@simulatte/doppler 0.1.9 → 0.2.0

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

@@ -1,298 +0,0 @@
-/**
- * Inference Schema Definitions
- *
- * Configuration for model inference behavior.
- * These are runtime settings that affect how the model executes.
- *
- * @module config/schema/inference
- */
-
-import type { ProbeStage } from './debug.schema.js';
-import type { KernelPathRef } from './kernel-path.schema.js';
-
-/** RoPE configuration for positional embeddings */
-export interface RoPEConfigSchema {
-  /** Base frequency for RoPE (default 10000, modern models use 1000000) */
-  ropeTheta?: number;
-
-  /** Local RoPE theta for sliding window layers (Gemma 3 uses 10000) */
-  ropeLocalTheta?: number;
-
-  /** Apply adjacent-pair rotary layout instead of rotate-half layout. */
-  mropeInterleaved?: boolean;
-
-  /** mRoPE section sizes before the Qwen doubling step. */
-  mropeSection?: number[] | null;
-
-  /** Fraction of the head dimension that participates in rotary embedding. */
-  partialRotaryFactor?: number | null;
-
-  /** RoPE scaling type */
-  ropeScalingType?: 'linear' | 'dynamic' | 'yarn' | null;
-
-  /** RoPE scaling factor */
-  ropeScalingFactor?: number;
-
-  /** Local RoPE scaling type for sliding window layers */
-  ropeLocalScalingType?: 'linear' | 'dynamic' | 'yarn' | null;
-
-  /** Local RoPE scaling factor for sliding window layers */
-  ropeLocalScalingFactor?: number;
-
-  /** YARN beta_fast parameter */
-  yarnBetaFast?: number;
-
-  /** YARN beta_slow parameter */
-  yarnBetaSlow?: number;
-
-  /** YARN original max position embeddings */
-  yarnOriginalMaxPos?: number;
-
-  /** Local YARN beta_fast parameter */
-  ropeLocalYarnBetaFast?: number;
-
-  /** Local YARN beta_slow parameter */
-  ropeLocalYarnBetaSlow?: number;
-
-  /** Local YARN original max position embeddings */
-  ropeLocalYarnOriginalMaxPos?: number;
-}
-
-/** Attention mechanism configuration */
-export interface AttentionSchema {
-  /** Use sliding window attention */
-  slidingWindow?: number | null;
-  /** Softcap attention logits before softmax */
-  attnLogitSoftcapping?: number | null;
-  /** Use query-key normalization */
-  queryKeyNorm?: boolean;
-  /** Apply sigmoid gate from q_proj split to attention output */
-  attentionOutputGate?: boolean;
-  /** @deprecated Use RoPEConfigSchema.ropeScalingType instead */
-  ropeScalingType?: 'linear' | 'dynamic' | 'yarn' | null;
-  /** @deprecated Use RoPEConfigSchema.ropeScalingFactor instead */
-  ropeScalingFactor?: number;
-}
-
-/** Normalization configuration */
-export interface NormalizationSchema {
-  /** Add 1.0 to RMSNorm weights (Gemma-style) */
-  rmsNormWeightOffset?: boolean;
-  /** RMSNorm epsilon */
-  rmsNormEps?: number;
-  /** Use post-attention norm */
-  postAttentionNorm?: boolean;
-  /** Use pre-feedforward norm */
-  preFeedforwardNorm?: boolean;
-  /** Use post-feedforward norm */
-  postFeedforwardNorm?: boolean;
-}
-
-/** Feed-forward network configuration */
-export interface FFNSchema {
-  /** Activation function */
-  activation?: 'silu' | 'gelu' | 'relu' | 'swiglu';
-  /** Whether activation is gated (e.g., SwiGLU, GeGLU) */
-  gatedActivation?: boolean;
-  /** Clamp SwiGLU output (null = disabled) */
-  swigluLimit?: number | null;
-}
-
-/** Built-in chat template types */
-export type ChatTemplateType =
-  | 'gemma'
-  | 'llama3'
-  | 'gpt-oss'
-  | 'chatml'
-  | 'qwen'
-  | 'translategemma'
-  | null;
-
-/** Chat template configuration for instruct models */
-export interface ChatTemplateSchema {
-  /** Template type identifier (gemma, llama3, gpt-oss, chatml, qwen, translategemma) */
-  type?: ChatTemplateType;
-
-  /** Whether to apply chat template by default (instruct models should set true) */
-  enabled?: boolean;
-
-  /** Custom template with {prompt} placeholder (overrides type) */
-  custom?: string;
-}
-
-export type LayerPipelineOp =
-  | 'save'
-  | 'load'
-  | 'conv'
-  | 'attention'
-  | 'rmsnorm'
-  | 'ffn'
-  | 'residual_add'
-  | 'cast'
-  | 'noop';
-
-export type LayerPipelinePhase = 'prefill' | 'decode' | 'both';
-export type LayerPipelineDtype = 'f16' | 'f32';
-
-export type LayerPipelineNormWeight =
-  | 'input'
-  | 'post_attention'
-  | 'post_attn'
-  | 'pre_ffn'
-  | 'post_ffn';
-
-export interface LayerPipelineStepSchema {
-  op: LayerPipelineOp;
-  /** Optional phase gate for this step (default: both) */
-  phase?: LayerPipelinePhase;
-  /** Source slot (default: "state") */
-  src?: string;
-  /** Destination slot (default: "state") */
-  dst?: string;
-  /** Slot name for save/load operations */
-  name?: string;
-  /** Norm weight selector (rmsnorm only) */
-  weight?: LayerPipelineNormWeight;
-  /** Residual slot for fused ops (optional) */
-  residual?: string | null;
-  /** Residual add inputs (defaults: a="state", b="residual") */
-  a?: string;
-  b?: string;
-  /** FFN variant override */
-  variant?: 'auto' | 'dense' | 'moe';
-  /** Skip input norm inside attention (use when providing explicit rmsnorm) */
-  skipInputNorm?: boolean;
-  /** Optional probe stage to emit for this step */
-  probeStage?: ProbeStage;
-  /** Explicit input dtype requirement for this step */
-  inputDtype?: LayerPipelineDtype;
-  /** Explicit output dtype contract for this step */
-  outputDtype?: LayerPipelineDtype;
-  /** Cast source dtype (cast op only) */
-  fromDtype?: LayerPipelineDtype;
-  /** Cast target dtype (cast op only) */
-  toDtype?: LayerPipelineDtype;
-}
-
-export interface LayerPipelineOverrideSchema {
-  layers: number[];
-  steps: LayerPipelineStepSchema[];
-}
-
-export interface LayerPipelineSchema {
-  steps: LayerPipelineStepSchema[];
-  overrides?: LayerPipelineOverrideSchema[];
-}
-
-/** Output/sampling configuration */
-export interface OutputSchema {
-  /** Softcap final logits */
-  finalLogitSoftcapping?: number | null;
-  /** Tie embeddings to output */
-  tieWordEmbeddings?: boolean;
-}
-
-/** Layer type for hybrid models */
-export type LayerType =
-  | 'attention'
-  | 'full_attention'
-  | 'sliding_attention'
-  | 'linear_attention'
-  | 'conv'
-  | 'mamba'
-  | 'rwkv';
-
-/** Linear-attention RMS norm vector layout */
-export type LinearNormMode =
-  | 'shared'      // one [head_v_dim] vector shared across all value heads
-  | 'per_head';   // one [num_v_heads * head_v_dim] vector
-
-/** Global layer pattern (computed at runtime from numLayers) */
-export type GlobalLayerPattern =
-  | 'even'       // Layers 0, 2, 4, ... are global (Gemma 2)
-  | 'odd'        // Layers 1, 3, 5, ... are global
-  | 'every_n';   // Every Nth layer is global (Gemma 3: every 6th)
-
-/** Layer pattern for hybrid architectures */
-export interface LayerPatternSchema {
-  /** Pattern type: 'all_attention', 'alternating', 'every_n', 'custom' */
-  type: 'all_attention' | 'alternating' | 'every_n' | 'custom';
-  /** For 'alternating': pattern for global/full attention layers */
-  globalPattern?: GlobalLayerPattern;
-  /** For 'every_n': the period value (e.g., 6 for Gemma 3) */
-  period?: number;
-  /** For 'every_n': first global layer index modulo period (default: 0) */
-  offset?: number;
-  /** @deprecated Use globalPattern/period instead */
-  attentionLayers?: number[];
-  /** For 'custom': explicit layer type mapping */
-  layerTypes?: LayerType[];
-}
-
-/**
- * Compute global attention layer indices from pattern.
- * Used at runtime when numLayers is known.
- */
-export declare function computeGlobalLayers(
-  pattern: LayerPatternSchema,
-  numLayers: number
-): number[];
-
-/** Complete inference configuration */
-export interface InferenceConfigSchema {
-  attention?: AttentionSchema;
-  normalization?: NormalizationSchema;
-  ffn?: FFNSchema;
-  moe?: {
-    kernelProfileId?: string | null;
-    supportedActivationDtypes?: string[];
-    preferredActivationDtype?: string | null;
-    shapeConstraints?: {
-      hiddenSizeDivisor?: number;
-      intermediateSizeDivisor?: number;
-      groupSize?: number;
-    };
-  } | null;
-  output?: OutputSchema;
-  layerPattern?: LayerPatternSchema;
-  rope?: RoPEConfigSchema;
-  pipeline?: LayerPipelineSchema | null;
-  /** Chat template for instruct models */
-  chatTemplate?: ChatTemplateSchema;
-  /**
-   * 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;
-}
-
-/** Sampling parameters */
-export interface SamplingSchema {
-  temperature?: number;
-  topK?: number;
-  topP?: number;
-  repetitionPenalty?: number;
-}
-
-/** Tokenizer runtime configuration */
-export interface TokenizerConfigSchema {
-  /** BOS token string */
-  bosToken?: string;
-  /** EOS token strings (can be multiple) */
-  eosTokens?: string[];
-  /** Pad token string */
-  padToken?: string;
-  /** Add BOS token to input */
-  addBosToken?: boolean;
-  /** Add EOS token to output */
-  addEosToken?: boolean;
-  /** Defer special-token validation until tokenizer metadata is fully loaded */
-  deferSpecialTokens?: boolean;
-  /** HuggingFace model ID for tokenizer fallback */
-  hfModel?: string;
-  /** Allow architecture-based fallback when hfModel is missing */
-  allowArchFallback?: boolean;
-  /** Chat template (jinja2-style) */
-  chatTemplate?: string;
-}

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

@@ -1,39 +0,0 @@
-// =============================================================================
-// Layer Pattern Schema
-// =============================================================================
-
-export function computeGlobalLayers(
-  pattern,
-  numLayers
-) {
-  if (pattern.attentionLayers) {
-    // Legacy: use explicit array (filtered to valid range)
-    return pattern.attentionLayers.filter(i => i < numLayers);
-  }
-
-  const patternType = pattern.type ?? null;
-  const globalPattern = pattern.globalPattern ?? (patternType === 'every_n' ? 'every_n' : null);
-
-  if (!globalPattern) {
-    // Default: all layers are global
-    return Array.from({ length: numLayers }, (_, i) => i);
-  }
-
-  switch (globalPattern) {
-    case 'even':
-      return Array.from({ length: numLayers }, (_, i) => i).filter(i => i % 2 === 0);
-    case 'odd':
-      return Array.from({ length: numLayers }, (_, i) => i).filter(i => i % 2 === 1);
-    case 'every_n': {
-      const n = pattern.period;
-      if (!Number.isFinite(n) || n <= 0) {
-        throw new Error('layerPattern.period must be a positive number for every_n patterns.');
-      }
-      const rawOffset = Number.isFinite(pattern.offset) ? Math.trunc(pattern.offset) : 0;
-      const offset = ((rawOffset % n) + n) % n;
-      return Array.from({ length: numLayers }, (_, i) => i).filter(i => (((i - offset) % n + n) % n) === 0);
-    }
-    default:
-      return Array.from({ length: numLayers }, (_, i) => i);
-  }
-}

package/src/config/schema/intent-bundle.schema.d.ts

@@ -1,28 +0,0 @@
-/**
- * Intent Bundle Config Schema
- *
- * Runtime policy for intent bundle gating.
- *
- * @module config/schema/intent-bundle
- */
-
-/**
- * Intent bundle configuration.
- */
-export interface IntentBundleConfigSchema {
-  /** Enable intent bundle checks */
-  enabled: boolean;
-  /** Optional URL for fetching bundle in test harness workflows */
-  bundleUrl: string | null;
-  /** Loaded intent bundle (runtime-only) */
-  bundle: Record<string, unknown> | null;
-  /** Require base model hash to match the loaded manifest */
-  requireBaseModelHash: boolean;
-  /** Require kernel registry version match */
-  requireKernelRegistryVersion: boolean;
-  /** Require deterministic output hash if specified in bundle */
-  enforceDeterministicOutput: boolean;
-}
-
-/** Default intent bundle configuration */
-export declare const DEFAULT_INTENT_BUNDLE_CONFIG: IntentBundleConfigSchema;

package/src/config/schema/intent-bundle.schema.js

@@ -1,12 +0,0 @@
-// =============================================================================
-// Intent Bundle Config
-// =============================================================================
-
-export const DEFAULT_INTENT_BUNDLE_CONFIG = {
-  enabled: false,
-  bundleUrl: null,
-  bundle: null,
-  requireBaseModelHash: true,
-  requireKernelRegistryVersion: true,
-  enforceDeterministicOutput: false,
-};

package/src/config/schema/kernel-path.schema.d.ts

@@ -1,184 +0,0 @@
-/**
- * Kernel Path Schema
- *
- * Defines explicit, ordered kernel dispatch sequences for inference.
- * Replaces the implicit q4kStrategy/fusedFFNQ4K configuration.
- *
- * A kernel path is a complete specification of:
- * - Which kernels run
- * - In what order
- * - With what override constants
- * - With what entry points
- *
- * @module config/schema/kernel-path
- */
-
-/**
- * A single kernel dispatch in the path.
- */
-export interface KernelStepSchema {
-  /**
-   * Logical operation name (for debugging/tracing).
-   * Examples: 'rmsnorm', 'q_proj', 'attention', 'ffn_fused'
-   */
-  op: string;
-
-  /**
-   * Kernel file name (without path).
-   * Examples: 'rmsnorm.wgsl', 'fused_matmul_q4.wgsl'
-   */
-  kernel: string;
-
-  /**
-   * Entry point function name.
-   * @default 'main'
-   */
-  entry?: string;
-
-  /**
-   * Override constants for pipeline creation.
-   * These are compile-time constants that affect code generation.
-   */
-  constants?: Record<string, number | boolean>;
-
-  /**
-   * Weight buffer reference (for matmul ops).
-   * Uses template syntax: 'layer.{L}.self_attn.q_proj'
-   * {L} is replaced with layer index at runtime.
-   */
-  weights?: string;
-
-  /**
-   * Input buffer slot name.
-   * @default 'hidden_state'
-   */
-  input?: string;
-
-  /**
-   * Output buffer slot name.
-   * @default 'hidden_state'
-   */
-  output?: string;
-}
-
-/**
- * Kernel sequence for a single transformer layer.
- */
-export interface LayerKernelPathSchema {
-  /** Ordered list of kernel dispatches */
-  steps: KernelStepSchema[];
-}
-
-/**
- * Override for specific layers (e.g., first/last layer differences).
- */
-export interface LayerOverrideSchema {
-  /** Layer indices this override applies to */
-  layers: number[];
-
-  /** Steps to use instead of default */
-  steps: KernelStepSchema[];
-}
-
-/**
- * Complete kernel path specification for a model.
- */
-export interface KernelPathSchema {
-  /** Path identifier */
-  id: string;
-
-  /** Human-readable name */
-  name: string;
-
-  /** Description of this path's characteristics */
-  description?: string;
-
-  /** Activation dtype for this path (e.g., 'f16', 'f32') */
-  activationDtype: string;
-
-  /**
-   * Output dtype for kernels in this path (e.g., 'f16', 'f32').
-   * Defaults to `activationDtype` when omitted.
-   */
-  outputDtype?: string;
-
-  /** KV cache dtype for this path; defaults to activationDtype when omitted. */
-  kvDtype?: string;
-
-  /**
-   * Explicit widening target used by the finiteness fallback execution plan.
-   * Required for inline/generated kernel paths that do not have a stable registry id.
-   */
-  finitenessFallbackKernelPathId?: string;
-
-  /**
-   * Prefill phase kernel sequence (M > 1).
-   * If not specified, uses decode with batched variants.
-   */
-  prefill?: LayerKernelPathSchema;
-
-  /**
-   * Decode phase kernel sequence (M = 1).
-   */
-  decode: LayerKernelPathSchema;
-
-  /**
-   * Layer-specific overrides.
-   * For models with different first/last layer behavior.
-   */
-  layerOverrides?: LayerOverrideSchema[];
-
-  /**
-   * Pre-layer operations (embedding lookup, initial norm).
-   */
-  preLayer?: KernelStepSchema[];
-
-  /**
-   * Post-layer operations (final norm, LM head).
-   */
-  postLayer?: KernelStepSchema[];
-
-  /**
-   * Sampling kernels.
-   */
-  sampling?: KernelStepSchema[];
-}
-
-/**
- * Built-in kernel path identifiers.
- * These are the known preset IDs, but custom presets can also be registered.
- */
-export type BuiltinKernelPathId =
-  | 'gemma2-q4k-fused-f32a'   // Gemma 2 Q4K weights, fused matmul, F32 activations
-  | 'gemma2-q4k-dequant-f16a' // Gemma 2 Q4K -> F16 dequant, F16 activations
-  | 'gemma2-q4k-dequant-f32a-nosubgroups' // Gemma 2 Q4K -> F32 dequant path with no subgroup requirement
-  | 'gemma2-q4k-dequant-f32a' // Legacy alias for gemma2-q4k-dequant-f32a-nosubgroups
-  | 'gemma2-f16-f16a'         // Gemma 2 F16 weights, F16 activations
-  | 'gemma2-f16-f32a'         // Gemma 2 F16 weights, F32 activations
-  | 'gemma3-f16-fused-f16a-online' // Gemma 3 F16 fused FFN online path
-  | 'gemma3-f16-fused-f32a-online' // Gemma 3 F16 fused FFN online path with F32 activations
-  | 'gemma3-f16-fused-f32a-online-streamingprefill' // Gemma 3 F16 fused FFN online path with streaming prefill attention
-  | 'gemma3-q4k-dequant-f16a-online' // Gemma 3 Q4K dequant online path (F16 activations)
-  | 'gemma3-q4k-dequant-f32a-online' // Gemma 3 Q4K dequant online path with F32 activations
-  | 'gemma3-q4k-dequant-f32w-f32a-online' // Gemma 3 Q4K path with F32 projection weights and F32 activations
-  | 'gemma3-q4k-dequant-f32a-nosubgroups' // Gemma 3 Q4K dequant path with no subgroup requirement
-  | 'gemma3-q4k-dequant-f32a' // Legacy alias for gemma3-q4k-dequant-f32a-nosubgroups
-  | 'lfm2-q4k-dequant-f32a-online' // LFM2 Q4K path with F32 activations and fast prefill
-  | 'embeddinggemma-f16-f32a' // EmbeddingGemma F16 weights, F32 activations
-  | 'embeddinggemma-f32-f32a' // EmbeddingGemma F32 weights, F32 activations
-  | 'embeddinggemma-q4k-dequant-f32a'; // EmbeddingGemma Q4K dequant, F32 activations
-
-/**
- * Kernel path reference - preset ID (string) or inline path schema.
- * Accepts any string for custom preset IDs, not just built-in IDs.
- */
-export type KernelPathRef = string | KernelPathSchema;
-
-/** Default entry point */
-export declare const DEFAULT_ENTRY: string;
-
-/** Default input slot */
-export declare const DEFAULT_INPUT: string;
-
-/** Default output slot */
-export declare const DEFAULT_OUTPUT: string;

package/src/config/schema/kernel-path.schema.js

@@ -1,9 +0,0 @@
-// =============================================================================
-// Defaults
-// =============================================================================
-
-export const DEFAULT_ENTRY = 'main';
-
-export const DEFAULT_INPUT = 'hidden_state';
-
-export const DEFAULT_OUTPUT = 'hidden_state';