@simulatte/doppler 0.1.6 → 0.1.8

package/src/config/presets/runtime/kernels/dequant-f16-q4k.json

@@ -1,20 +1,13 @@
 {
   "id": "kernels/dequant-f16-q4k",
   "name": "dequant-f16-q4k",
-  "description": "Runtime alias preset for Gemma 2 Q4K dequant-to-F16 kernel path.",
+  "description": "Deprecated alias for kernels/gemma2-q4k-dequant-f16a.",
   "intent": "investigate",
-  "stability": "canonical",
+  "stability": "deprecated",
   "owner": "doppler-core",
   "createdAtUtc": "2026-02-25T00:00:00Z",
-  "extends": "default",
-  "runtime": {
-    "inference": {
-      "kernelPath": "gemma2-q4k-dequant-f16a",
-      "kernelPathPolicy": {
-        "mode": "capability-aware",
-        "sourceScope": ["config", "model", "manifest", "execution-v0"],
-        "onIncompatible": "remap"
-      }
-    }
-  }
+  "deprecatedAtUtc": "2026-03-08T00:00:00Z",
+  "replacementId": "kernels/gemma2-q4k-dequant-f16a",
+  "extends": "kernels/gemma2-q4k-dequant-f16a",
+  "runtime": {}
 }

package/src/config/presets/runtime/kernels/dequant-f32-q4k.json

@@ -1,20 +1,13 @@
 {
   "id": "kernels/dequant-f32-q4k",
   "name": "dequant-f32-q4k",
-  "description": "Runtime alias preset for Gemma 2 Q4K dequant-to-F32 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/kernels/embeddinggemma-q4k-dequant-f32a.json

@@ -0,0 +1,37 @@
+{
+  "id": "kernels/embeddinggemma-q4k-dequant-f32a",
+  "name": "embeddinggemma-q4k-dequant-f32a",
+  "description": "Explicit runtime override preset for the EmbeddingGemma Q4K dequant-to-F32 kernel path.",
+  "intent": "investigate",
+  "stability": "canonical",
+  "owner": "doppler-core",
+  "createdAtUtc": "2026-03-09T00:00:00Z",
+  "extends": "default",
+  "runtime": {
+    "inference": {
+      "compute": {
+        "activationDtype": "f32"
+      },
+      "kvcache": {
+        "kvDtype": "f32"
+      },
+      "session": {
+        "compute": {
+          "defaults": {
+            "activationDtype": "f32",
+            "outputDtype": "f32"
+          }
+        },
+        "kvcache": {
+          "kvDtype": "f32"
+        }
+      },
+      "kernelPath": "embeddinggemma-q4k-dequant-f32a",
+      "kernelPathPolicy": {
+        "mode": "capability-aware",
+        "sourceScope": ["config", "model", "manifest", "execution-v0"],
+        "onIncompatible": "remap"
+      }
+    }
+  }
+}

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

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

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/model/qwen3-5-layer-probe.json

@@ -0,0 +1,52 @@
+{
+  "id": "model/qwen3-5-layer-probe",
+  "name": "qwen3-5-layer-probe",
+  "description": "Probe all 24 layer outputs in Qwen 3.5 to isolate where the hidden state distribution collapses.",
+  "intent": "investigate",
+  "stability": "canonical",
+  "owner": "doppler-core",
+  "createdAtUtc": "2026-03-13T00:00:00Z",
+  "extends": "modes/debug",
+  "runtime": {
+    "inference": {
+      "prompt": "What color is the sky on a clear day? Answer in one word.",
+      "batching": {
+        "maxTokens": 1
+      },
+      "sampling": {
+        "temperature": 0
+      }
+    },
+    "shared": {
+      "debug": {
+        "trace": {
+          "enabled": true,
+          "categories": ["attn", "ffn", "logits"],
+          "layers": null,
+          "maxDecodeSteps": 1
+        },
+        "probes": [
+          {
+            "id": "embed",
+            "stage": "embed_out",
+            "tokens": [-1],
+            "dims": [0, 1, 2, 3, 512, 513]
+          },
+          {
+            "id": "layer_out",
+            "stage": "layer_out",
+            "layers": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23],
+            "tokens": [-1],
+            "dims": [0, 1, 2, 3]
+          },
+          {
+            "id": "logits",
+            "stage": "logits_final",
+            "tokens": [-1],
+            "dims": [271, 0, 1, 2, 3, 496, 138]
+          }
+        ]
+      }
+    }
+  }
+}

package/src/config/presets/runtime/model/qwen3-5-linear-attn-debug.json

@@ -0,0 +1,90 @@
+{
+  "id": "model/qwen3-5-linear-attn-debug",
+  "name": "qwen3-5-linear-attn-debug",
+  "description": "Probe linear attention intermediates in Qwen 3.5 layer 0 for comparison with HF reference.",
+  "intent": "investigate",
+  "stability": "canonical",
+  "owner": "doppler-core",
+  "createdAtUtc": "2026-03-13T00:00:00Z",
+  "extends": "modes/debug",
+  "runtime": {
+    "inference": {
+      "prompt": "Hello",
+      "batching": {
+        "maxTokens": 1
+      },
+      "sampling": {
+        "temperature": 0
+      },
+      "chatTemplate": {
+        "enabled": false
+      }
+    },
+    "shared": {
+      "debug": {
+        "trace": {
+          "enabled": true,
+          "categories": ["attn", "logits"],
+          "layers": null,
+          "maxDecodeSteps": 1
+        },
+        "probes": [
+          {
+            "id": "embed",
+            "stage": "embed_out",
+            "tokens": [-1],
+            "dims": [0, 1, 2, 3, 4, 5, 6, 7]
+          },
+          {
+            "id": "qkv",
+            "stage": "linear_qkv_proj",
+            "layers": [0],
+            "tokens": [-1],
+            "dims": [0, 1, 2, 3, 4, 5, 6, 7]
+          },
+          {
+            "id": "z",
+            "stage": "linear_z_proj",
+            "layers": [0],
+            "tokens": [-1],
+            "dims": [0, 1, 2, 3, 4, 5, 6, 7]
+          },
+          {
+            "id": "a",
+            "stage": "linear_a_proj",
+            "layers": [0],
+            "tokens": [-1],
+            "dims": [0, 1, 2, 3, 4, 5, 6, 7]
+          },
+          {
+            "id": "b",
+            "stage": "linear_b_proj",
+            "layers": [0],
+            "tokens": [-1],
+            "dims": [0, 1, 2, 3, 4, 5, 6, 7]
+          },
+          {
+            "id": "core",
+            "stage": "linear_core_out",
+            "layers": [0],
+            "tokens": [-1],
+            "dims": [0, 1, 2, 3, 4, 5, 6, 7]
+          },
+          {
+            "id": "layer0",
+            "stage": "layer_out",
+            "layers": [0],
+            "tokens": [-1],
+            "dims": [0, 1, 2, 3, 4, 5, 6, 7]
+          },
+          {
+            "id": "logits",
+            "stage": "logits_final",
+            "tokens": [-1],
+            "dims": [0, 1, 2, 3]
+          }
+        ]
+      }
+    }
+  }
+}

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/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/conversion.schema.d.ts

@@ -32,6 +32,7 @@ export interface ParsedModelSchema {
   config: RawModelConfigSchema;
   architecture?: string;
   quantization?: string;
+  generationConfig?: unknown;
   tokenizerJson?: unknown;
   tokenizerConfig?: unknown;
   tokenizerModel?: unknown;

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

@@ -172,9 +172,9 @@ export const DEFAULT_PRESET_INFERENCE_CONFIG = {
     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/kernel-path.schema.d.ts

@@ -151,13 +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;
 
@@ -162,7 +163,7 @@ export interface ArchitectureSchema {
  * Use `null` to indicate "not applicable" (e.g., no softcapping).
  */
 export interface ManifestAttentionSchema {
-  /** Query pre-attention scalar (Gemma 2: 256, standard: sqrt(headDim)) */
+  /** Query pre-attention scalar: attnScale = 1/sqrt(scalar). Standard = headDim. */
   queryPreAttnScalar: number;
   /** Attention logit softcapping (Gemma 2: 50, null = disabled) */
   attnLogitSoftcapping: number | null;