@simulatte/doppler 0.1.5 → 0.1.7

package/src/config/presets/runtime/kernels/gemma2-q4k-dequant-f16a.json

@@ -0,0 +1,33 @@
+{
+  "id": "kernels/gemma2-q4k-dequant-f16a",
+  "name": "gemma2-q4k-dequant-f16a",
+  "description": "Explicit runtime override preset for the Gemma 2 Q4K dequant-to-F16 kernel path.",
+  "intent": "investigate",
+  "stability": "canonical",
+  "owner": "doppler-core",
+  "createdAtUtc": "2026-03-08T00:00:00Z",
+  "extends": "default",
+  "runtime": {
+    "inference": {
+      "compute": {
+        "activationDtype": "f16"
+      },
+      "kvcache": {
+        "kvDtype": "f16"
+      },
+      "session": {
+        "compute": {
+          "defaults": {
+            "outputDtype": "f16"
+          }
+        }
+      },
+      "kernelPath": "gemma2-q4k-dequant-f16a",
+      "kernelPathPolicy": {
+        "mode": "capability-aware",
+        "sourceScope": ["config", "model", "manifest", "execution-v0"],
+        "onIncompatible": "remap"
+      }
+    }
+  }
+}

package/src/config/presets/runtime/kernels/gemma2-q4k-dequant-f32a-nosubgroups.json

@@ -0,0 +1,33 @@
+{
+  "id": "kernels/gemma2-q4k-dequant-f32a-nosubgroups",
+  "name": "gemma2-q4k-dequant-f32a-nosubgroups",
+  "description": "Explicit runtime override preset for the Gemma 2 Q4K dequant-to-F32 no-subgroups kernel path.",
+  "intent": "investigate",
+  "stability": "canonical",
+  "owner": "doppler-core",
+  "createdAtUtc": "2026-03-08T00:00:00Z",
+  "extends": "default",
+  "runtime": {
+    "inference": {
+      "compute": {
+        "activationDtype": "f32"
+      },
+      "kvcache": {
+        "kvDtype": "f32"
+      },
+      "session": {
+        "compute": {
+          "defaults": {
+            "outputDtype": "f32"
+          }
+        }
+      },
+      "kernelPath": "gemma2-q4k-dequant-f32a-nosubgroups",
+      "kernelPathPolicy": {
+        "mode": "capability-aware",
+        "sourceScope": ["config", "model", "manifest", "execution-v0"],
+        "onIncompatible": "remap"
+      }
+    }
+  }
+}

package/src/config/presets/runtime/kernels/gemma2-q4k-fused-f32a.json

@@ -0,0 +1,33 @@
+{
+  "id": "kernels/gemma2-q4k-fused-f32a",
+  "name": "gemma2-q4k-fused-f32a",
+  "description": "Explicit runtime override preset for the Gemma 2 fused Q4K F32 kernel path.",
+  "intent": "investigate",
+  "stability": "canonical",
+  "owner": "doppler-core",
+  "createdAtUtc": "2026-03-08T00:00:00Z",
+  "extends": "default",
+  "runtime": {
+    "inference": {
+      "compute": {
+        "activationDtype": "f32"
+      },
+      "kvcache": {
+        "kvDtype": "f32"
+      },
+      "session": {
+        "compute": {
+          "defaults": {
+            "outputDtype": "f32"
+          }
+        }
+      },
+      "kernelPath": "gemma2-q4k-fused-f32a",
+      "kernelPathPolicy": {
+        "mode": "capability-aware",
+        "sourceScope": ["config", "model", "manifest", "execution-v0"],
+        "onIncompatible": "remap"
+      }
+    }
+  }
+}

package/src/config/presets/runtime/kernels/safe-q4k.json

@@ -1,20 +1,13 @@
 {
   "id": "kernels/safe-q4k",
   "name": "safe-q4k",
-  "description": "Runtime alias preset for Gemma 2 Q4K dequant-to-F32 compatibility kernel path.",
+  "description": "Deprecated alias for kernels/gemma2-q4k-dequant-f32a-nosubgroups.",
   "intent": "investigate",
-  "stability": "canonical",
+  "stability": "deprecated",
   "owner": "doppler-core",
   "createdAtUtc": "2026-02-25T00:00:00Z",
-  "extends": "default",
-  "runtime": {
-    "inference": {
-      "kernelPath": "gemma2-q4k-dequant-f32a",
-      "kernelPathPolicy": {
-        "mode": "capability-aware",
-        "sourceScope": ["config", "model", "manifest", "execution-v0"],
-        "onIncompatible": "remap"
-      }
-    }
-  }
+  "deprecatedAtUtc": "2026-03-08T00:00:00Z",
+  "replacementId": "kernels/gemma2-q4k-dequant-f32a-nosubgroups",
+  "extends": "kernels/gemma2-q4k-dequant-f32a-nosubgroups",
+  "runtime": {}
 }

package/src/config/presets/runtime/platform/metal-apple-q4k.json

@@ -6,6 +6,6 @@
   "stability": "canonical",
   "owner": "doppler-core",
   "createdAtUtc": "2026-02-25T00:00:00Z",
-  "extends": "kernels/dequant-f16-q4k",
+  "extends": "kernels/gemma2-q4k-dequant-f16a",
   "runtime": {}
 }

package/src/config/required-inference-fields-contract-check.js

@@ -50,6 +50,9 @@ function createValidInferenceFixture() {
       ropeScalingFactor: 1.0,
       ropeScalingType: null,
       ropeLocalTheta: null,
+      mropeInterleaved: false,
+      mropeSection: null,
+      partialRotaryFactor: null,
       yarnBetaFast: null,
       yarnBetaSlow: null,
       yarnOriginalMaxPos: null,
@@ -94,6 +97,9 @@ const FIELD_CASES = Object.freeze([
   { kind: 'nonNullable', path: ['rope', 'ropeScalingFactor'], message: 'rope.ropeScalingFactor is required' },
   { kind: 'nullable', path: ['rope', 'ropeScalingType'], message: 'rope.ropeScalingType must be explicitly set' },
   { kind: 'nullable', path: ['rope', 'ropeLocalTheta'], message: 'rope.ropeLocalTheta must be explicitly set' },
+  { kind: 'nonNullable', path: ['rope', 'mropeInterleaved'], message: 'rope.mropeInterleaved is required' },
+  { kind: 'nullable', path: ['rope', 'mropeSection'], message: 'rope.mropeSection must be explicitly set' },
+  { kind: 'nullable', path: ['rope', 'partialRotaryFactor'], message: 'rope.partialRotaryFactor must be explicitly set' },
   { kind: 'nullable', path: ['rope', 'yarnBetaFast'], message: 'rope.yarnBetaFast must be explicitly set' },
   { kind: 'nullable', path: ['rope', 'yarnBetaSlow'], message: 'rope.yarnBetaSlow must be explicitly set' },
   { kind: 'nullable', path: ['rope', 'yarnOriginalMaxPos'], message: 'rope.yarnOriginalMaxPos must be explicitly set' },

package/src/config/runtime.js

@@ -1,5 +1,6 @@
 import { createDopplerConfig, setKernelThresholds } from './schema/index.js';
 import { validateRuntimeConfig, validateRuntimeOverrides } from './param-validator.js';
+import { isPlainObject } from '../utils/plain-object.js';
 
 let runtimeConfig = createDopplerConfig().runtime;
 setKernelThresholds(runtimeConfig.shared.kernelThresholds);
@@ -9,12 +10,16 @@ export function getRuntimeConfig() {
 }
 
 export function setRuntimeConfig(overrides) {
-  if (!overrides) {
+  if (overrides === undefined || overrides === null) {
     runtimeConfig = createDopplerConfig().runtime;
     setKernelThresholds(runtimeConfig.shared.kernelThresholds);
     return runtimeConfig;
   }
 
+  if (!isPlainObject(overrides)) {
+    throw new Error('DopplerConfigError: runtime overrides must be an object when provided.');
+  }
+
   assertNoDeprecatedRuntimeKeys(overrides);
   validateRuntimeOverrides(overrides);
 

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

@@ -181,6 +181,11 @@ export type ProbeStage =
   // Attention stages (per-layer)
   | 'attn_input'      // Input to attention (after residual from previous layer)
   | 'attn_normed'     // After input RMSNorm
+  | 'linear_qkv_proj' // Linear-attention fused QKV projection output
+  | 'linear_z_proj'   // Linear-attention z projection output
+  | 'linear_a_proj'   // Linear-attention a projection output
+  | 'linear_b_proj'   // Linear-attention b projection output
+  | 'linear_core_out' // Linear-attention recurrent core output (before o_proj)
   | 'q_proj'          // Q projection output
   | 'k_proj'          // K projection output
   | 'v_proj'          // V projection output

package/src/config/schema/doppler.schema.js

@@ -5,6 +5,7 @@ import { DEFAULT_EMULATION_CONFIG, createEmulationConfig } from './emulation.sch
 import { mergeEcosystemConfig } from './ecosystem.schema.js';
 import {
   chooseNullish,
+  chooseDefined,
   mergeExecutionPatchLists,
   mergeKernelPathPolicy,
   mergeShallowObject,
@@ -31,6 +32,13 @@ export const DEFAULT_DOPPLER_CONFIG = {
   runtime: DEFAULT_RUNTIME_CONFIG,
 };
 
+function cloneConfigTree(value) {
+  if (typeof structuredClone === 'function') {
+    return structuredClone(value);
+  }
+  return JSON.parse(JSON.stringify(value));
+}
+
 // =============================================================================
 // Factory Function
 // =============================================================================
@@ -39,19 +47,21 @@ export function createDopplerConfig(
   overrides
 ) {
   if (!overrides) {
-    return { ...DEFAULT_DOPPLER_CONFIG };
+    return {
+      model: DEFAULT_DOPPLER_CONFIG.model,
+      runtime: cloneConfigTree(DEFAULT_RUNTIME_CONFIG),
+    };
   }
 
   const runtimeOverrides = overrides.runtime ?? {};
+  const runtimeBase = cloneConfigTree(DEFAULT_RUNTIME_CONFIG);
   const runtime = overrides.runtime
-    ? mergeRuntimeConfig(DEFAULT_RUNTIME_CONFIG, runtimeOverrides)
-    : { ...DEFAULT_RUNTIME_CONFIG };
+    ? mergeRuntimeConfig(runtimeBase, runtimeOverrides)
+    : runtimeBase;
   const config = {
     model: overrides.model ?? DEFAULT_DOPPLER_CONFIG.model,
     runtime,
   };
-
-  applyCalibrateDefaults(config.runtime, runtimeOverrides);
   return config;
 }
 
@@ -239,7 +249,7 @@ function mergeInferenceConfig(
     speculative: { ...base.speculative, ...overrides.speculative },
     generation: { ...base.generation, ...overrides.generation },
     pipeline: overrides.pipeline ?? base.pipeline,
-    kernelPath: overrides.kernelPath ?? base.kernelPath,
+    kernelPath: chooseDefined(overrides.kernelPath, base.kernelPath),
     kernelPathSource: overrides.kernelPathSource ?? base.kernelPathSource,
     kernelPathPolicy: mergeKernelPathPolicy(baseKernelPathPolicy, overrideKernelPathPolicy),
     chatTemplate: mergeShallowObject(base.chatTemplate, overrides.chatTemplate),
@@ -302,21 +312,6 @@ function mergeDebugConfig(
   };
 }
 
-function applyCalibrateDefaults(runtime, runtimeOverrides) {
-  const intent = runtime?.shared?.tooling?.intent;
-  if (intent !== 'calibrate') return;
-
-  const warmupOverrides = runtimeOverrides?.shared?.kernelWarmup;
-  const hasPrewarmOverride = warmupOverrides
-    && Object.prototype.hasOwnProperty.call(warmupOverrides, 'prewarm');
-  if (!hasPrewarmOverride) {
-    runtime.shared.kernelWarmup = {
-      ...runtime.shared.kernelWarmup,
-      prewarm: true,
-    };
-  }
-}
-
 function mergeBenchmarkConfig(
   base,
   overrides

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

@@ -165,13 +165,16 @@ export const DEFAULT_PRESET_INFERENCE_CONFIG = {
   rope: {
     ropeTheta: 10000,
     ropeLocalTheta: null,
+    mropeInterleaved: false,
+    mropeSection: null,
+    partialRotaryFactor: null,
     ropeScalingType: null,
     ropeScalingFactor: 1.0,
     ropeLocalScalingType: null,
     ropeLocalScalingFactor: 1.0,
-    yarnBetaFast: 32,
-    yarnBetaSlow: 1,
-    yarnOriginalMaxPos: 4096,
+    yarnBetaFast: null,
+    yarnBetaSlow: null,
+    yarnOriginalMaxPos: null,
     ropeLocalYarnBetaFast: null,
     ropeLocalYarnBetaSlow: null,
     ropeLocalYarnOriginalMaxPos: null,

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

@@ -18,6 +18,15 @@ export interface RoPEConfigSchema {
   /** 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;
 

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

@@ -105,6 +105,12 @@ export interface KernelPathSchema {
   /** 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.
@@ -145,13 +151,17 @@ export interface KernelPathSchema {
 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-f32a' // Gemma 3 Q4K dequant path with 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

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

@@ -155,18 +155,25 @@ export const DEFAULT_KERNEL_THRESHOLDS = {
   tuner: DEFAULT_TUNER_LIMITS,
 };
 
+function cloneThresholdTree(value) {
+  if (typeof structuredClone === 'function') {
+    return structuredClone(value);
+  }
+  return JSON.parse(JSON.stringify(value));
+}
+
 // =============================================================================
 // Runtime Access
 // =============================================================================
 
-let currentThresholds = { ...DEFAULT_KERNEL_THRESHOLDS };
+let currentThresholds = cloneThresholdTree(DEFAULT_KERNEL_THRESHOLDS);
 
 export function getKernelThresholds() {
-  return currentThresholds;
+  return cloneThresholdTree(currentThresholds);
 }
 
 export function setKernelThresholds(overrides) {
-  currentThresholds = {
+  const nextThresholds = {
     ...currentThresholds,
     ...overrides,
     matmul: { ...currentThresholds.matmul, ...overrides.matmul },
@@ -180,8 +187,9 @@ export function setKernelThresholds(overrides) {
     cast: { ...currentThresholds.cast, ...overrides.cast },
     tuner: { ...currentThresholds.tuner, ...overrides.tuner },
   };
+  currentThresholds = cloneThresholdTree(nextThresholds);
 }
 
 export function resetKernelThresholds() {
-  currentThresholds = { ...DEFAULT_KERNEL_THRESHOLDS };
+  currentThresholds = cloneThresholdTree(DEFAULT_KERNEL_THRESHOLDS);
 }

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

@@ -88,7 +88,8 @@ export interface QuantizationInfoSchema {
   tts?: QuantizationValue;         // TTS decoder
   projector?: QuantizationValue;   // Cross-modal projection layers
 
-  // Runtime hints (NOT included in variantTag - these are runtime, not storage)
+  // Runtime hints. `compute` may be included in variantTag when artifact naming
+  // treats activation dtype as part of the published variant identity.
   kvCache?: QuantizationValue;
   compute?: QuantizationValue;
 
@@ -217,6 +218,12 @@ export interface ManifestRoPESchema {
   ropeTheta: number;
   /** Local theta for sliding window layers (null = same as ropeTheta) */
   ropeLocalTheta: number | null;
+  /** Use 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 (null = no scaling, 'linear', 'dynamic', 'yarn') */
   ropeScalingType: string | null;
   /** RoPE scaling factor (1.0 if no scaling) */

package/src/config/schema/manifest.schema.js

@@ -1,4 +1,5 @@
 import { MB } from './units.schema.js';
+import { validateRequiredInferenceFields } from '../../inference/pipelines/text/config.js';
 
 // =============================================================================
 // Hash & Versioning
@@ -62,14 +63,17 @@ export const DEFAULT_MANIFEST_INFERENCE = {
   rope: {
     ropeTheta: 10000,
     ropeLocalTheta: null,  // Same as ropeTheta (null = use ropeTheta)
+    mropeInterleaved: false,
+    mropeSection: null,
+    partialRotaryFactor: null,
     ropeScalingType: null,  // No scaling (null = disabled)
     ropeScalingFactor: 1.0,
     ropeLocalScalingType: null,  // Local scaling policy (null = no scaling)
     ropeLocalScalingFactor: 1.0,
     // YARN parameters - only relevant when ropeScalingType='yarn'
-    yarnBetaFast: 32,
-    yarnBetaSlow: 1,
-    yarnOriginalMaxPos: 4096,
+    yarnBetaFast: null,
+    yarnBetaSlow: null,
+    yarnOriginalMaxPos: null,
     // Local YARN parameters - only relevant when ropeLocalScalingType='yarn'
     ropeLocalYarnBetaFast: null,
     ropeLocalYarnBetaSlow: null,
@@ -121,6 +125,18 @@ export function validateManifestInference(
       `Please re-convert the model using the latest converter.`
     );
   }
+
+  if (manifest.modelType === 'diffusion' || manifest.modelType === 'energy') {
+    return;
+  }
+
+  const inference = typeof structuredClone === 'function'
+    ? structuredClone(manifest.inference)
+    : JSON.parse(JSON.stringify(manifest.inference));
+  validateRequiredInferenceFields(
+    inference,
+    manifest.modelId ?? 'unknown'
+  );
 }
 
 export function hasInferenceConfig(

package/src/config/training-defaults.js

@@ -2,9 +2,17 @@ import { createDopplerConfig, DEFAULT_TRAINING_SETTINGS } from './schema/index.j
 import { validateDistillTrainingConfig } from './schema/distill-training.schema.js';
 import { validateUlTrainingConfig } from './schema/ul-training.schema.js';
 
+function cloneConfigTree(value) {
+  if (typeof structuredClone === 'function') {
+    return structuredClone(value);
+  }
+  return JSON.parse(JSON.stringify(value));
+}
+
 function mergeTrainingSettings(base, overrides) {
+  const baseConfig = cloneConfigTree(base);
   if (!overrides) {
-    const merged = { ...base };
+    const merged = baseConfig;
     validateDistillTrainingConfig(merged.distill);
     validateUlTrainingConfig(merged.ul);
     if (merged.distill.enabled === true && merged.ul.enabled === true) {
@@ -14,42 +22,42 @@ function mergeTrainingSettings(base, overrides) {
   }
 
   const merged = {
-    enabled: overrides.enabled ?? base.enabled,
-    lora: { ...base.lora, ...overrides.lora },
+    enabled: overrides.enabled ?? baseConfig.enabled,
+    lora: { ...baseConfig.lora, ...overrides.lora },
     optimizer: {
-      ...base.optimizer,
+      ...baseConfig.optimizer,
       ...overrides.optimizer,
-      scheduler: { ...base.optimizer.scheduler, ...overrides.optimizer?.scheduler },
+      scheduler: { ...baseConfig.optimizer.scheduler, ...overrides.optimizer?.scheduler },
     },
-    gradient: { ...base.gradient, ...overrides.gradient },
-    precision: { ...base.precision, ...overrides.precision },
-    attention: { ...base.attention, ...overrides.attention },
+    gradient: { ...baseConfig.gradient, ...overrides.gradient },
+    precision: { ...baseConfig.precision, ...overrides.precision },
+    attention: { ...baseConfig.attention, ...overrides.attention },
     telemetry: {
-      ...base.telemetry,
+      ...baseConfig.telemetry,
       ...overrides.telemetry,
       alerts: {
-        ...base.telemetry.alerts,
+        ...baseConfig.telemetry.alerts,
         ...overrides.telemetry?.alerts,
         thresholds: {
-          ...base.telemetry.alerts.thresholds,
+          ...baseConfig.telemetry.alerts.thresholds,
           ...overrides.telemetry?.alerts?.thresholds,
         },
       },
     },
-    lossScaling: { ...base.lossScaling, ...overrides.lossScaling },
+    lossScaling: { ...baseConfig.lossScaling, ...overrides.lossScaling },
     distill: {
-      ...base.distill,
+      ...baseConfig.distill,
       ...overrides.distill,
-      freeze: { ...base.distill.freeze, ...overrides.distill?.freeze },
+      freeze: { ...baseConfig.distill.freeze, ...overrides.distill?.freeze },
     },
     ul: {
-      ...base.ul,
+      ...baseConfig.ul,
       ...overrides.ul,
-      noiseSchedule: { ...base.ul.noiseSchedule, ...overrides.ul?.noiseSchedule },
-      priorAlignment: { ...base.ul.priorAlignment, ...overrides.ul?.priorAlignment },
-      decoderSigmoidWeight: { ...base.ul.decoderSigmoidWeight, ...overrides.ul?.decoderSigmoidWeight },
-      lossWeights: { ...base.ul.lossWeights, ...overrides.ul?.lossWeights },
-      freeze: { ...base.ul.freeze, ...overrides.ul?.freeze },
+      noiseSchedule: { ...baseConfig.ul.noiseSchedule, ...overrides.ul?.noiseSchedule },
+      priorAlignment: { ...baseConfig.ul.priorAlignment, ...overrides.ul?.priorAlignment },
+      decoderSigmoidWeight: { ...baseConfig.ul.decoderSigmoidWeight, ...overrides.ul?.decoderSigmoidWeight },
+      lossWeights: { ...baseConfig.ul.lossWeights, ...overrides.ul?.lossWeights },
+      freeze: { ...baseConfig.ul.freeze, ...overrides.ul?.freeze },
     },
   };
   validateDistillTrainingConfig(merged.distill);
@@ -74,7 +82,7 @@ export function createTrainingConfig(overrides = {}) {
 
 export const DEFAULT_TRAINING_CONFIG = createTrainingConfig();
 
-let trainingConfig = DEFAULT_TRAINING_CONFIG;
+let trainingConfig = createTrainingConfig();
 
 export function getTrainingConfig() {
   return trainingConfig;
@@ -86,6 +94,6 @@ export function setTrainingConfig(overrides) {
 }
 
 export function resetTrainingConfig() {
-  trainingConfig = DEFAULT_TRAINING_CONFIG;
+  trainingConfig = createTrainingConfig();
   return trainingConfig;
 }