: boolean;
-  // the generation temperature (higher means more randomness).
-  // Set to 0 for greedy decoding.
-  temperature: number;
-  // only consider the topk most likely tokens for sampling.
-  // used if doSample is true.
-  topk: number;
-  // random seed for sampling.
-  seed: number;
-}
-
-export const DefaultGenerationConfig: Required<GenerationConfig> = {
-  temperature: 1.0,
-  doSample: false,
-  seed: Math.random(),
-  topk: 50,
-};

discojs/src/models/implementations/gpt/gpt.spec.ts

@@ -38,7 +38,7 @@ describe("gpt-tfjs", () => {
     const inputTokens = tokenizer.tokenize(data);
 
     const outputToken = (
-      await model.predict(List.of(inputTokens), { seed })
+      await model.predict(List.of(inputTokens), { seed, doSample: false })
     ).first();
     if (outputToken === undefined) throw new Error("empty prediction");
     const output = tokenizer.decode([outputToken]);

discojs/src/models/implementations/gpt/gpt.ts

@@ -15,14 +15,10 @@ import { EpochLogs } from "#models/logs";
 import { Model } from "#models/model";
 import { GPTModel } from "#models/implementations/gpt/model";
 import evaluate from "#models/implementations/gpt/evaluate";
-import {
-  DefaultGPTConfig,
-  DefaultGenerationConfig,
-} from "#models/implementations/gpt/config";
-import type {
-  GPTConfig,
-  GenerationConfig,
-} from "#models/implementations/gpt/config";
+import { DefaultGPTConfig } from "#models/implementations/gpt/config";
+import type { GPTConfig } from "#models/implementations/gpt/config";
+import { DefaultGenerationConfig } from "#models/generation";
+import type { GenerationConfig } from "#models/generation";
 
 const debug = createDebug("discojs:models:gpt");
 
@@ -199,7 +195,9 @@ export class GPT extends Model<"text"> {
       logits
         .slice([logits.shape[0] - 1])
         .squeeze<tf.Tensor1D>([0])
-        .div<tf.Tensor1D>(config.temperature)
+        .div<tf.Tensor1D>(
+          config.doSample && config.temperature > 0 ? config.temperature : 1,
+        )
         .softmax(),
     );
     logits.dispose();
@@ -263,7 +261,7 @@ export class GPT extends Model<"text"> {
     return this.model;
   }
 
-  [Symbol.dispose](): void {
+  dispose(): void {
     if (this.model.optimizer !== undefined) {
       this.model.optimizer.dispose();
     }

discojs/src/models/index.ts

@@ -2,6 +2,8 @@ export { Model } from "./model.js";
 export type { BatchLogs, ValidationMetrics } from "./logs.js";
 export { EpochLogs } from "./logs.js";
 export { Tokenizer } from "./tokenizer.js";
+export { DefaultGenerationConfig } from "./generation.js";
+export type { GenerationConfig } from "./generation.js";
 
 export type {
   GPTConfig,

discojs/src/models/model.ts

@@ -54,5 +54,9 @@ export abstract class Model<D extends DataType> implements Disposable {
    * }
    * Calling f() will call the model's dispose method when exiting the function.
    */
-  abstract [Symbol.dispose](): void;
+  [Symbol.dispose](): void {
+    this.dispose();
+  }
+
+  abstract dispose(): void;
 }

discojs/src/models/onnx.spec.ts

@@ -3,7 +3,7 @@ import { describe, expect, it } from "vitest";
 import { List } from "immutable";
 import { AutoTokenizer } from "@xenova/transformers";
 import { ONNXModel } from "#models/onnx";
-import { DefaultGenerationConfig } from "#models/implementations/gpt/config";
+import { DefaultGenerationConfig } from "#models/generation";
 
 describe("ONNXModel.predict", { timeout: 50_000 }, () => {
   it("should generate the next token ID from a prompt", async () => {

discojs/src/models/onnx.ts

@@ -5,9 +5,9 @@ import { List } from "immutable";
 import type { WeightsContainer } from "#weights/index";
 import type { Batched } from "#dataset/index";
 import type { DataFormat } from "#types/index";
-import type { GenerationConfig as TFJSGenerationConfig } from "#models/implementations/gpt/config";
 import { Model } from "#models/model";
-import { DefaultGenerationConfig } from "#models/implementations/gpt/config";
+import { DefaultGenerationConfig } from "#models/generation";
+import type { GenerationConfig } from "#models/generation";
 
 export class ONNXModel extends Model<"text"> {
   readonly datatype = "text" as const;
@@ -30,7 +30,7 @@ export class ONNXModel extends Model<"text"> {
 
   override async predict(
     batch: Batched<DataFormat.ModelEncoded["text"][0]>,
-    options?: Partial<TFJSGenerationConfig>,
+    options?: Partial<GenerationConfig>,
   ): Promise<Batched<DataFormat.ModelEncoded["text"][1]>> {
     const config = Object.assign({}, DefaultGenerationConfig, options);
 
@@ -43,7 +43,7 @@ export class ONNXModel extends Model<"text"> {
 
   async #predictSingle(
     tokens: DataFormat.ModelEncoded["text"][0],
-    config: TFJSGenerationConfig,
+    config: GenerationConfig,
   ): Promise<DataFormat.ModelEncoded["text"][1]> {
     const contextLength =
       (this.model.config as { max_position_embeddings?: number })
@@ -125,7 +125,7 @@ export class ONNXModel extends Model<"text"> {
     throw new Error("Weights setting not supported in ONNX models");
   }
 
-  [Symbol.dispose](): void {
+  dispose() {
     // Dispose of the model to free up memory
     void this.model.dispose();
   }

discojs/src/models/tfjs.ts

@@ -209,7 +209,7 @@ export class TFJS<D extends "image" | "tabular"> extends Model<D> {
     return [this.datatype, await ret];
   }
 
-  [Symbol.dispose](): void {
+  dispose(): void {
     this.model.dispose();
   }
 

discojs/src/types/datatype.ts

@@ -1,2 +1,9 @@
 export const dataTypeValues = ["image", "tabular", "text"] as const;
+
 export type DataType = (typeof dataTypeValues)[number];
+
+export function isDataType(x: unknown): x is DataType {
+  return (
+    typeof x == "string" && (dataTypeValues as readonly string[]).includes(x)
+  );
+}

discojs/src/types/index.ts

@@ -1,7 +1,7 @@
 // eslint-disable-next-line no-restricted-syntax -- namespace re-export acceptable here
 export * as DataFormat from "./data_format.js";
 
-export { dataTypeValues } from "./datatype.js";
+export { dataTypeValues, isDataType } from "./datatype.js";
 export type { DataType } from "./datatype.js";
 
 export type { Network } from "./network.js";

package.json

@@ -4,7 +4,7 @@
   "scripts": {
     "lint": "eslint",
     "format:check": "prettier -c .",
-    "format:fix": "prettier -w .",
+    "format:fix": "prettier -w --list-different .",
     "check_cycles": "dpdm --tsconfig discojs/tsconfig.lib.json --circular --exit-code circular:1 'discojs/src/**/*.ts'"
   },
   "dependencies": {

server/package.json

@@ -10,7 +10,7 @@
   },
   "scripts": {
     "watch": "nodemon --ext ts --ignore dist --watch ../discojs-node/dist --watch . --exec pnpm run",
-    "start": "node dist/main.js",
+    "start": "pnpm run build && node dist/main.js",
     "build": "tsc --build",
     "test": "cd .. && vitest --run --project=server"
   },

webapp/.env.development

@@ -1 +1,4 @@
-VITE_SERVER_URL=http://localhost:8080
+# When set to localhost, the server host is derived from whichever
+# host serves the webapp (localhost, or your LAN IP when running `vite --host`).
+VITE_SERVER_URL=localhost
+VITE_SERVER_PORT=8080