@simulatte/doppler 0.1.6 → 0.1.8

package/src/inference/pipelines/text/generator-runtime.js

@@ -213,6 +213,10 @@ export function resolvePrefillEmbeddingOptions(state, options = {}) {
     ? state.manifest.modelType.toLowerCase()
     : '';
   const generationDefaults = state.runtimeConfig.inference.generation;
+  // Embedding models default to 'mean' pooling — this is a model-category behavior,
+  // not a model-family identity check. Ideally embedding model presets would set
+  // generation.embeddingMode='mean' in their runtime config; the modelType fallback
+  // provides this default for manifests that predate runtime-preset embedding mode.
   const defaultEmbeddingMode = modelType === 'embedding'
     ? 'mean'
     : generationDefaults.embeddingMode;
@@ -226,6 +230,7 @@ export function resolveAdvanceEmbeddingMode(state, options = {}) {
   const modelType = typeof state.manifest?.modelType === 'string'
     ? state.manifest.modelType.toLowerCase()
     : '';
+  // See resolvePrefillEmbeddingOptions for embedding-model pooling rationale.
   const configuredMode = state.runtimeConfig.inference.generation.embeddingMode;
   return resolveConfiguredValue(
     options.embeddingMode,

package/src/inference/pipelines/text/generator-steps.d.ts

@@ -19,6 +19,58 @@ export declare function resolveBatchStop(
   eosTokenId: number | undefined | null
 ): number;
 
+export declare function findInvalidGeneratedToken(
+  tokens: number[],
+  vocabSize: number,
+  padTokenId?: number | null
+): { index: number; tokenId: number } | null;
+
+export interface SampledTokenStagingBuffer {
+  mapAsync(mode: number): Promise<void>;
+  getMappedRange(): ArrayBufferLike;
+  unmap(): void;
+  destroy(): void;
+}
+
+export declare function readSampledTokenFromStagingBuffer(
+  stagingBuffer: SampledTokenStagingBuffer,
+  options?: {
+    ownsStagingBuffer?: boolean;
+    hasFinitenessBuffer?: boolean;
+    ring?: { advance(): void } | null;
+  }
+): Promise<{
+  nextToken: number;
+  finitenessStatus: {
+    triggered: boolean;
+    metadata: string;
+  };
+}>;
+
+export declare function readMappedBufferCopy(
+  stagingBuffer: SampledTokenStagingBuffer,
+  options?: {
+    ownsStagingBuffer?: boolean;
+  }
+): Promise<ArrayBuffer>;
+
+export declare function readBatchTokensFromStagingBuffers(options: {
+  tokensStagingBuffer: SampledTokenStagingBuffer;
+  stopStagingBuffer?: SampledTokenStagingBuffer | null;
+  finitenessStagingBuffer?: SampledTokenStagingBuffer | null;
+  tokenCount: number;
+  ownsTokensStaging?: boolean;
+  ownsStopStaging?: boolean;
+  ring?: { advance(): void } | null;
+}): Promise<{
+  tokens: number[];
+  stopFlags: Uint32Array | null;
+  finitenessStatus: {
+    triggered: boolean;
+    metadata: string;
+  };
+}>;
+
 export declare function decodeStep(
   state: unknown,
   currentIds: number[],