bios-sdk 0.1.1-dev.16 → 0.1.1-dev.19

package/dist/resources/gpu.d.ts

@@ -33,7 +33,7 @@ export declare class GPU {
      * Uses the model's parameter count and architecture to suggest
      * the best GPU type and count for training.
      *
-     * @param modelId - Full HuggingFace model ID (e.g. "meta-llama/Llama-3.1-8B-Instruct").
+     * @param modelId - Full model id (e.g. "meta-llama/Llama-3.1-8B-Instruct").
      *
      * @example
      * ```ts

package/dist/resources/gpu.js

@@ -54,7 +54,7 @@ export class GPU {
      * Uses the model's parameter count and architecture to suggest
      * the best GPU type and count for training.
      *
-     * @param modelId - Full HuggingFace model ID (e.g. "meta-llama/Llama-3.1-8B-Instruct").
+     * @param modelId - Full model id (e.g. "meta-llama/Llama-3.1-8B-Instruct").
      *
      * @example
      * ```ts

package/dist/resources/models.d.ts

@@ -1,7 +1,7 @@
 import type { HttpClient } from '../client.js';
 import type { ModelSearchParams, ModelSearchResponse, ModelConfig, AdapterCompatibilityParams, AdapterCompatibilityResponse, ArchitectureScope, SupportedArchitecturesResponse } from '../types.js';
 /**
- * Access the BIOS model catalog -- search HuggingFace models, fetch
+ * Access the BIOS model catalog -- search models, fetch
  * training-relevant configuration, and check adapter compatibility.
  */
 export declare class Models {
@@ -9,7 +9,7 @@ export declare class Models {
     /** @internal */
     constructor(_http: HttpClient);
     /**
-     * Search the HuggingFace model catalog for fine-tunable models.
+     * Search the model catalog for fine-tunable models.
      *
      * Results are filtered to exclude quantized derivatives (GGUF, GPTQ, AWQ)
      * and adapters by default -- pass `kind: "adapter"` to find LoRA adapters.
@@ -24,13 +24,13 @@ export declare class Models {
      */
     search(params?: ModelSearchParams): Promise<ModelSearchResponse>;
     /**
-     * Fetch the training configuration for a specific model from HuggingFace.
+     * Fetch the training configuration for a specific model from the catalog.
      *
      * Returns parameter counts, architecture type, and whether the model
      * uses Mixture-of-Experts -- information needed to choose the right
      * GPU and adapter configuration.
      *
-     * @param modelId - Full HuggingFace model ID (e.g. "meta-llama/Llama-3.1-8B").
+     * @param modelId - Full model id (e.g. "meta-llama/Llama-3.1-8B").
      *
      * @example
      * ```ts
@@ -58,7 +58,7 @@ export declare class Models {
     /**
      * List the architectures the platform supports for a given scope. This is
      * the authoritative registry the platform gates on:
-     *  - `inference` — HF architecture classes that can be served (deployed).
+     *  - `inference` — architecture classes that can be served (deployed).
      *  - `training` — architecture keys the fine-tuning wizard/gate accepts.
      *
      * A `count` of 0 means the registry is empty and nothing is explicitly

package/dist/resources/models.js

@@ -1,5 +1,5 @@
 /**
- * Access the BIOS model catalog -- search HuggingFace models, fetch
+ * Access the BIOS model catalog -- search models, fetch
  * training-relevant configuration, and check adapter compatibility.
  */
 export class Models {
@@ -9,7 +9,7 @@ export class Models {
         this._http = _http;
     }
     /**
-     * Search the HuggingFace model catalog for fine-tunable models.
+     * Search the model catalog for fine-tunable models.
      *
      * Results are filtered to exclude quantized derivatives (GGUF, GPTQ, AWQ)
      * and adapters by default -- pass `kind: "adapter"` to find LoRA adapters.
@@ -49,13 +49,13 @@ export class Models {
         return this._http.fetchGet(`/api/public/model-search?${q}`);
     }
     /**
-     * Fetch the training configuration for a specific model from HuggingFace.
+     * Fetch the training configuration for a specific model from the catalog.
      *
      * Returns parameter counts, architecture type, and whether the model
      * uses Mixture-of-Experts -- information needed to choose the right
      * GPU and adapter configuration.
      *
-     * @param modelId - Full HuggingFace model ID (e.g. "meta-llama/Llama-3.1-8B").
+     * @param modelId - Full model id (e.g. "meta-llama/Llama-3.1-8B").
      *
      * @example
      * ```ts
@@ -96,7 +96,7 @@ export class Models {
     /**
      * List the architectures the platform supports for a given scope. This is
      * the authoritative registry the platform gates on:
-     *  - `inference` — HF architecture classes that can be served (deployed).
+     *  - `inference` — architecture classes that can be served (deployed).
      *  - `training` — architecture keys the fine-tuning wizard/gate accepts.
      *
      * A `count` of 0 means the registry is empty and nothing is explicitly

package/dist/resources/training.d.ts

@@ -10,11 +10,26 @@ export declare class Training {
     /** Return the authoritative BIOS image, method, adapter, and hyperparameter contract. */
     capabilities(): Promise<TrainingCapabilities>;
     /**
-     * Create a new training job.
-     *
-     * Provisions GPU resources, downloads the model and dataset, and begins
-     * training. The job starts in "pending" status and progresses through
-     * provisioning, downloading, and running states.
+     * Create a new training job (book-before-reveal).
+     *
+     * The call blocks while the ranked GPU ladder is booked (~40s typical). A
+     * training id and the "training started" email exist only once a real pod is
+     * secured, so the returned `status` is one of:
+     *
+     * - `"booked"` — a GPU was secured (booked == secured); the job then
+     *   progresses through provisioning, downloading, and running on its own.
+     * - `"securing"` — still booking at the deadline; the async tail continues.
+     *   Poll {@link get} until it reaches `booked`/`running` (or terminal).
+     *   Nothing is charged until a pod is real.
+     * - `"queued"` — returned only with explicit queue consent
+     *   (`queueIfUnavailable: true`); waits for stock at zero charge and books
+     *   via the same path.
+     *
+     * If the ladder is exhausted at booking time without queue consent, this
+     * rejects with a `CAPACITY_UNAVAILABLE` (HTTP 409) carrying neutral
+     * alternatives — no phantom job remains and nothing is billed. The SDK never
+     * auto-substitutes a GPU; a transient 503 is a retry, never a capacity
+     * verdict.
      *
      * @example
      * ```ts
@@ -120,7 +135,12 @@ export declare class Training {
      */
     stop(id: string, keepData?: boolean): Promise<TrainingStopResponse>;
     /**
-     * Resume a stopped training job from its last checkpoint.
+     * Resume a stopped or `interrupted` training job from its last checkpoint.
+     *
+     * Use this on a `stopped` job, or one resting at `interrupted` — the
+     * self-heal state a started job enters when its pod is lost and billing has
+     * already been stopped (distinct from `failed`). Resume re-books on secured
+     * capacity before reporting resumed, so book-before-reveal still applies.
      *
      * @param id - Training job ID.
      *

package/dist/resources/training.js

@@ -191,11 +191,26 @@ export class Training {
         return this._http.fetchGet('/api/training/capabilities');
     }
     /**
-     * Create a new training job.
+     * Create a new training job (book-before-reveal).
      *
-     * Provisions GPU resources, downloads the model and dataset, and begins
-     * training. The job starts in "pending" status and progresses through
-     * provisioning, downloading, and running states.
+     * The call blocks while the ranked GPU ladder is booked (~40s typical). A
+     * training id and the "training started" email exist only once a real pod is
+     * secured, so the returned `status` is one of:
+     *
+     * - `"booked"` — a GPU was secured (booked == secured); the job then
+     *   progresses through provisioning, downloading, and running on its own.
+     * - `"securing"` — still booking at the deadline; the async tail continues.
+     *   Poll {@link get} until it reaches `booked`/`running` (or terminal).
+     *   Nothing is charged until a pod is real.
+     * - `"queued"` — returned only with explicit queue consent
+     *   (`queueIfUnavailable: true`); waits for stock at zero charge and books
+     *   via the same path.
+     *
+     * If the ladder is exhausted at booking time without queue consent, this
+     * rejects with a `CAPACITY_UNAVAILABLE` (HTTP 409) carrying neutral
+     * alternatives — no phantom job remains and nothing is billed. The SDK never
+     * auto-substitutes a GPU; a transient 503 is a retry, never a capacity
+     * verdict.
      *
      * @example
      * ```ts
@@ -363,7 +378,12 @@ export class Training {
         return this._http.fetchPost(`/api/training/jobs/${encodeURIComponent(id)}/stop`, { keep_data: keepData });
     }
     /**
-     * Resume a stopped training job from its last checkpoint.
+     * Resume a stopped or `interrupted` training job from its last checkpoint.
+     *
+     * Use this on a `stopped` job, or one resting at `interrupted` — the
+     * self-heal state a started job enters when its pod is lost and billing has
+     * already been stopped (distinct from `failed`). Resume re-books on secured
+     * capacity before reporting resumed, so book-before-reveal still applies.
      *
      * @param id - Training job ID.
      *

package/dist/types.d.ts

@@ -99,9 +99,9 @@ export interface AvailableGpuAlternative {
     tier?: string;
     recommended?: boolean;
 }
-/** A model returned from the HuggingFace-backed model search. */
+/** A model returned from the catalog model search. */
 export interface Model {
-    /** Full HuggingFace model ID (e.g. "meta-llama/Llama-3.1-8B"). */
+    /** Full model id (e.g. "meta-llama/Llama-3.1-8B"). */
     id: string;
     /** Model author / organization. */
     author: string;
@@ -115,11 +115,11 @@ export interface Model {
     architecture: 'dense' | 'moe';
     /** Model type: "llm" or "vlm". */
     modelType: 'llm' | 'vlm';
-    /** HuggingFace model_type from config.json. */
+    /** Raw model_type from config.json. */
     hfModelType: string | null;
-    /** Download count on HuggingFace. */
+    /** Download count from the catalog. */
     downloads: number;
-    /** Like count on HuggingFace. */
+    /** Like count from the catalog. */
     likes: number;
     /** Whether this is a private model. */
     isPrivate: boolean;
@@ -148,7 +148,7 @@ export interface ModelSearchParams {
     visibility?: 'all' | 'public' | 'private';
     /** Model kind: "full" (base models only) or "adapter". */
     kind?: 'full' | 'adapter';
-    /** HuggingFace integration ID for accessing private models. */
+    /** Integration ID for accessing private models. */
     integrationId?: string;
 }
 /** Response from a model search. */
@@ -160,9 +160,9 @@ export interface ModelSearchResponse {
     hasMore: boolean;
     hfUsername: string;
 }
-/** Model configuration from HuggingFace config.json. */
+/** Model configuration from the model's config.json. */
 export interface ModelConfig {
-    /** Full HuggingFace model ID. */
+    /** Full model id. */
     modelId: string;
     /** Model type from config.json (e.g. "llama", "qwen2"). */
     model_type: string | null;
@@ -389,9 +389,9 @@ export type RLHFAlgorithm = 'dpo' | 'simpo' | 'cpo' | 'orpo' | 'kto' | 'rm';
 /** Supported adapter types. */
 export type AdapterType = 'lora' | 'qlora' | 'adalora' | 'full' | 'loha' | 'lokr' | 'boft' | 'oft' | 'vera' | 'fourierft' | 'bone' | 'adapter' | 'reft' | 'llamapro' | 'longlora';
 /** Training job status values. */
-export type TrainingJobStatus = 'pending' | 'queued' | 'provisioning' | 'preparing' | 'starting' | 'downloading' | 'running' | 'completed' | 'failed' | 'saving' | 'stopped' | 'stopping' | 'resuming' | 'cancelled';
+export type TrainingJobStatus = 'pending' | 'queued' | 'securing' | 'booked' | 'provisioning' | 'preparing' | 'starting' | 'downloading' | 'running' | 'completed' | 'failed' | 'interrupted' | 'saving' | 'stopped' | 'stopping' | 'resuming' | 'cancelled';
 /** Status values accepted by GET /api/training/jobs?status=. */
-export type TrainingStatusFilter = 'pending' | 'queued' | 'provisioning' | 'preparing' | 'starting' | 'running' | 'saving' | 'completed' | 'failed' | 'stopped';
+export type TrainingStatusFilter = 'pending' | 'queued' | 'securing' | 'booked' | 'provisioning' | 'preparing' | 'starting' | 'running' | 'interrupted' | 'saving' | 'completed' | 'failed' | 'stopped';
 /** One ranked GPU fallback choice. The first choice is the primary. */
 export interface GPUChoice {
     gpuType: string;
@@ -407,9 +407,9 @@ export interface GPUChoice {
 export interface TrainingCreateParams {
     /** Stable retry key. Reuse it after a timeout; a new value creates a new job. Auto-generated when omitted. */
     idempotencyKey?: string;
-    /** HuggingFace model ID to fine-tune. */
+    /** Model id to fine-tune. */
     model: string;
-    /** Optional Hugging Face branch, tag, or commit; the API canonicalizes it to an exact commit. */
+    /** Optional model branch, tag, or commit; the API canonicalizes it to an exact commit. */
     modelRevision?: string;
     /** One dataset ID to train on. Kept for compatibility with single-dataset jobs. */
     datasetId?: string;
@@ -490,7 +490,7 @@ export interface TrainingCreateParams {
     modelParamsB?: number;
     /** Active-parameter hint for MoE models in billions. */
     modelActiveParamsB?: number;
-    /** HuggingFace integration used for gated/private model access. */
+    /** Integration used for gated/private model access. */
     integrationId?: string;
     /** Existing network volume to attach. */
     networkVolumeId?: string;
@@ -882,9 +882,9 @@ export interface GPUPricingResponse {
 /** Parameters understood by the authenticated training GPU-options endpoint. */
 export interface GPUOptionsParams {
     modelId: string;
-    /** Requested Hugging Face branch, tag, or commit; response returns the exact commit. */
+    /** Requested model branch, tag, or commit; response returns the exact commit. */
     modelRevision?: string;
-    /** Stored Hugging Face integration used for gated or private repositories. */
+    /** Stored integration used for gated or private models. */
     integrationId?: string;
     /** Canonical adapter/train type. */
     trainType?: AdapterType | string;
@@ -928,9 +928,9 @@ export interface GPUOptionSuggestion {
 }
 export interface GPUOptionsResponse {
     options: GPUOption[];
-    /** Canonical Hugging Face owner/repository identity. */
+    /** Canonical owner/model identity. */
     model_id: string;
-    /** Exact immutable 40-hex Hugging Face commit used for sizing. */
+    /** Exact immutable 40-hex model commit used for sizing. */
     model_revision: string;
     model_params_b: number;
     min_vram_gb: number;
@@ -1253,7 +1253,7 @@ export interface InferenceGPUAlternative {
 export interface InferenceGPUOptionsParams {
     /**
      * MODEL-ADDRESSED sizing (recommended): the SERVER resolves the model facts
-     * from this HF repo id (optionally with `revision`) and computes
+     * from this model id (optionally with `revision`) and computes
      * min_gpus/valid_counts/bookable_counts — the same single implementation the
      * create gate enforces. When set, the deprecated client-fact params below
      * are unnecessary.
@@ -1262,7 +1262,7 @@ export interface InferenceGPUOptionsParams {
     revision?: string;
     /** Size from an existing deployment's pinned facts instead of a model id. */
     inferenceId?: string;
-    /** HF integration for gated models when using `model`. */
+    /** Integration for gated models when using `model`. */
     hfIntegrationId?: string;
     /** DEPRECATED client-fact path; required only when `model`/`inferenceId` are absent. */
     paramsB?: number;
@@ -1406,10 +1406,10 @@ export interface AdapterCompatibilityResponse {
 }
 /** One entry in the supported-architecture registry. */
 export interface SupportedArchitecture {
-    /** For inference: the HF architecture class (e.g. "LlamaForCausalLM"). For training: the arch key (e.g. "llama"). */
+    /** For inference: the architecture class (e.g. "LlamaForCausalLM"). For training: the arch key (e.g. "llama"). */
     architecture: string;
     display_name: string;
-    /** Comma-separated HF model_type aliases this entry covers. */
+    /** Comma-separated model_type aliases this entry covers. */
     model_types: string;
     enabled: boolean;
     notes: string;
@@ -1426,7 +1426,7 @@ export interface SupportedArchitecturesResponse {
 export interface AdapterCompatibilityParams {
     /** Architecture key (e.g. "llama", "qwen2"). */
     architecture?: string;
-    /** HuggingFace model_type for automatic architecture detection. */
+    /** The model's model_type for automatic architecture detection. */
     modelType?: string;
     /** Training method. Defaults to "sft". */
     trainingMethod?: TrainingMethod;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "bios-sdk",
-  "version": "0.1.1-dev.16",
+  "version": "0.1.1-dev.19",
   "description": "Official TypeScript SDK for the BIOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",