@simulatte/doppler 0.1.7 → 0.1.8

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/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/manifest.schema.d.ts

@@ -163,7 +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;

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

@@ -40,7 +40,7 @@ export const DEFAULT_MANIFEST_INFERENCE = {
   schema: null,
   presetId: null,
   attention: {
-    queryPreAttnScalar: 8,  // sqrt(64) for standard 64-dim heads
+    queryPreAttnScalar: 64, // headDim for standard 64-dim heads; attnScale = 1/sqrt(scalar)
     attnLogitSoftcapping: null,  // No softcapping (null = disabled)
     slidingWindow: null,  // Full attention (null = no sliding window)
     queryKeyNorm: false,

package/src/config/schema/storage.schema.js

@@ -35,7 +35,7 @@ export const DEFAULT_STORAGE_ALIGNMENT_CONFIG = {
 export const DEFAULT_STORAGE_BACKEND_CONFIG = {
   backend: 'auto', // auto | opfs | indexeddb | memory
   opfs: {
-    useSyncAccessHandle: true,
+    useSyncAccessHandle: false,
     maxConcurrentHandles: 2,
   },
   indexeddb: {

package/src/converter/conversion-plan.js

@@ -117,7 +117,10 @@ function isLikelyEmbeddingGemma(rawConfig, architectureHint) {
 
 export function inferSourceWeightQuantization(tensors) {
   if (!Array.isArray(tensors) || tensors.length === 0) {
-    return 'f16';
+    throw new Error(
+      'Cannot infer source weight quantization: no tensors provided. ' +
+      'Set converterConfig.quantization.weights explicitly.'
+    );
   }
   const weightTensors = [];
   for (const tensor of tensors) {
@@ -128,7 +131,12 @@ export function inferSourceWeightQuantization(tensors) {
     weightTensors.push({ name, dtype });
   }
   const dtypes = new Set(weightTensors.map((tensor) => tensor.dtype));
-  if (dtypes.size === 0) return 'f16';
+  if (dtypes.size === 0) {
+    throw new Error(
+      'Cannot infer source weight quantization: no recognizable weight dtypes found. ' +
+      'Set converterConfig.quantization.weights explicitly.'
+    );
+  }
   if (dtypes.size > 1) {
     const detail = Array.from(dtypes)
       .sort()

package/src/converter/core.js

@@ -983,6 +983,7 @@ export function createManifest(
     isDiffusion ? 'diffusion' : extractArchitecture(model.config, model.ggufConfig)
   );
   const rawConfig = model.config || {};
+  const generationConfig = model.generationConfig ?? null;
   const resolvedArchitecture = isDiffusion
     ? architecture
     : resolveIntermediateSizeFromTensors(architecture, model, tensorLocations, rawConfig, modelId);
@@ -1037,6 +1038,7 @@ export function createManifest(
       ? null
       : resolveEosTokenId({
           config: rawConfig,
+          generationConfig,
           tokenizer: model.tokenizer ?? model.tokenizerConfig ?? null,
           tokenizerJson: model.tokenizerJson ?? null,
         });

package/src/converter/manifest-inference.js

@@ -240,16 +240,6 @@ function detectAttentionOutputGate(presetInference, modelConfig, defaults) {
     return modelConfig.attn_output_gate;
   }
 
-  const modelType = normalizeLayerTypeName(modelConfig?.model_type);
-  const hasLinearAttentionLayers = Array.isArray(modelConfig?.layer_types)
-    && modelConfig.layer_types.some((entry) => normalizeCustomLayerType(entry) === 'linear_attention');
-  if (
-    hasLinearAttentionLayers
-    && (modelType === 'qwen2' || modelType === 'qwen3_5' || modelType === 'qwen3_5_text')
-  ) {
-    return true;
-  }
-
   return defaults.attention.attentionOutputGate;
 }
 
@@ -259,21 +249,18 @@ function resolveQueryPreAttnScalar(preset, modelConfig, headDim) {
     return explicit;
   }
 
-  const modelType = normalizeLayerTypeName(modelConfig?.model_type);
-  const presetId = normalizeLayerTypeName(preset?.id);
-  if (modelType.startsWith('qwen') || presetId === 'qwen3') {
-    return headDim;
+  // Standard attention scaling: attnScale = 1/sqrt(queryPreAttnScalar).
+  // For standard transformers queryPreAttnScalar = headDim, giving 1/sqrt(headDim).
+  // Preset may override for non-standard models.
+  const presetScalar = Number(preset?.inference?.attention?.queryPreAttnScalar);
+  if (Number.isFinite(presetScalar) && presetScalar > 0) {
+    return presetScalar;
   }
 
-  return Math.sqrt(headDim);
+  return headDim;
 }
 
 function detectRmsNormWeightOffset(presetInference, modelConfig, defaults) {
-  const modelType = normalizeLayerTypeName(modelConfig?.model_type);
-  if (modelType === 'qwen3_5' || modelType === 'qwen3_5_text') {
-    return true;
-  }
-
   if (typeof presetInference?.normalization?.rmsNormWeightOffset === 'boolean') {
     return presetInference.normalization.rmsNormWeightOffset;
   }
@@ -385,8 +372,8 @@ export function buildManifestInference(preset, config, headDim = 64, quantizatio
       queryPreAttnScalar: resolveQueryPreAttnScalar(preset, modelConfig, headDim),
       attnLogitSoftcapping: presetInference.attention?.attnLogitSoftcapping ??
         modelConfig.attn_logit_softcapping ?? defaults.attention.attnLogitSoftcapping,
-      slidingWindow: presetInference.attention?.slidingWindow ??
-        modelConfig.sliding_window ?? defaults.attention.slidingWindow,
+      slidingWindow: modelConfig.sliding_window ??
+        presetInference.attention?.slidingWindow ?? defaults.attention.slidingWindow,
       queryKeyNorm: presetInference.attention?.queryKeyNorm ?? defaults.attention.queryKeyNorm,
       attentionOutputGate: detectAttentionOutputGate(presetInference, modelConfig, defaults),
       causal: detectedCausalAttention ?? presetInference.attention?.causal ?? defaults.attention.causal,
@@ -459,6 +446,9 @@ export function buildManifestInference(preset, config, headDim = 64, quantizatio
         );
       }
       globalPattern = null;
+      // Default offset 0 means first global layer at index 0 (most common pattern).
+      // This is the every_n pattern default, distinct from layerPattern.offset=null
+      // which means "not applicable" in the schema.
       offset = (
         detectEveryNOffsetFromLayerTypes(modelConfig.layer_types, period)
         ?? normalizeEveryNOffset(presetPattern.offset, period)

package/src/converter/parsers/transformer.js

@@ -7,6 +7,9 @@ export async function parseTransformerModel(adapter) {
   } = adapter;
 
   const config = await readJson('config.json', 'config.json');
+  const generationConfig = await fileExists('generation_config.json')
+    ? await readJson('generation_config.json', 'generation_config.json')
+    : null;
   const architectureHint = config.architectures?.[0] ?? config.model_type ?? '';
 
   let tensors = null;
@@ -19,6 +22,7 @@ export async function parseTransformerModel(adapter) {
 
   return {
     config,
+    generationConfig,
     tensors,
     architectureHint,
   };

package/src/converter/quantization-info.js

@@ -2,6 +2,10 @@
 import { DEFAULT_QUANTIZATION_DEFAULTS, DEFAULT_Q4K_LAYOUT } from '../config/index.js';
 import { classifyTensorRole } from '../formats/rdrr/index.js';
 
+// Default quantization tag when no explicit dtype is provided.
+// F16 is the canonical unquantized storage format for WebGPU inference.
+const DEFAULT_QUANT_TAG = 'f16';
+
 // Quantization tag aliases mapped to canonical names.
 // Add new aliases here rather than adding if/else branches.
 const QUANT_TAG_ALIASES = {
@@ -47,7 +51,7 @@ const QUANT_TAG_ALIASES = {
 };
 
 export function normalizeQuantTag(value) {
-  if (!value) return 'f16';
+  if (!value) return DEFAULT_QUANT_TAG;
   const lower = value.toLowerCase();
   return QUANT_TAG_ALIASES[lower] ?? lower;
 }

package/src/converter/quantizer.js

@@ -74,9 +74,10 @@ function findMinMax(data, offset, length) {
   return { min, max };
 }
 
-export function quantizeQ4KBlock(data, offset) {
+function quantizeQ4KBlockWithValidLength(data, offset, validLength = QK_K) {
   const block = new Uint8Array(QK4_K_BLOCK_SIZE);
   const blockView = new DataView(block.buffer);
+  const clampedValidLength = Math.max(0, Math.min(QK_K, Math.trunc(validLength)));
 
   const scales = new Float32Array(8);
   const minOffsets = new Float32Array(8);
@@ -84,14 +85,22 @@ export function quantizeQ4KBlock(data, offset) {
 
   for (let sb = 0; sb < 8; sb++) {
     const sbOffset = offset + sb * 32;
-    const { min, max } = findMinMax(data, sbOffset, 32);
+    const subblockStart = sb * 32;
+    const validInSubblock = Math.max(0, Math.min(32, clampedValidLength - subblockStart));
+    if (validInSubblock === 0) {
+      scales[sb] = 0;
+      minOffsets[sb] = 0;
+      continue;
+    }
+
+    const { min, max } = findMinMax(data, sbOffset, validInSubblock);
 
     minOffsets[sb] = -min;
     const range = max - min;
     scales[sb] = range > 0 ? range / 15 : 0;
 
     const invScale = scales[sb] > 0 ? 1 / scales[sb] : 0;
-    for (let i = 0; i < 32; i++) {
+    for (let i = 0; i < validInSubblock; i++) {
       const val = data[sbOffset + i];
       let q = Math.round((val - min) * invScale);
       q = Math.max(0, Math.min(15, q));
@@ -155,6 +164,10 @@ export function quantizeQ4KBlock(data, offset) {
   return block;
 }
 
+export function quantizeQ4KBlock(data, offset) {
+  return quantizeQ4KBlockWithValidLength(data, offset, QK_K);
+}
+
 function dequantizeQ4KBlock(block) {
   const blockView = new DataView(block.buffer, block.byteOffset);
   const result = new Float32Array(256);
@@ -245,22 +258,16 @@ export function quantizeToQ4KMRowWise(data, shape) {
   }
 
   const blocksPerRow = Math.ceil(cols / QK_K);
-  const paddedColsPerRow = blocksPerRow * QK_K;
   const totalBlocks = rows * blocksPerRow;
 
   const quantized = new Uint8Array(totalBlocks * QK4_K_BLOCK_SIZE);
 
   for (let row = 0; row < rows; row++) {
-    // Extract and pad this row
-    const rowData = new Float32Array(paddedColsPerRow);
-    const srcOffset = row * cols;
-    for (let c = 0; c < cols; c++) {
-      rowData[c] = data[srcOffset + c];
-    }
-
     // Quantize each block in this row
     for (let b = 0; b < blocksPerRow; b++) {
-      const block = quantizeQ4KBlock(rowData, b * QK_K);
+      const validLength = Math.max(0, Math.min(QK_K, cols - b * QK_K));
+      const srcOffset = row * cols + b * QK_K;
+      const block = quantizeQ4KBlockWithValidLength(data, srcOffset, validLength);
       const dstOffset = (row * blocksPerRow + b) * QK4_K_BLOCK_SIZE;
       quantized.set(block, dstOffset);
     }

package/src/converter/rope-config.js

@@ -1,3 +1,5 @@
+import { DEFAULT_MANIFEST_INFERENCE } from '../config/schema/index.js';
+
 function asObject(value) {
   if (value == null || typeof value !== 'object' || Array.isArray(value)) {
     return null;
@@ -50,7 +52,7 @@ function resolveScalingConfig(ropeScalingConfig, options = {}) {
     }
     return {
       ropeScalingType: null,
-      ropeScalingFactor: 1.0,
+      ropeScalingFactor: DEFAULT_MANIFEST_INFERENCE.rope.ropeScalingFactor,
       yarnBetaFast: null,
       yarnBetaSlow: null,
       yarnOriginalMaxPos: null,
@@ -58,7 +60,7 @@ function resolveScalingConfig(ropeScalingConfig, options = {}) {
   }
 
   let ropeScalingType = scalingType;
-  let ropeScalingFactor = 1.0;
+  let ropeScalingFactor = DEFAULT_MANIFEST_INFERENCE.rope.ropeScalingFactor;
   let yarnBetaFast = null;
   let yarnBetaSlow = null;
   let yarnOriginalMaxPos = null;
@@ -110,7 +112,7 @@ function hasScalingDirective(ropeScalingConfig) {
 function hasMeaningfulScalingConfig(resolvedScaling) {
   if (!resolvedScaling) return false;
   return resolvedScaling.ropeScalingType != null
-    || resolvedScaling.ropeScalingFactor !== 1.0
+    || resolvedScaling.ropeScalingFactor !== DEFAULT_MANIFEST_INFERENCE.rope.ropeScalingFactor
     || resolvedScaling.yarnBetaFast != null
     || resolvedScaling.yarnBetaSlow != null
     || resolvedScaling.yarnOriginalMaxPos != null;
@@ -159,7 +161,7 @@ export function buildRoPEConfig(presetInference, config) {
       ?? null,
     ropeScalingFactor: presetRoPE.ropeScalingFactor
       ?? presetAttn?.ropeScalingFactor  // Deprecated location
-      ?? 1.0,
+      ?? DEFAULT_MANIFEST_INFERENCE.rope.ropeScalingFactor,
     yarnBetaFast: presetRoPE.yarnBetaFast ?? null,
     yarnBetaSlow: presetRoPE.yarnBetaSlow ?? null,
     yarnOriginalMaxPos: presetRoPE.yarnOriginalMaxPos ?? null,
@@ -223,7 +225,7 @@ export function buildRoPEConfig(presetInference, config) {
     ?? asFiniteNumber(flatRoPEParameters?.rope_theta)
     ?? asFiniteNumber(config.rope_theta)
     ?? presetInference.rope?.ropeTheta
-    ?? 10000;
+    ?? DEFAULT_MANIFEST_INFERENCE.rope.ropeTheta;
 
   // For Gemma 3, local sliding attention theta comes from rope_parameters.sliding_attention.
   const ropeLocalTheta = asFiniteNumber(slidingAttentionRoPE?.rope_theta)
@@ -232,7 +234,7 @@ export function buildRoPEConfig(presetInference, config) {
 
   const mropeInterleaved = asBoolean(flatRoPEParameters?.mrope_interleaved)
     ?? presetInference.rope?.mropeInterleaved
-    ?? false;
+    ?? DEFAULT_MANIFEST_INFERENCE.rope.mropeInterleaved;
   const mropeSection = asNumberArray(flatRoPEParameters?.mrope_section)
     ?? presetInference.rope?.mropeSection
     ?? null;

package/src/converter/tokenizer-utils.d.ts

@@ -1,5 +1,6 @@
 export declare function resolveEosTokenId(options: {
   config?: Record<string, unknown> | null;
+  generationConfig?: Record<string, unknown> | null;
   tokenizer?: {
     eosTokenId?: number;
     eos_token_id?: number;

package/src/converter/tokenizer-utils.js

@@ -1,6 +1,8 @@
-export function resolveEosTokenId({ config, tokenizer, tokenizerJson }) {
+export function resolveEosTokenId({ config, generationConfig, tokenizer, tokenizerJson }) {
   const nestedTextConfig = getNestedTextConfig(config);
   const candidateSources = [
+    generationConfig?.eos_token_id,
+    generationConfig?.eos_token_ids,
     tokenizer?.eosTokenId,
     tokenizer?.eos_token_id,
     tokenizerJson?.specialTokens?.eos,
@@ -19,6 +21,7 @@ export function resolveEosTokenId({ config, tokenizer, tokenizerJson }) {
   }
 
   const eosTokenStringCandidates = [
+    generationConfig?.eos_token,
     tokenizer?.eosToken,
     tokenizer?.eos_token,
     tokenizerJson?.specialTokens?.eos_token,