@supersoniks/concorde 4.9.95 → 4.9.97-visual-stack.0

package/src/core/components/functional/infer/hugging-face/hf-schema.ts

@@ -0,0 +1,390 @@
+/**
+ * Schémas des options d'appel, par tâche.
+ *
+ * Ces schémas servent à valider à l'exécution ce qu'un descripteur SDUI (ou
+ * un attribut HTML) passe dans `options.call`, et à documenter le catalogue
+ * MCP (`hfCallOptionsJsonSchema`).
+ *
+ * Ils ne peuvent pas diverger des types de Transformers.js :
+ * - chaque tâche doit lister TOUTES les options sérialisables de son pipeline
+ *   (clé manquante = erreur de compilation) ;
+ * - une clé qui n'existe pas dans le pipeline est refusée ;
+ * - tout ce qu'un schéma accepte doit être une valeur valide du type de
+ *   l'option (sinon erreur de compilation « schema-trop-large »).
+ *
+ * Les bornes (`maximum`, `maxItems`…) protègent le poste : un descripteur
+ * SDUI ne peut pas demander 10 000 jetons ou 200 faisceaux.
+ */
+import type { HfCallOptions, HfTask } from "./types";
+
+/* ------------------------------------------------------------------ */
+/* Langage de schéma (sous-ensemble JSON Schema)                        */
+/* ------------------------------------------------------------------ */
+
+export type HfValueSchema =
+  | { readonly type: "boolean" }
+  | {
+      readonly type: "number";
+      readonly integer?: boolean;
+      readonly minimum?: number;
+      readonly maximum?: number;
+    }
+  | { readonly type: "string"; readonly maxLength?: number }
+  | { readonly type: "null" }
+  | { readonly enum: readonly (string | number | boolean)[] }
+  | {
+      readonly type: "array";
+      readonly items: HfValueSchema;
+      readonly maxItems?: number;
+    }
+  | {
+      readonly type: "tuple";
+      readonly prefixItems: readonly HfValueSchema[];
+    }
+  | { readonly anyOf: readonly HfValueSchema[] };
+
+/** Valeurs acceptées par un schéma (niveau types). */
+export type HfSchemaValue<S> = S extends { type: "boolean" }
+  ? boolean
+  : S extends { type: "number" }
+    ? number
+    : S extends { type: "string" }
+      ? string
+      : S extends { type: "null" }
+        ? null
+        : S extends { enum: readonly (infer E)[] }
+          ? E
+          : S extends { type: "array"; items: infer I }
+            ? HfSchemaValue<I>[]
+            : S extends { type: "tuple"; prefixItems: infer P }
+              ? { -readonly [K in keyof P]: HfSchemaValue<P[K]> }
+              : S extends { anyOf: readonly (infer A)[] }
+                ? HfSchemaValue<A>
+                : never;
+
+type SchemaTable = {
+  readonly [T in HfTask]: {
+    readonly [K in keyof HfCallOptions<T>]-?: HfValueSchema;
+  };
+};
+
+/**
+ * Vérifie que chaque schéma est « sûr » : ce qu'il accepte est assignable au
+ * type de l'option dans Transformers.js.
+ */
+type Sound<S extends SchemaTable> = {
+  readonly [T in HfTask]: {
+    readonly [K in keyof HfCallOptions<T>]-?: K extends keyof S[T]
+      ? HfSchemaValue<S[T][K]> extends Exclude<HfCallOptions<T>[K], undefined>
+        ? S[T][K]
+        : { "schema-trop-large": K; accepte: HfSchemaValue<S[T][K]> }
+      : never;
+  };
+};
+
+function defineCallSchemas<const S extends SchemaTable>(
+  schemas: S & Sound<S>,
+): S {
+  return schemas;
+}
+
+/* ------------------------------------------------------------------ */
+/* Briques partagées                                                   */
+/* ------------------------------------------------------------------ */
+
+const bool = { type: "boolean" } as const;
+const num = (minimum?: number, maximum?: number) =>
+  ({ type: "number", minimum, maximum }) as const;
+const int = (minimum?: number, maximum?: number) =>
+  ({ type: "number", integer: true, minimum, maximum }) as const;
+const str = (maxLength = 500) => ({ type: "string", maxLength }) as const;
+const intList = (maxItems = 256) =>
+  ({ type: "array", items: int(0), maxItems }) as const;
+const intMatrix = (maxItems = 64) =>
+  ({ type: "array", items: intList(), maxItems }) as const;
+const labelList = { type: "array", items: str(200), maxItems: 50 } as const;
+const labels = { anyOf: [str(200), labelList] } as const;
+const topK = { anyOf: [int(1, 50), { type: "null" }] } as const;
+
+/**
+ * Paramètres de génération (`GenerationConfig`) sérialisables, communs à
+ * `translation`, `image-to-text` et `automatic-speech-recognition`. Bornés
+ * pour rester raisonnables dans un navigateur.
+ */
+const generation = {
+  max_length: int(1, 1024),
+  max_new_tokens: int(1, 1024),
+  min_length: int(0, 1024),
+  min_new_tokens: int(0, 1024),
+  early_stopping: { anyOf: [bool, { enum: ["never"] }] },
+  max_time: num(0, 120),
+  do_sample: bool,
+  num_beams: int(1, 8),
+  num_beam_groups: int(1, 8),
+  penalty_alpha: num(0, 1),
+  use_cache: bool,
+  temperature: num(0, 5),
+  top_k: int(0, 200),
+  top_p: num(0, 1),
+  typical_p: num(0, 1),
+  epsilon_cutoff: num(0, 1),
+  eta_cutoff: num(0, 1),
+  diversity_penalty: num(0, 10),
+  repetition_penalty: num(0, 10),
+  encoder_repetition_penalty: num(0, 10),
+  length_penalty: num(-10, 10),
+  no_repeat_ngram_size: int(0, 32),
+  bad_words_ids: intMatrix(),
+  force_words_ids: {
+    anyOf: [
+      intMatrix(),
+      { type: "array", items: intMatrix(), maxItems: 64 },
+    ],
+  },
+  renormalize_logits: bool,
+  forced_bos_token_id: int(0),
+  forced_eos_token_id: { anyOf: [int(0), intList()] },
+  remove_invalid_values: bool,
+  exponential_decay_length_penalty: {
+    type: "tuple",
+    prefixItems: [int(0), num()],
+  },
+  suppress_tokens: intList(),
+  begin_suppress_tokens: intList(),
+  forced_decoder_ids: {
+    type: "array",
+    items: { type: "tuple", prefixItems: [int(0), int(0)] },
+    maxItems: 64,
+  },
+  guidance_scale: num(0, 50),
+  num_return_sequences: int(1, 8),
+  output_attentions: bool,
+  output_hidden_states: bool,
+  output_scores: bool,
+  return_dict_in_generate: bool,
+  pad_token_id: int(0),
+  bos_token_id: int(0),
+  eos_token_id: { anyOf: [int(0), intList()] },
+  encoder_no_repeat_ngram_size: int(0, 32),
+  decoder_start_token_id: int(0),
+} as const;
+
+/* ------------------------------------------------------------------ */
+/* Table par tâche                                                     */
+/* ------------------------------------------------------------------ */
+
+export const HF_CALL_OPTION_SCHEMAS = defineCallSchemas({
+  "feature-extraction": {
+    pooling: {
+      enum: ["none", "mean", "cls", "first_token", "eos", "last_token"],
+    },
+    normalize: bool,
+    quantize: bool,
+    precision: { enum: ["binary", "ubinary"] },
+  },
+  "text-classification": { top_k: topK },
+  "zero-shot-classification": {
+    candidate_labels: labels,
+    hypothesis_template: str(300),
+    multi_label: bool,
+  },
+  translation: generation,
+  "token-classification": {
+    ignore_labels: { type: "array", items: str(100), maxItems: 50 },
+    aggregation_strategy: { enum: ["none", "simple"] },
+  },
+  "fill-mask": { top_k: int(1, 50) },
+  "question-answering": {
+    context: { anyOf: [str(20_000), { type: "array", items: str(20_000), maxItems: 20 }] },
+    top_k: int(1, 20),
+  },
+  "image-classification": { top_k: topK },
+  "zero-shot-image-classification": {
+    candidate_labels: labelList,
+    hypothesis_template: str(300),
+  },
+  "object-detection": { threshold: num(0, 1), percentage: bool },
+  "zero-shot-object-detection": {
+    candidate_labels: labelList,
+    threshold: num(0, 1),
+    top_k: int(1, 100),
+    percentage: bool,
+  },
+  "image-feature-extraction": { pool: bool },
+  "image-segmentation": {
+    threshold: num(0, 1),
+    mask_threshold: num(0, 1),
+    overlap_mask_area_threshold: num(0, 1),
+    subtask: { anyOf: [str(50), { type: "null" }] },
+    label_ids_to_fuse: intList(),
+    target_sizes: intMatrix(),
+  },
+  "depth-estimation": {},
+  "background-removal": {},
+  "image-to-text": generation,
+  "audio-classification": { top_k: int(1, 50) },
+  "automatic-speech-recognition": {
+    ...generation,
+    return_timestamps: { anyOf: [bool, { enum: ["word"] }] },
+    chunk_length_s: num(0, 60),
+    stride_length_s: num(0, 30),
+    force_full_sequences: bool,
+    language: str(40),
+    task: { enum: ["transcribe", "translate"] },
+    num_frames: int(0),
+  },
+});
+
+/* ------------------------------------------------------------------ */
+/* Validation à l'exécution                                            */
+/* ------------------------------------------------------------------ */
+
+export class HfOptionsError extends Error {
+  constructor(message: string) {
+    super(message);
+    this.name = "HfOptionsError";
+  }
+}
+
+function describe(value: unknown): string {
+  if (value === null) return "null";
+  if (Array.isArray(value)) return "array";
+  return typeof value;
+}
+
+function check(
+  schema: HfValueSchema,
+  value: unknown,
+  path: string,
+): string | null {
+  if ("anyOf" in schema) {
+    const errors = schema.anyOf.map((s) => check(s, value, path));
+    return errors.some((e) => e === null) ? null : errors[0];
+  }
+  if ("enum" in schema) {
+    return schema.enum.includes(value as never)
+      ? null
+      : `${path} : valeur ${JSON.stringify(value)} non autorisée (attendu : ${schema.enum
+          .map((v) => JSON.stringify(v))
+          .join(" | ")}).`;
+  }
+  switch (schema.type) {
+    case "boolean":
+      return typeof value === "boolean"
+        ? null
+        : `${path} : booléen attendu, reçu ${describe(value)}.`;
+    case "null":
+      return value === null ? null : `${path} : null attendu.`;
+    case "string":
+      if (typeof value !== "string")
+        return `${path} : texte attendu, reçu ${describe(value)}.`;
+      if (schema.maxLength !== undefined && value.length > schema.maxLength)
+        return `${path} : texte trop long (${value.length} > ${schema.maxLength}).`;
+      return null;
+    case "number":
+      if (typeof value !== "number" || !Number.isFinite(value))
+        return `${path} : nombre attendu, reçu ${describe(value)}.`;
+      if (schema.integer && !Number.isInteger(value))
+        return `${path} : entier attendu.`;
+      if (schema.minimum !== undefined && value < schema.minimum)
+        return `${path} : minimum ${schema.minimum}.`;
+      if (schema.maximum !== undefined && value > schema.maximum)
+        return `${path} : maximum ${schema.maximum}.`;
+      return null;
+    case "array": {
+      if (!Array.isArray(value))
+        return `${path} : liste attendue, reçu ${describe(value)}.`;
+      if (schema.maxItems !== undefined && value.length > schema.maxItems)
+        return `${path} : trop d'éléments (${value.length} > ${schema.maxItems}).`;
+      for (let i = 0; i < value.length; i++) {
+        const err = check(schema.items, value[i], `${path}[${i}]`);
+        if (err) return err;
+      }
+      return null;
+    }
+    case "tuple": {
+      if (!Array.isArray(value) || value.length !== schema.prefixItems.length)
+        return `${path} : liste de ${schema.prefixItems.length} éléments attendue.`;
+      for (let i = 0; i < value.length; i++) {
+        const err = check(schema.prefixItems[i], value[i], `${path}[${i}]`);
+        if (err) return err;
+      }
+      return null;
+    }
+  }
+}
+
+/**
+ * Valide des options d'appel venues de l'extérieur (SDUI, attribut HTML).
+ * Clés inconnues refusées, types et bornes vérifiés.
+ */
+export function validateCallOptions<T extends HfTask>(
+  task: T,
+  options: unknown,
+): HfCallOptions<T> {
+  if (options === undefined || options === null) return {} as HfCallOptions<T>;
+  if (typeof options !== "object" || Array.isArray(options)) {
+    throw new HfOptionsError(
+      `[sonic-hugging-face-infer] options.call : objet attendu, reçu ${describe(options)}.`,
+    );
+  }
+  const table = HF_CALL_OPTION_SCHEMAS[task] as Record<string, HfValueSchema>;
+  for (const [key, value] of Object.entries(options)) {
+    const schema = table[key];
+    if (!schema) {
+      const allowed = Object.keys(table);
+      throw new HfOptionsError(
+        `[sonic-hugging-face-infer] option "${key}" inconnue pour la tâche "${task}"` +
+          (allowed.length ? ` (autorisées : ${allowed.join(", ")}).` : " (aucune option)."),
+      );
+    }
+    if (value === undefined) continue;
+    const err = check(schema, value, `options.call.${key}`);
+    if (err) throw new HfOptionsError(`[sonic-hugging-face-infer] ${err}`);
+  }
+  return options as HfCallOptions<T>;
+}
+
+/** JSON Schema (draft 2020-12) des options d'appel d'une tâche, pour le catalogue MCP. */
+export function hfCallOptionsJsonSchema(task: HfTask): Record<string, unknown> {
+  const toJson = (s: HfValueSchema): Record<string, unknown> => {
+    if ("anyOf" in s) return { anyOf: s.anyOf.map(toJson) };
+    if ("enum" in s) return { enum: [...s.enum] };
+    switch (s.type) {
+      case "number": {
+        const out: Record<string, unknown> = {
+          type: s.integer ? "integer" : "number",
+        };
+        if (s.minimum !== undefined) out.minimum = s.minimum;
+        if (s.maximum !== undefined) out.maximum = s.maximum;
+        return out;
+      }
+      case "string":
+        return s.maxLength !== undefined
+          ? { type: "string", maxLength: s.maxLength }
+          : { type: "string" };
+      case "array":
+        return s.maxItems !== undefined
+          ? { type: "array", items: toJson(s.items), maxItems: s.maxItems }
+          : { type: "array", items: toJson(s.items) };
+      case "tuple":
+        return {
+          type: "array",
+          prefixItems: s.prefixItems.map(toJson),
+          minItems: s.prefixItems.length,
+          maxItems: s.prefixItems.length,
+        };
+      default:
+        return { type: s.type };
+    }
+  };
+  const table = HF_CALL_OPTION_SCHEMAS[task] as Record<string, HfValueSchema>;
+  return {
+    $schema: "https://json-schema.org/draft/2020-12/schema",
+    type: "object",
+    additionalProperties: false,
+    properties: Object.fromEntries(
+      Object.entries(table).map(([k, v]) => [k, toJson(v)]),
+    ),
+  };
+}

package/src/core/components/functional/infer/hugging-face/hf-worker-runtime.spec.ts

@@ -0,0 +1,147 @@
+import { describe, expect, it } from "vitest";
+import { buildHfWorkerSource, hfWorkerMain, type HfLibrary } from "./hf-worker-runtime";
+
+class FakeTensor {
+  constructor(
+    private rows: number[][],
+    public dims = [rows.length, rows[0]?.length ?? 0],
+  ) {}
+  tolist() {
+    return this.rows;
+  }
+}
+
+function harness(lib: HfLibrary) {
+  const sent: unknown[] = [];
+  let listener: (event: { data: unknown }) => void = () => {};
+  hfWorkerMain(lib, {
+    postMessage: (m) => sent.push(m),
+    addEventListener: (_t, l) => (listener = l),
+  });
+  let id = 0;
+  const request = async (body: Record<string, unknown>) => {
+    const myId = ++id;
+    listener({ data: { ...body, id: myId } });
+    for (let i = 0; i < 50; i++) {
+      await new Promise((r) => setTimeout(r, 1));
+      const reply = sent.find((m) => (m as { id?: number }).id === myId);
+      if (reply) return reply as { ok: boolean; result?: unknown; error?: { code: string } };
+    }
+    throw new Error("pas de réponse");
+  };
+  return { request, sent };
+}
+
+function fakeLib(): HfLibrary & { created: unknown[] } {
+  const created: unknown[] = [];
+  const lib = {
+    created,
+    env: { backends: { onnx: { wasm: {} } } } as HfLibrary["env"],
+    async pipeline(task: string, model: string, options?: Record<string, unknown>) {
+      created.push({ task, model, options });
+      (options?.progress_callback as (i: unknown) => void)?.({
+        status: "progress_total",
+        loaded: 5,
+        total: 10,
+        progress: 50,
+      });
+      const pipe = async (input: unknown, a?: unknown, b?: unknown) => {
+        if (task === "feature-extraction") {
+          const texts = Array.isArray(input) ? input : [input];
+          return new FakeTensor(texts.map((t) => [String(t).length, 1]));
+        }
+        if (task === "zero-shot-classification") {
+          return { sequence: input, labels: a, scores: [0.9, 0.1], opts: b };
+        }
+        return { input };
+      };
+      return pipe;
+    },
+  };
+  return lib;
+}
+
+describe("hfWorkerMain", () => {
+  it("configure l'environnement de Transformers.js (cache, hôte, WASM)", async () => {
+    const lib = fakeLib();
+    const { request } = harness(lib);
+    await request({
+      type: "configure",
+      env: { cacheKey: "concorde-hf-4.3.0", remoteHost: "https://cdn/models/", wasmPaths: "https://cdn/rt/" },
+    });
+    expect(lib.env).toMatchObject({
+      allowLocalModels: false,
+      useBrowserCache: true,
+      useWasmCache: true,
+      cacheKey: "concorde-hf-4.3.0",
+      remoteHost: "https://cdn/models/",
+    });
+    expect(lib.env.backends?.onnx?.wasm?.wasmPaths).toEqual({
+      mjs: "https://cdn/rt/ort-wasm-simd-threaded.asyncify.mjs",
+      wasm: "https://cdn/rt/ort-wasm-simd-threaded.asyncify.wasm",
+    });
+  });
+
+  it("repli WASM quand WebGPU manque, progression relayée", async () => {
+    const lib = fakeLib();
+    const { request, sent } = harness(lib);
+    const model = { alias: "m", task: "feature-extraction", repo: "r", revision: "sha", options: { dtype: "q8" } };
+    const reply = await request({ type: "load", model, devices: ["webgpu", "wasm"] });
+    expect(reply).toMatchObject({ ok: true, result: { backend: "wasm" } });
+    expect(lib.created[0]).toMatchObject({ options: { device: "wasm", revision: "sha", dtype: "q8" } });
+    expect(sent).toContainEqual({
+      type: "progress",
+      alias: "m",
+      progress: { loadedBytes: 5, totalBytes: 10, percent: 50 },
+    });
+  });
+
+  it("aucun backend autorisé utilisable → unsupported", async () => {
+    const { request } = harness(fakeLib());
+    const model = { alias: "m", task: "feature-extraction", repo: "r", revision: "sha", options: {} };
+    const reply = await request({ type: "load", model, devices: ["webgpu"] });
+    expect(reply).toMatchObject({ ok: false, error: { code: "unsupported" } });
+  });
+
+  it("run : lot d'embeddings sérialisé, argument positionnel replacé", async () => {
+    const { request } = harness(fakeLib());
+    const fe = { alias: "fe", task: "feature-extraction", repo: "r", revision: "s", options: {} };
+    const zs = { alias: "zs", task: "zero-shot-classification", repo: "r", revision: "s", options: {} };
+    await request({ type: "load", model: fe, devices: ["wasm"] });
+    await request({ type: "load", model: zs, devices: ["wasm"] });
+    const batch = await request({
+      type: "run",
+      request: { alias: "fe", inputs: ["ab", "abc"], call: { pooling: "mean" }, batch: true },
+    });
+    expect(batch.result).toEqual([
+      [2, 1],
+      [3, 1],
+    ]);
+    const single = await request({
+      type: "run",
+      request: { alias: "fe", inputs: ["abcd"], call: {}, batch: false },
+    });
+    expect(single.result).toEqual([[4, 1]]);
+    const labels = await request({
+      type: "run",
+      request: {
+        alias: "zs",
+        inputs: ["drôle"],
+        call: { candidate_labels: ["humour", "drame"], multi_label: true },
+        positional: "candidate_labels",
+        batch: false,
+      },
+    });
+    expect(labels.result).toEqual([
+      { sequence: "drôle", labels: ["humour", "drame"], scores: [0.9, 0.1], opts: { multi_label: true } },
+    ]);
+  });
+
+  it("la source du worker blob est autonome et syntaxiquement valide", () => {
+    const source = buildHfWorkerSource("https://cdn/hf/4.3.0/transformers.min.js");
+    expect(source).toContain('await import("https://cdn/hf/4.3.0/transformers.min.js")');
+    // Vérifie la syntaxe (module avec top-level await) sans l'exécuter.
+    const AsyncFunction = Object.getPrototypeOf(async () => {}).constructor;
+    expect(() => new AsyncFunction(source.replace(/\bself\b/g, "globalThis"))).not.toThrow();
+  });
+});