runbios-sdk 0.2.1-rc.98 → 0.2.2-dev.171

package/dist/resources/training.d.ts

@@ -1,5 +1,5 @@
 import type { HttpClient } from '../client.js';
-import type { TrainingCreateParams, TrainingListParams, TrainingJob, TrainingListResponse, TrainingMetrics, TrainingCheckpoint, TrainingLogs, TrainingStopResponse, TrainingResumeResponse, TrainingPreflightResponse, TrainingCapabilities } from '../types.js';
+import type { TrainingCreateParams, TrainingListParams, TrainingJob, TrainingListResponse, TrainingMetrics, TrainingCheckpoint, TrainingLogs, TrainingStopResponse, TrainingResumeResponse, TrainingPreflightResponse, TrainingRecommendParams, TrainingAdvisorVerdict, TrainingCapabilities } from '../types.js';
 /**
  * Create, monitor, and manage fine-tuning training jobs.
  */
@@ -12,8 +12,11 @@ export declare class Training {
     /**
      * 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 machine is
+     * Choose method, adapter, datasets and the complete hyperparameter config
+     * first, call {@link preflight}, then choose one returned GPU configuration
+     * last. Its minimum is derived from every earlier choice. This call blocks
+     * while that placement is booked (~40s typical). A training id and the
+     * "training started" email exist only once a real machine is
      * secured, so the returned `status` is one of:
      *
      * - `"booked"` — a GPU was secured (booked == secured); the job then
@@ -25,11 +28,12 @@ export declare class Training {
      *   (`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.
+     * If booking fails, this rejects with `CAPACITY_UNAVAILABLE` (HTTP 409)
+     * carrying fresh qualifying alternatives — no phantom job and no charge.
+     * Keep every earlier field unchanged, select one alternative and retry. The
+     * rejected type is never recommended back. Queue consent is appropriate only
+     * when the alternatives list is empty; waiting is unbilled. The SDK never
+     * auto-substitutes a GPU; a transient 503 is a retry, never a capacity verdict.
      *
      * @example
      * ```ts
@@ -51,6 +55,14 @@ export declare class Training {
     create(params: TrainingCreateParams): Promise<TrainingJob>;
     /** Validate and canonicalize a training request without creating or billing a job. */
     preflight(params: TrainingCreateParams): Promise<TrainingPreflightResponse>;
+    /**
+     * Ask the trainer's own sizing model what to run: per-device batch,
+     * accumulation, learning rate, warmup, predicted peak memory, minimum GPU
+     * count and a wall-clock estimate for this model on this GPU type, with the
+     * basis of every number. Side-effect free. When `available` is false no
+     * advisor is deployed and the other fields are absent -- nothing is guessed.
+     */
+    recommend(params: TrainingRecommendParams): Promise<TrainingAdvisorVerdict>;
     /** Return one server-driven page with pagination metadata. */
     listPage(params?: TrainingListParams): Promise<TrainingListResponse>;
     /**

package/dist/resources/training.js

@@ -67,10 +67,12 @@ function buildTrainingRequest(params) {
         body.integration_id = params.integrationId;
     if (params.networkVolumeId !== undefined)
         body.network_volume_id = params.networkVolumeId;
-    if (params.cacheDataset !== undefined)
-        body.cache_dataset = params.cacheDataset;
     if (params.datasetSampleLimits !== undefined)
         body.dataset_sample_limits = params.datasetSampleLimits;
+    if (params.datasetSampling !== undefined)
+        body.dataset_sampling = params.datasetSampling;
+    if (params.datasetSamplingStrategy !== undefined)
+        body.dataset_sampling_strategy = params.datasetSamplingStrategy;
     if (params.datasetMixing !== undefined)
         body.dataset_mixing = params.datasetMixing;
     if (params.mixing !== undefined)
@@ -78,6 +80,8 @@ function buildTrainingRequest(params) {
     const config = {};
     if (params.epochs !== undefined)
         config.num_train_epochs = params.epochs;
+    if (params.maxSteps !== undefined)
+        config.max_steps = params.maxSteps;
     if (params.batchSize !== undefined)
         config.per_device_train_batch_size = params.batchSize;
     if (params.gradientAccumulation !== undefined)
@@ -183,8 +187,11 @@ export class Training {
     /**
      * 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 machine is
+     * Choose method, adapter, datasets and the complete hyperparameter config
+     * first, call {@link preflight}, then choose one returned GPU configuration
+     * last. Its minimum is derived from every earlier choice. This call blocks
+     * while that placement is booked (~40s typical). A training id and the
+     * "training started" email exist only once a real machine is
      * secured, so the returned `status` is one of:
      *
      * - `"booked"` — a GPU was secured (booked == secured); the job then
@@ -196,11 +203,12 @@ export class Training {
      *   (`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.
+     * If booking fails, this rejects with `CAPACITY_UNAVAILABLE` (HTTP 409)
+     * carrying fresh qualifying alternatives — no phantom job and no charge.
+     * Keep every earlier field unchanged, select one alternative and retry. The
+     * rejected type is never recommended back. Queue consent is appropriate only
+     * when the alternatives list is empty; waiting is unbilled. The SDK never
+     * auto-substitutes a GPU; a transient 503 is a retry, never a capacity verdict.
      *
      * @example
      * ```ts
@@ -240,6 +248,41 @@ export class Training {
         const { body } = buildTrainingRequest(params);
         return this._http.fetchPost('/api/training/preflight', body);
     }
+    /**
+     * Ask the trainer's own sizing model what to run: per-device batch,
+     * accumulation, learning rate, warmup, predicted peak memory, minimum GPU
+     * count and a wall-clock estimate for this model on this GPU type, with the
+     * basis of every number. Side-effect free. When `available` is false no
+     * advisor is deployed and the other fields are absent -- nothing is guessed.
+     */
+    async recommend(params) {
+        const q = new URLSearchParams({ model_id: params.model, gpu_type: params.gpuType });
+        if (params.modelRevision)
+            q.set('model_revision', params.modelRevision);
+        if (params.integrationId)
+            q.set('integration_id', params.integrationId);
+        if (params.gpuCount !== undefined)
+            q.set('gpu_count', String(params.gpuCount));
+        if (params.adapter)
+            q.set('train_type', params.adapter);
+        if (params.method)
+            q.set('method', params.method === 'cpt' ? 'pt' : params.method);
+        if (params.maxLength !== undefined)
+            q.set('max_length', String(params.maxLength));
+        if (params.epochs !== undefined)
+            q.set('num_train_epochs', String(params.epochs));
+        if (params.maxSteps !== undefined)
+            q.set('max_steps', String(params.maxSteps));
+        if (params.perDeviceTrainBatchSize !== undefined)
+            q.set('per_device_train_batch_size', String(params.perDeviceTrainBatchSize));
+        if (params.gradientAccumulationSteps !== undefined)
+            q.set('gradient_accumulation_steps', String(params.gradientAccumulationSteps));
+        if (params.datasetIds?.length)
+            q.set('dataset_ids', params.datasetIds.join(','));
+        if (params.workspaceId)
+            q.set('workspace_id', params.workspaceId);
+        return this._http.fetchGet(`/api/training/recommend?${q}`);
+    }
     /** Return one server-driven page with pagination metadata. */
     async listPage(params = {}) {
         const q = new URLSearchParams();