@simulatte/doppler 0.1.8 → 0.2.0

package/src/config/schema/inference-defaults.schema.d.ts

@@ -1,276 +0,0 @@
-/**
- * Inference Defaults Config Schema
- *
- * Default values for inference pipeline: batching, sampling, generation.
- * These defaults are used when no model-specific or user overrides are provided.
- *
- * Note: SamplingDefaultsSchema provides defaults for fields from SamplingSchema
- * (in inference.schema.js), plus greedyThreshold and repetitionPenaltyWindow
- * which are unique to defaults.
- *
- * @module config/schema/inference-defaults
- */
-
-import type {
-  ChatTemplateSchema,
-  InferenceConfigSchema,
-  LayerPipelineSchema,
-  SamplingSchema,
-  TokenizerConfigSchema,
-} from './inference.schema.js';
-import type {
-  ExecutionV0PatchSchema,
-  ExecutionV0SessionDefaultsSchema,
-} from './execution-v0.schema.js';
-import type { KVCacheConfigSchema } from './kvcache.schema.js';
-import type { MoERuntimeConfigSchema } from './moe.schema.js';
-import type { KernelPathRef } from './kernel-path.schema.js';
-import type { ManifestInferenceSchema } from './manifest.schema.js';
-import type { SpeculativeConfigSchema } from './speculative.schema.js';
-import type { DiffusionConfigSchema } from './diffusion.schema.js';
-import type { EnergyConfigSchema } from './energy.schema.js';
-
-// Re-export for convenience
-export type { ManifestInferenceSchema };
-
-/**
- * Deep partial type for runtime overrides of model-specific inference config.
- * Allows overriding any nested field in ManifestInferenceSchema.
- */
-export type ModelInferenceOverrides = {
-  [P in keyof ManifestInferenceSchema]?: ManifestInferenceSchema[P] extends object
-    ? { [K in keyof ManifestInferenceSchema[P]]?: ManifestInferenceSchema[P][K] }
-    : ManifestInferenceSchema[P];
-};
-
-/**
- * Default batching configuration for inference.
- *
- * Controls how tokens are batched during generation.
- */
-export interface BatchingDefaultsSchema {
-  /** Number of sequences to process in parallel (default: 1) */
-  batchSize: number;
-
-  /** Maximum tokens to generate per sequence (default: 512) */
-  maxTokens: number;
-
-  /** When to check for stop conditions: per-token or per-batch (default: 'batch') */
-  stopCheckMode: 'per-token' | 'batch';
-
-  /**
-   * Number of GPU decode batches to run before readback.
-   * null means read back after every batch (default: 1).
-   */
-  readbackInterval: number | null;
-
-  /** Ring size for token output buffers (null = disable ring) */
-  ringTokens: number | null;
-
-  /** Ring size for per-token stop buffers (null = disable ring) */
-  ringStop: number | null;
-
-  /** Ring size for staging readback buffers (null = disable ring) */
-  ringStaging: number | null;
-}
-
-/** Default batching configuration */
-export declare const DEFAULT_BATCHING_DEFAULTS: BatchingDefaultsSchema;
-
-/**
- * Default compute precision configuration.
- *
- * Controls dtype for intermediate activations and compute operations.
- * F16 reduces memory bandwidth by 2x but may have precision implications.
- */
-export interface ComputeDefaultsSchema {
-  /** Dtype for hidden state activations (default: 'f32', experimental: 'f16') */
-  activationDtype: 'f16' | 'f32';
-
-  /** Parameter count threshold for "large model" classification (default: 4e9 = 4B params) */
-  largeModelParamThreshold: number;
-
-  /** Multiplier for estimating model params from hidden^2 × layers (default: 12) */
-  paramEstimationMultiplier: number;
-
-  /** Keep weights in F32 (skip downcast even when F16 is available) */
-  keepF32Weights: boolean;
-
-  /**
-   * Range-aware selective widening policy used by finiteness guard.
-   * When enabled, decode checks for abs(value) > absThreshold and widens
-   * dangerous tokens to F32 fallback before KV corruption.
-   */
-  rangeAwareSelectiveWidening: RangeAwareSelectiveWideningDefaultsSchema;
-
-  /**
-   * Number of decode tokens to keep in widened (F32) mode after an overflow
-   * hit before restoring the original dtype path.
-   */
-  deferredRoundingWindowTokens: number;
-}
-
-export interface RangeAwareSelectiveWideningDefaultsSchema {
-  /** Enable range-aware widening guard for F16 activations. */
-  enabled: boolean;
-  /** Absolute-value threshold used for early widening checks. */
-  absThreshold: number;
-  /** Whether non-finite values should always trigger widening. */
-  includeNonFinite: boolean;
-}
-
-/** Default compute configuration */
-export declare const DEFAULT_COMPUTE_DEFAULTS: ComputeDefaultsSchema;
-
-/**
- * Configuration for oversized weights (embeddings, LM head).
- *
- * When weights exceed device binding limits, DOPPLER can keep them on CPU
- * and stream chunks to the GPU for matmul or gather operations.
- */
-export interface LargeWeightConfigSchema {
-  /** Enable CPU-backed chunking for oversized weights */
-  enabled: boolean;
-  /** Safety ratio applied to GPU binding limits (0..1). Default: 0.9 */
-  safetyRatio: number;
-  /** Prefer uploading F16 chunks when supported (reduces chunk size) */
-  preferF16: boolean;
-  /** Optional override for LM head chunk rows (null = auto) */
-  lmHeadChunkRows?: number | null;
-}
-
-/** Default large-weight configuration */
-export declare const DEFAULT_LARGE_WEIGHT_CONFIG: LargeWeightConfigSchema;
-
-/**
- * Default sampling configuration for token selection.
- *
- * Extends Required<SamplingSchema> with greedyThreshold and repetitionPenaltyWindow
- * for runtime decisions.
- * SamplingSchema (in inference.schema.js) uses optional fields for partial overrides;
- * this schema provides concrete defaults for all sampling parameters.
- */
-export interface SamplingDefaultsSchema extends Required<SamplingSchema> {
-  /** Temperature below this uses greedy decoding (default: 0.01) */
-  greedyThreshold: number;
-
-  /** Number of recent tokens to consider for repetition penalty (default: 100) */
-  repetitionPenaltyWindow: number;
-}
-
-/** Default sampling configuration */
-export declare const DEFAULT_SAMPLING_DEFAULTS: SamplingDefaultsSchema;
-
-/** Default generation behavior flags used by generator runtime paths. */
-export interface GenerationDefaultsSchema {
-  useSpeculative: boolean;
-  profile: boolean;
-  benchmark: boolean;
-  disableCommandBatching: boolean;
-  disableMultiTokenDecode: boolean;
-  embeddingMode: 'last' | 'mean';
-}
-
-/** Default generation behavior flags */
-export declare const DEFAULT_GENERATION_CONFIG: GenerationDefaultsSchema;
-
-/**
- * Default tokenizer configuration.
- *
- * Provides defaults for common tokenizer behavior. Actual token strings
- * come from the model's tokenizer config; these control runtime behavior.
- */
-export interface TokenizerDefaultsSchema {
-  /** Add BOS token to input (default: true for most models) */
-  addBosToken: boolean;
-
-  /** Add EOS token to output (default: false, model decides) */
-  addEosToken: boolean;
-
-  /** Defer token validation until token metadata is fully loaded */
-  deferSpecialTokens: boolean;
-}
-
-/** Default tokenizer configuration */
-export declare const DEFAULT_TOKENIZER_DEFAULTS: TokenizerDefaultsSchema;
-
-/**
- * Complete inference defaults configuration schema.
- *
- * Combines batching, sampling, compute, and tokenizer defaults for the inference pipeline.
- */
-export interface InferenceDefaultsConfigSchema {
-  batching: BatchingDefaultsSchema;
-  sampling: SamplingDefaultsSchema;
-  compute: ComputeDefaultsSchema;
-  tokenizer: TokenizerDefaultsSchema;
-  /** Handling for oversized embeddings/LM head */
-  largeWeights: LargeWeightConfigSchema;
-  /** KV cache configuration */
-  kvcache: KVCacheConfigSchema;
-  /** MoE routing and cache configuration */
-  moe: MoERuntimeConfigSchema;
-  /** Speculative decoding configuration */
-  speculative: SpeculativeConfigSchema;
-  /** Generation behavior defaults */
-  generation: GenerationDefaultsSchema;
-  /** Diffusion pipeline configuration */
-  diffusion: DiffusionConfigSchema;
-  /** Energy pipeline configuration */
-  energy: EnergyConfigSchema;
-  /** Optional default prompt text for test harnesses */
-  prompt?: string | null;
-  /** Log token ids in test harness output */
-  debugTokens?: boolean;
-  pipeline?: LayerPipelineSchema | null;
-  /** Session overrides for execution v0 (session-only runtime knobs). */
-  session?: ExecutionV0SessionDefaultsSchema;
-  /** Atomic execution graph patch for execution v0. */
-  executionPatch?: ExecutionV0PatchSchema;
-  /**
-   * Kernel path for explicit kernel dispatch ordering.
-   * Specifies exactly which kernels run, in what order, with what configs.
-   * Can be a preset ID (e.g., 'gemma2-q4k-fused-f32a') or inline KernelPathSchema.
-   */
-  kernelPath?: KernelPathRef;
-  /**
-   * Internal kernel-path source hint used by execution-v0 runtime patching.
-   * Not intended for direct user configuration.
-   */
-  kernelPathSource?: 'runtime' | 'config' | 'model' | 'manifest' | 'execution-v0' | 'none';
-  /**
-   * Kernel-path auto-selection policy.
-   * locked: preserve configured kernelPath as-is.
-   * capability-aware: allow rule-driven remap for selected sources.
-   */
-  kernelPathPolicy?: KernelPathPolicySchema;
-  /**
-   * Chat template override for runtime config.
-   * When set, overrides the model preset's chatTemplate.enabled setting.
-   */
-  chatTemplate?: ChatTemplateSchema;
-  /**
-   * Model-specific inference overrides.
-   * Allows runtime override of any manifest inference field (attention, normalization, rope, etc.).
-   * These are merged with manifest.inference via mergeConfig().
-   */
-  modelOverrides?: ModelInferenceOverrides;
-}
-
-export interface KernelPathPolicySchema {
-  mode: 'locked' | 'capability-aware';
-  sourceScope: Array<'model' | 'manifest' | 'config' | 'runtime' | 'execution-v0'>;
-  /**
-   * Backward-compatible alias for sourceScope.
-   * Prefer sourceScope in new configs.
-   */
-  allowSources?: Array<'model' | 'manifest' | 'config' | 'runtime' | 'execution-v0'>;
-  onIncompatible: 'error' | 'remap';
-}
-
-/** Default inference configuration */
-export declare const DEFAULT_INFERENCE_DEFAULTS_CONFIG: InferenceDefaultsConfigSchema;
-export declare const DEFAULT_KERNEL_PATH_POLICY: KernelPathPolicySchema;
-
-/** Default inference configuration for model presets */
-export declare const DEFAULT_PRESET_INFERENCE_CONFIG: InferenceConfigSchema;

package/src/config/schema/inference-defaults.schema.js

@@ -1,188 +0,0 @@
-import { DEFAULT_KVCACHE_CONFIG } from './kvcache.schema.js';
-import { DEFAULT_MOE_RUNTIME_CONFIG } from './moe.schema.js';
-import { DEFAULT_SPECULATIVE_CONFIG } from './speculative.schema.js';
-import { DEFAULT_RMS_NORM_EPS } from './manifest.schema.js';
-import { DEFAULT_DIFFUSION_CONFIG } from './diffusion.schema.js';
-import { DEFAULT_ENERGY_CONFIG } from './energy.schema.js';
-import {
-  DEFAULT_EXECUTION_V0_PATCH,
-  DEFAULT_EXECUTION_V0_SESSION_DEFAULTS,
-} from './execution-v0.schema.js';
-
-// =============================================================================
-// Generation Defaults (generator.js runtime options)
-// =============================================================================
-
-export const DEFAULT_GENERATION_CONFIG = {
-  useSpeculative: false,
-  profile: false,
-  benchmark: false,
-  disableCommandBatching: false,
-  disableMultiTokenDecode: false,
-  embeddingMode: 'last',
-};
-
-// =============================================================================
-// Batching Defaults
-// =============================================================================
-
-export const DEFAULT_BATCHING_DEFAULTS = {
-  batchSize: 4,  // GPU run-ahead window: record N tokens, submit once, readback once
-  maxTokens: 256,
-  stopCheckMode: 'batch',
-  readbackInterval: 1,
-  ringTokens: 1,
-  ringStop: 1,
-  ringStaging: 1,
-};
-
-// =============================================================================
-// Compute Defaults
-// =============================================================================
-
-export const DEFAULT_COMPUTE_DEFAULTS = {
-  activationDtype: 'f16',  // Default to F16 for web inference; fallback to F32 when unsupported
-  largeModelParamThreshold: 4e9,  // 4B parameters
-  paramEstimationMultiplier: 12,  // Rough approximation: 12 * hidden^2 * layers
-  keepF32Weights: false,  // Skip weight downcast (debug/compat)
-  rangeAwareSelectiveWidening: {
-    enabled: true,
-    absThreshold: 65500,
-    includeNonFinite: true,
-  },
-  deferredRoundingWindowTokens: 1,
-};
-
-// =============================================================================
-// Large Weight Handling
-// =============================================================================
-
-export const DEFAULT_LARGE_WEIGHT_CONFIG = {
-  enabled: true,
-  safetyRatio: 0.9,
-  preferF16: true,
-  lmHeadChunkRows: null,
-};
-
-// =============================================================================
-// Sampling Defaults
-// =============================================================================
-
-export const DEFAULT_SAMPLING_DEFAULTS = {
-  temperature: 1.0,
-  topP: 0.95,
-  topK: 50,
-  repetitionPenalty: 1.1,
-  greedyThreshold: 0.01,
-  repetitionPenaltyWindow: 100,
-};
-
-// =============================================================================
-// Tokenizer Defaults
-// =============================================================================
-
-export const DEFAULT_TOKENIZER_DEFAULTS = {
-  addBosToken: true,
-  addEosToken: false,
-  deferSpecialTokens: true,
-  addSpacePrefix: null,  // null = auto-detect from tokenizer.json
-};
-
-// =============================================================================
-// Chat Template Defaults
-// =============================================================================
-
-export const DEFAULT_CHAT_TEMPLATE_CONFIG = {
-  enabled: false,
-};
-
-export const DEFAULT_KERNEL_PATH_POLICY = {
-  mode: 'locked',
-  sourceScope: ['model', 'manifest'],
-  onIncompatible: 'error',
-};
-
-// =============================================================================
-// Complete Inference Defaults Config
-// =============================================================================
-
-export const DEFAULT_INFERENCE_DEFAULTS_CONFIG = {
-  batching: DEFAULT_BATCHING_DEFAULTS,
-  sampling: DEFAULT_SAMPLING_DEFAULTS,
-  compute: DEFAULT_COMPUTE_DEFAULTS,
-  tokenizer: DEFAULT_TOKENIZER_DEFAULTS,
-  largeWeights: DEFAULT_LARGE_WEIGHT_CONFIG,
-  kvcache: DEFAULT_KVCACHE_CONFIG,
-  moe: DEFAULT_MOE_RUNTIME_CONFIG,
-  speculative: DEFAULT_SPECULATIVE_CONFIG,
-  generation: DEFAULT_GENERATION_CONFIG,
-  chatTemplate: DEFAULT_CHAT_TEMPLATE_CONFIG,
-  diffusion: DEFAULT_DIFFUSION_CONFIG,
-  energy: DEFAULT_ENERGY_CONFIG,
-  pipeline: null,
-  kernelPath: undefined,
-  kernelPathSource: undefined,
-  kernelPathPolicy: DEFAULT_KERNEL_PATH_POLICY,
-  session: DEFAULT_EXECUTION_V0_SESSION_DEFAULTS,
-  executionPatch: DEFAULT_EXECUTION_V0_PATCH,
-};
-
-// =============================================================================
-// Preset Inference Defaults
-// =============================================================================
-
-export const DEFAULT_PRESET_INFERENCE_CONFIG = {
-  attention: {
-    slidingWindow: null,
-    attnLogitSoftcapping: null,
-    queryKeyNorm: false,
-    attentionOutputGate: false,
-    causal: true,
-    ropeScalingType: null,
-    ropeScalingFactor: 1.0,
-  },
-  normalization: {
-    rmsNormWeightOffset: false,
-    rmsNormEps: DEFAULT_RMS_NORM_EPS,
-    postAttentionNorm: false,
-    preFeedforwardNorm: false,
-    postFeedforwardNorm: false,
-  },
-  ffn: {
-    activation: 'silu',
-    gatedActivation: true,
-  },
-  output: {
-    finalLogitSoftcapping: null,
-    tieWordEmbeddings: false,
-    scaleEmbeddings: false,
-    embeddingTranspose: false,
-    embeddingVocabSize: null,
-  },
-  layerPattern: {
-    type: 'all_attention',
-  },
-  rope: {
-    ropeTheta: 10000,
-    ropeLocalTheta: null,
-    mropeInterleaved: false,
-    mropeSection: null,
-    partialRotaryFactor: null,
-    ropeScalingType: null,
-    ropeScalingFactor: 1.0,
-    ropeLocalScalingType: null,
-    ropeLocalScalingFactor: 1.0,
-    yarnBetaFast: null,
-    yarnBetaSlow: null,
-    yarnOriginalMaxPos: null,
-    ropeLocalYarnBetaFast: null,
-    ropeLocalYarnBetaSlow: null,
-    ropeLocalYarnOriginalMaxPos: null,
-  },
-  moe: null,
-  pipeline: null,
-  chatTemplate: {
-    type: null,
-  },
-  kernelPath: undefined,
-};