openpond-sdk 0.12.8 → 0.12.10

package/dist/training-bundle.js

@@ -1,7 +1,7 @@
 import {
   prepareImportedTasksetPackage,
   projectLearningBatchGraders
-} from "./chunk-EEZ5YHJQ.js";
+} from "./chunk-ICFFZ4WC.js";
 import {
   projectPortableTaskRecord,
   resolvePortableTasksetRewardExecution
@@ -10,15 +10,15 @@ import {
   TasksetDraftOutputContractSchema,
   TasksetSchema,
   computeTasksetHash
-} from "./chunk-IOZCYSJ6.js";
+} from "./chunk-UAYXORU5.js";
 import {
   LearningBatchDatasetSourceRefSchema,
   MAX_TASKSET_PACKAGE_BYTES,
   decodeTasksetPackageFile,
   validateTasksetPackage
-} from "./chunk-RFKZSSLB.js";
-import "./chunk-TK3DXH7A.js";
+} from "./chunk-7WKCQ4V5.js";
 import {
+  ManagedSftRecipeSchema,
   TRAINING_EVALUATION_SOURCE_PATH,
   TrainingCapabilitiesSchema,
   TrainingEvaluationSourceSchema,
@@ -30,11 +30,15 @@ import {
   trainingEvaluationSourceRef,
   trainingInputArtifactUploadHash,
   trainingJobSubmissionHash
-} from "./chunk-EYMZQML2.js";
+} from "./chunk-WOD72SFI.js";
+import "./chunk-EZME7DV4.js";
+import "./chunk-GX6KMHW2.js";
+import "./chunk-TK3DXH7A.js";
+import "./chunk-H3367ISL.js";
 import {
   TrainingHandoffOriginSchema
-} from "./chunk-MRUDFIB6.js";
-import "./chunk-UWM2ZPGS.js";
+} from "./chunk-5QMREXMU.js";
+import "./chunk-7FYPFHCQ.js";
 import {
   HarnessSourceSelectionSchema,
   createAgentSnapshot,
@@ -42,13 +46,13 @@ import {
   validateHarnessPolicySourcePackage,
   validateHarnessSourcePackage
 } from "./chunk-Q4TKI2MV.js";
-import "./chunk-H3367ISL.js";
-import "./chunk-WYR3MZHU.js";
-import "./chunk-EWRWGFKD.js";
-import "./chunk-CJDEVI3T.js";
+import "./chunk-HDKP6KE2.js";
+import "./chunk-IIQENUOK.js";
+import "./chunk-LT2YQCZI.js";
 import {
   TrainingGpuSelectionSchema
-} from "./chunk-YSEYPMQ3.js";
+} from "./chunk-BQ2C2Y54.js";
+import "./chunk-B3ZHJ2KK.js";
 import "./chunk-CL3IOOVL.js";
 import {
   ImmutableReleaseRefSchema,
@@ -371,6 +375,7 @@ async function prepareManagedTrainingSubmission(raw) {
   const manifest = HarnessRunManifestSchema.parse(input.manifest);
   const bundle = ResolvedTrainingBundleManifestSchema.parse(input.bundleManifest);
   const qualification = input.executionMode === "qualification";
+  const gpuSelection = input.gpuSelection ?? project.trainingSetup.managedGpuSelection ? TrainingGpuSelectionSchema.parse(input.gpuSelection ?? project.trainingSetup.managedGpuSelection) : void 0;
   const physicalPlacement = input.physicalPlacement ? TrainingPhysicalPlacementSchema.parse(input.physicalPlacement) : void 0;
   const qualificationGroupSize = manifest.qualificationSamples === void 0 ? null : z2.object({ rollout: z2.object({ groupSize: z2.number().int().min(2).max(32) }).passthrough() }).passthrough().parse(recipe).rollout.groupSize;
   if (manifest.qualificationSamples !== bundle.qualificationSamples || manifest.qualificationSamples !== void 0 && (!qualification || manifest.qualificationSamples > qualificationGroupSize)) {
@@ -393,8 +398,17 @@ async function prepareManagedTrainingSubmission(raw) {
   if (!baseModel || baseModel.modelId !== manifest.model.source || baseModel.revision !== manifest.model.revision || baseModel.tokenizerRevision !== manifest.model.tokenizerRevision || baseModel.chatTemplateHash !== manifest.model.chatTemplateHash) {
     throw new Error("The synced Model Project base Model does not match the Run manifest.");
   }
-  if (recipe.method !== "grpo" || contentHash(recipe) !== manifest.recipe.configHash || approval.approvalHash !== manifest.approval.approvalHash || approval.maximumSpendUsd !== manifest.approval.maximumSpendUsd) {
-    throw new Error("Managed training requires the exact approved GRPO recipe and spend ceiling.");
+  if (!["grpo", "sft"].includes(recipe.method) || contentHash(recipe) !== manifest.recipe.configHash || approval.approvalHash !== manifest.approval.approvalHash || approval.maximumSpendUsd !== manifest.approval.maximumSpendUsd) {
+    throw new Error("Managed training requires the exact approved supported recipe and spend ceiling.");
+  }
+  if (recipe.method === "sft") {
+    const supervised = ManagedSftRecipeSchema.parse(recipe);
+    if (contentHash(supervised) !== contentHash(recipe))
+      throw new Error("SFT recipe must be fully materialized before approval.");
+    if (qualification || input.resumeFrom || physicalPlacement || (gpuSelection?.gpuCount ?? 1) !== 1)
+      throw new Error(
+        "Initial managed SFT requires a fresh single-GPU run without reuse-only placement or rollout qualification."
+      );
   }
   const limits = z2.object({ wallTimeMs: z2.number().positive() }).parse(recipe.resourceLimits);
   if (input.idempotencyKey !== void 0)
@@ -472,7 +486,8 @@ async function prepareManagedTrainingSubmission(raw) {
       resumeFrom: input.resumeFrom
     },
     requestedCapabilities: [
-      { id: qualification ? "managed_rl.policy.qualification" : "managed_rl.policy.grpo", version: "1", required: true },
+      { id: qualification ? "managed_rl.policy.qualification" : recipe.method === "sft" ? "managed_training.sft" : "managed_rl.policy.grpo", version: "1", required: true },
+      ...gpuSelection?.gpuCount === 2 ? [{ id: "managed_rl.policy.data_parallel", version: "1", required: true }] : [],
       {
         id: `managed_rl.rollouts.${manifest.runtimeTarget.placement}`,
         version: "1",
@@ -481,7 +496,7 @@ async function prepareManagedTrainingSubmission(raw) {
       ...project.trainingSetup.managedGpuRequirement === "h100_hbm3" ? [{ id: "managed_rl.gpu.h100_hbm3", version: "1", required: true }] : []
     ],
     placementObjective: project.trainingSetup.managedGpuPlacementObjective,
-    ...input.gpuSelection ?? project.trainingSetup.managedGpuSelection ? { gpuSelection: TrainingGpuSelectionSchema.parse(input.gpuSelection ?? project.trainingSetup.managedGpuSelection) } : {},
+    ...gpuSelection ? { gpuSelection } : {},
     ...physicalPlacement ? { physicalPlacement } : {},
     budget: {
       maximumSpendUsd: approval.maximumSpendUsd,