@simulatte/doppler 0.1.7 → 0.1.8

package/CHANGELOG.md

@@ -6,6 +6,25 @@ This changelog is package-facing and release-oriented. Entries before `0.1.7`
 were retrofitted from package version history, release commits, and release
 docs so the `0.1.x` line has one conventional npm-visible history surface.
 
+## [0.1.8] - 2026-03-13
+
+### Changed
+
+- Simplified demo to show only verified Q4K models (Gemma 3 270M, Gemma 3 1B).
+  Hidden Translate, Diffusion, and Embedding tabs until models are ready.
+- Trimmed hosted HF registry and quickstart registry to the two verified models.
+- Aligned catalog, HF registry, and quickstart registry to the canonical
+  external support registry as single source of truth for HF revisions.
+
+### Fixed
+
+- Fixed Qwen 3.5 conversion configs using wrong model preset (`qwen3` instead
+  of `qwen3_5`), which caused support matrix check failures.
+- Fixed catalog lifecycle metadata inconsistencies: corrected `local`, `hf`,
+  `curated`, and `demo` fields to match actual artifact availability.
+- Removed failing and unverified models from demo visibility (TranslateGemma 4B,
+  EmbeddingGemma 300M with broken HF manifest, Qwen 3.5 0.8B/2B, F16 variant).
+
 ## [0.1.7] - 2026-03-10
 
 ### Added

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@simulatte/doppler",
-  "version": "0.1.7",
+  "version": "0.1.8",
   "description": "Browser-native WebGPU inference engine for local intent and inference loops",
   "main": "src/index.js",
   "types": "src/index.d.ts",
@@ -29,22 +29,22 @@
     "bench:chart": "node ./benchmarks/vendors/compare-chart.js",
     "bench:chart:readme": "node ./benchmarks/vendors/compare-chart.js --preset readme-evidence",
     "bench:architecture:chart": "node ./benchmarks/vendors/generate-architecture-overview-svg.js",
-    "ci:diffusion:contract": "node tools/ci-diffusion-contract-gates.mjs",
-    "ci:diffusion:contract:list": "node tools/ci-diffusion-contract-gates.mjs --list",
-    "ci:training:contract": "node tools/ci-training-contract-gates.mjs",
-    "ci:training:contract:list": "node tools/ci-training-contract-gates.mjs --list",
-    "training:contract:delta": "node tools/emit-training-contract-delta.mjs",
-    "training:workloads:verify": "node tools/verify-training-workload-packs.mjs --registry tools/configs/training-workloads/registry.json",
-    "training:report-ids:publish": "node tools/publish-training-report-ids.mjs --registry tools/configs/training-workloads/registry.json",
-    "distill:studio:mvp": "node tools/distill-studio-mvp.mjs",
-    "distill:quality-gate": "node tools/distill-studio-quality-gate.mjs",
-    "p2p:observability": "node tools/p2p-delivery-observability.mjs",
-    "p2p:drill": "node tools/p2p-resilience-drill.mjs",
+    "ci:diffusion:contract": "node tools/ci-diffusion-contract-gates.js",
+    "ci:diffusion:contract:list": "node tools/ci-diffusion-contract-gates.js --list",
+    "ci:training:contract": "node tools/ci-training-contract-gates.js",
+    "ci:training:contract:list": "node tools/ci-training-contract-gates.js --list",
+    "training:contract:delta": "node tools/emit-training-contract-delta.js",
+    "training:workloads:verify": "node tools/verify-training-workload-packs.js --registry tools/configs/training-workloads/registry.json",
+    "training:report-ids:publish": "node tools/publish-training-report-ids.js --registry tools/configs/training-workloads/registry.json",
+    "distill:studio:mvp": "node tools/distill-studio-mvp.js",
+    "distill:quality-gate": "node tools/distill-studio-quality-gate.js",
+    "p2p:observability": "node tools/p2p-delivery-observability.js",
+    "p2p:drill": "node tools/p2p-resilience-drill.js",
     "test": "npm run test:unit",
-    "test:unit": "node tools/run-node-tests.mjs --suite unit",
-    "test:gpu": "node tools/run-node-tests.mjs --suite gpu",
-    "test:coverage": "node tools/run-node-coverage.mjs",
-    "test:coverage:report": "node tools/run-node-coverage.mjs --no-threshold",
+    "test:unit": "node tools/run-node-tests.js --suite unit",
+    "test:gpu": "node tools/run-node-tests.js --suite gpu",
+    "test:coverage": "node tools/run-node-coverage.js",
+    "test:coverage:report": "node tools/run-node-coverage.js --no-threshold",
     "test:gpu:browser": "node tools/doppler-cli.js verify --config '{\"request\":{\"suite\":\"kernels\"},\"run\":{\"surface\":\"browser\",\"browser\":{\"opfsCache\":false,\"headless\":true,\"channel\":\"chromium\",\"browserArgs\":[\"--use-angle=swiftshader\",\"--disable-vulkan-surface\"],\"console\":true}}}'",
     "agents:verify": "node tools/verify-agent-parity.js",
     "agents:freshness": "node tools/verify-agent-freshness.js",
@@ -74,9 +74,10 @@
     "ci:catalog:check": "npm run registry:sync:scripts:check && npm run support:matrix:check && npm run registry:hf:check",
     "external:rdrr:index": "node tools/sync-external-rdrr-index.js",
     "external:rdrr:index:check": "node tools/sync-external-rdrr-index.js --check",
-    "verify:embeddinggemma-300m": "node tools/run-registry-verify.js embeddinggemma-300m",
-    "verify:gemma-3-1b-it-f16": "node tools/run-registry-verify.js gemma-3-1b-it-f16",
-    "verify:gemma-3-1b-it-f16-af32": "node tools/run-registry-verify.js gemma-3-1b-it-f16-af32",
+    "external:support:sync": "node tools/sync-external-support-registry.js",
+    "external:support:check": "node tools/sync-external-support-registry.js --check",
+    "catalog:sync:external": "node tools/sync-catalog-from-external-support.js",
+    "catalog:sync:external:check": "node tools/sync-catalog-from-external-support.js --check",
     "verify:gemma-3-1b-it-q4k-ehf16-af32": "node tools/run-registry-verify.js gemma-3-1b-it-q4k-ehf16-af32",
     "verify:gemma-3-1b-it-wq4k-ef16-hf16": "node tools/run-registry-verify.js gemma-3-1b-it-wq4k-ef16-hf16",
     "verify:gemma-3-270m-it-q4k-ehf16-af32": "node tools/run-registry-verify.js gemma-3-270m-it-q4k-ehf16-af32",
@@ -84,25 +85,9 @@
     "verify:gemma-3-270m-it-wq4k-ef16-hf16": "node tools/run-registry-verify.js gemma-3-270m-it-wq4k-ef16-hf16",
     "verify:gemma-3-270m-it-wq4k-ef16-hf16-f32": "node tools/run-registry-verify.js gemma-3-270m-it-wq4k-ef16-hf16-f32",
     "verify:gemma3-1b": "node tools/run-registry-verify.js gemma3-1b",
-    "verify:gemma3-1b-f16": "node tools/run-registry-verify.js gemma3-1b-f16",
     "verify:gemma3-270m": "node tools/run-registry-verify.js gemma3-270m",
-    "verify:google-embeddinggemma-300m": "node tools/run-registry-verify.js google-embeddinggemma-300m",
-    "verify:google-embeddinggemma-300m-q4k-ehf16-af32": "node tools/run-registry-verify.js google-embeddinggemma-300m-q4k-ehf16-af32",
-    "verify:google-embeddinggemma-300m-wq4k-ef16": "node tools/run-registry-verify.js google-embeddinggemma-300m-wq4k-ef16",
     "verify:google-gemma-3-1b-it": "node tools/run-registry-verify.js google-gemma-3-1b-it",
-    "verify:google-gemma-3-270m-it": "node tools/run-registry-verify.js google-gemma-3-270m-it",
-    "verify:google-translategemma-4b-it": "node tools/run-registry-verify.js google-translategemma-4b-it",
-    "verify:qwen-3-5-0-8b": "node tools/run-registry-verify.js qwen-3-5-0-8b",
-    "verify:qwen-3-5-0-8b-wq4k-ef16-hf16-f16": "node tools/run-registry-verify.js qwen-3-5-0-8b-wq4k-ef16-hf16-f16",
-    "verify:qwen-3-5-2b": "node tools/run-registry-verify.js qwen-3-5-2b",
-    "verify:qwen-3-5-2b-wq4k-ef16-hf16-f16": "node tools/run-registry-verify.js qwen-3-5-2b-wq4k-ef16-hf16-f16",
-    "verify:qwen-qwen3.5-0.8b": "node tools/run-registry-verify.js qwen-qwen3.5-0.8b",
-    "verify:qwen-qwen3.5-2b": "node tools/run-registry-verify.js qwen-qwen3.5-2b",
-    "verify:qwen3-0.8b": "node tools/run-registry-verify.js qwen3-0.8b",
-    "verify:qwen3-2b": "node tools/run-registry-verify.js qwen3-2b",
-    "verify:translategemma": "node tools/run-registry-verify.js translategemma",
-    "verify:translategemma-4b": "node tools/run-registry-verify.js translategemma-4b",
-    "verify:translategemma-4b-it-wq4k-ef16-hf16": "node tools/run-registry-verify.js translategemma-4b-it-wq4k-ef16-hf16"
+    "verify:google-gemma-3-270m-it": "node tools/run-registry-verify.js google-gemma-3-270m-it"
   },
   "exports": {
     ".": {

package/src/browser/browser-converter.js

@@ -408,6 +408,7 @@ export async function convertModel(files, options = {}) {
     // Parse based on format
     let modelInfo;
     let config = null;
+    let generationConfig = null;
     let tokenizerJson = null;
     let tokenizerConfig = null;
     let tokenizerModel = null;
@@ -455,6 +456,10 @@ export async function convertModel(files, options = {}) {
         tokenizerConfig = await parseTokenizerConfigJson(auxiliary.tokenizerConfig);
         modelInfo.tokenizerConfig = tokenizerConfig;
       }
+      if (auxiliary.generationConfig) {
+        generationConfig = await parseConfigJson(auxiliary.generationConfig);
+        modelInfo.generationConfig = generationConfig;
+      }
       if (auxiliary.tokenizerModel) {
         const source = normalizeTensorSource(auxiliary.tokenizerModel);
         tokenizerModel = await source.readRange(0, source.size);

package/src/client/doppler-registry.json

@@ -16,25 +16,9 @@
       ],
       "hf": {
         "repoId": "Clocksmith/rdrr",
-        "revision": "cd6c12be0e83e92d6dbd92598a0aa94391ec7e94",
+        "revision": "ca6f0dbdf3882d3893a65cf48f2bb6f1520df162",
         "path": "models/gemma-3-270m-it-q4k-ehf16-af32"
       }
-    },
-    {
-      "modelId": "google-embeddinggemma-300m-q4k-ehf16-af32",
-      "aliases": [
-        "embeddinggemma-300m",
-        "google/embeddinggemma-300m",
-        "google-embeddinggemma-300m-wq4k-ef16"
-      ],
-      "modes": [
-        "embedding"
-      ],
-      "hf": {
-        "repoId": "Clocksmith/rdrr",
-        "revision": "b23aca921ea11729d6f34b9484555968a5ab0e42",
-        "path": "models/google-embeddinggemma-300m-q4k-ehf16-af32"
-      }
     }
   ]
 }

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

@@ -134,6 +134,11 @@ export function getKernelPathStrict(): boolean;
  */
 export function isKernelPathFusedQ4K(path?: KernelPathSchema | null): boolean;
 
+/**
+ * Check if a kernel path requires matmul weights to stay in F32.
+ */
+export function kernelPathRequiresF32MatmulWeights(path?: KernelPathSchema | null): boolean;
+
 /**
  * Check if the active kernel path uses fused Q4K matmul.
  */

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

@@ -503,6 +503,19 @@ export function isKernelPathFusedQ4K(path = undefined) {
   return kernelSteps.some((step) => step.kernel.includes('fused_matmul_q4'));
 }
 
+export function kernelPathRequiresF32MatmulWeights(path = undefined) {
+  const lookupPath = path === undefined ? activeKernelPath : path;
+  if (!lookupPath) return false;
+  const kernelSteps = [
+    ...(lookupPath.decode?.steps ?? []),
+    ...(lookupPath.prefill?.steps ?? []),
+    ...(lookupPath.preLayer ?? []),
+    ...(lookupPath.postLayer ?? []),
+    ...(lookupPath.layerOverrides?.flatMap((override) => override.steps) ?? []),
+  ];
+  return kernelSteps.some((step) => normalizeKernelFile(step.kernel) === 'matmul_f32.wgsl');
+}
+
 export function isActiveKernelPathFusedQ4K() {
   return isKernelPathFusedQ4K(activeKernelPath);
 }

package/src/config/kernels/registry.json

@@ -4322,6 +4322,80 @@
         }
       }
     },
+    "split_qg": {
+      "description": "De-interleave Q and Gate projections from q_proj output for attentionOutputGate models",
+      "baseBindings": [
+        {
+          "index": 0,
+          "name": "uniforms",
+          "type": "uniform"
+        },
+        {
+          "index": 1,
+          "name": "qg_interleaved",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 2,
+          "name": "Q",
+          "type": "storage"
+        },
+        {
+          "index": 3,
+          "name": "G",
+          "type": "storage"
+        }
+      ],
+      "baseUniforms": {
+        "size": 16,
+        "fields": [
+          {
+            "name": "num_tokens",
+            "type": "u32",
+            "offset": 0
+          },
+          {
+            "name": "num_heads",
+            "type": "u32",
+            "offset": 4
+          },
+          {
+            "name": "head_dim",
+            "type": "u32",
+            "offset": 8
+          },
+          {
+            "name": "_pad",
+            "type": "u32",
+            "offset": 12
+          }
+        ]
+      },
+      "variants": {
+        "default": {
+          "wgsl": "split_qg.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": []
+        },
+        "f16": {
+          "wgsl": "split_qg_f16.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [
+            "shader-f16"
+          ]
+        }
+      }
+    },
     "sample": {
       "description": "GPU-side sampling kernels",
       "baseBindings": [

package/src/config/loader.js

@@ -23,6 +23,7 @@ const mambaPreset = await loadJson('./presets/models/mamba.json', import.meta.ur
 const modernbertPreset = await loadJson('./presets/models/modernbert.json', import.meta.url, 'Failed to load preset');
 const lfm2Preset = await loadJson('./presets/models/lfm2.json', import.meta.url, 'Failed to load preset');
 const qwen3Preset = await loadJson('./presets/models/qwen3.json', import.meta.url, 'Failed to load preset');
+const qwen35Preset = await loadJson('./presets/models/qwen3_5.json', import.meta.url, 'Failed to load preset');
 const kimiK2Preset = await loadJson('./presets/models/kimi-k2.json', import.meta.url, 'Failed to load preset');
 const gptOssPreset = await loadJson('./presets/models/gpt-oss.json', import.meta.url, 'Failed to load preset');
 
@@ -46,6 +47,7 @@ export const PRESET_REGISTRY = {
   modernbert: modernbertPreset,
   lfm2: lfm2Preset,
   qwen3: qwen3Preset,
+  qwen3_5: qwen35Preset,
   kimi_k2: kimiK2Preset,
   gpt_oss: gptOssPreset,
 };
@@ -97,6 +99,7 @@ export const PRESET_DETECTION_ORDER = [
   'gemma3',
   'llama3',
   'lfm2',
+  'qwen3_5',
   'qwen3',
   'kimi_k2',
   'gpt_oss',

package/src/config/merge-contract-check.js

@@ -171,6 +171,13 @@ export function buildMergeContractArtifact() {
     `configA=${isolatedConfigA.runtime.inference.compute.activationDtype}, configB=${isolatedConfigB.runtime.inference.compute.activationDtype}`,
     'actual'
   );
+  recordCheck(
+    checks,
+    'runtime.schema.storage.opfs_sync_access_handle_defaults_off',
+    isolatedConfigB.runtime.loading.storage.backend.opfs.useSyncAccessHandle === false,
+    `value=${String(isolatedConfigB.runtime.loading.storage.backend.opfs.useSyncAccessHandle)}`,
+    'actual'
+  );
 
   const calibrateConfig = createDopplerConfig({
     runtime: {

package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f32w-f32a-online.json

@@ -0,0 +1,56 @@
+{
+  "id": "gemma3-q4k-dequant-f32w-f32a-online",
+  "name": "Gemma 3 Q4K Dequant (F32 projection weights, F32 activations, online decode)",
+  "description": "Q4K projection weights dequantized to F32 with F32 activations. Tied embeddings and LM head stay on the native F16 path. Decode uses online attention; prefill uses streaming attention.",
+  "activationDtype": "f32",
+  "kvDtype": "f16",
+  "decode": {
+    "steps": [
+      { "op": "input_norm", "kernel": "rmsnorm.wgsl", "entry": "main" },
+      { "op": "q_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.q_proj" },
+      { "op": "k_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.k_proj" },
+      { "op": "v_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.v_proj" },
+      { "op": "rope_q", "kernel": "rope.wgsl", "entry": "main" },
+      { "op": "rope_k", "kernel": "rope.wgsl", "entry": "main" },
+      { "op": "attention", "kernel": "attention_decode_online_f16kv.wgsl", "entry": "main" },
+      { "op": "o_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.o_proj" },
+      { "op": "attn_residual", "kernel": "residual.wgsl", "entry": "main" },
+      { "op": "post_attn_norm", "kernel": "rmsnorm.wgsl", "entry": "main" },
+      { "op": "gate_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.mlp.gate_proj" },
+      { "op": "up_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.mlp.up_proj" },
+      { "op": "activation", "kernel": "gelu.wgsl", "entry": "main", "constants": { "HAS_GATE": true } },
+      { "op": "down_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.mlp.down_proj" },
+      { "op": "ffn_residual", "kernel": "residual.wgsl", "entry": "main" }
+    ]
+  },
+  "prefill": {
+    "steps": [
+      { "op": "input_norm", "kernel": "rmsnorm.wgsl", "entry": "main" },
+      { "op": "q_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.q_proj" },
+      { "op": "k_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.k_proj" },
+      { "op": "v_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.v_proj" },
+      { "op": "rope_q", "kernel": "rope.wgsl", "entry": "main" },
+      { "op": "rope_k", "kernel": "rope.wgsl", "entry": "main" },
+      { "op": "attention", "kernel": "attention_streaming_f16kv.wgsl", "entry": "main" },
+      { "op": "o_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.o_proj" },
+      { "op": "attn_residual", "kernel": "residual.wgsl", "entry": "main" },
+      { "op": "post_attn_norm", "kernel": "rmsnorm.wgsl", "entry": "main" },
+      { "op": "gate_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.mlp.gate_proj" },
+      { "op": "up_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.mlp.up_proj" },
+      { "op": "activation", "kernel": "gelu.wgsl", "entry": "main", "constants": { "HAS_GATE": true } },
+      { "op": "down_proj", "kernel": "matmul_f32.wgsl", "entry": "main", "weights": "layer.{L}.mlp.down_proj" },
+      { "op": "ffn_residual", "kernel": "residual.wgsl", "entry": "main" }
+    ]
+  },
+  "preLayer": [
+    { "op": "embed", "kernel": "gather_f16.wgsl", "entry": "main", "weights": "embed_tokens" }
+  ],
+  "postLayer": [
+    { "op": "final_norm", "kernel": "rmsnorm.wgsl", "entry": "main" },
+    { "op": "lm_head", "kernel": "matmul_gemv_subgroup.wgsl", "entry": "main_multicol", "weights": "lm_head", "constants": { "MULTICOL_COLS_PER_WG": 64, "MULTICOL_THREADS_PER_COL": 4 } },
+    { "op": "lm_head_prefill", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "lm_head" }
+  ],
+  "sampling": [
+    { "op": "sample", "kernel": "sample.wgsl", "entry": "sample_single_pass" }
+  ]
+}

package/src/config/presets/kernel-paths/lfm2-q4k-dequant-f32a-nosubgroups.json

@@ -0,0 +1,61 @@
+{
+  "id": "lfm2-q4k-dequant-f32a-nosubgroups",
+  "name": "LFM2 Q4K Dequant (F32 activations, no subgroups)",
+  "description": "Subgroup-free LFM2 Q4K path: F32 activations with tiled prefill matmul and small-kernel prefill attention. Still requires shader-f16 kernels.",
+  "activationDtype": "f32",
+  "kvDtype": "f16",
+
+  "decode": {
+    "steps": [
+      { "op": "input_norm",    "kernel": "rmsnorm.wgsl",                        "entry": "main" },
+      { "op": "q_proj",        "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.self_attn.q_proj" },
+      { "op": "k_proj",        "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.self_attn.k_proj" },
+      { "op": "v_proj",        "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.self_attn.v_proj" },
+      { "op": "rope_q",        "kernel": "rope.wgsl",                           "entry": "main" },
+      { "op": "rope_k",        "kernel": "rope.wgsl",                           "entry": "main" },
+      { "op": "attention",     "kernel": "attention_decode_chunked_f16kv.wgsl",  "entry": "main" },
+      { "op": "o_proj",        "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
+      { "op": "attn_residual", "kernel": "residual.wgsl",                       "entry": "main" },
+      { "op": "post_attn_norm","kernel": "rmsnorm.wgsl",                        "entry": "main" },
+      { "op": "gate_proj",     "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.mlp.gate_proj" },
+      { "op": "up_proj",       "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.mlp.up_proj" },
+      { "op": "activation",    "kernel": "gelu.wgsl",                           "entry": "main", "constants": { "HAS_GATE": true } },
+      { "op": "down_proj",     "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
+      { "op": "ffn_residual",  "kernel": "residual.wgsl",                       "entry": "main" }
+    ]
+  },
+
+  "prefill": {
+    "steps": [
+      { "op": "input_norm",    "kernel": "rmsnorm.wgsl",                        "entry": "main" },
+      { "op": "q_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.q_proj" },
+      { "op": "k_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.k_proj" },
+      { "op": "v_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.v_proj" },
+      { "op": "rope_q",        "kernel": "rope.wgsl",                           "entry": "main" },
+      { "op": "rope_k",        "kernel": "rope.wgsl",                           "entry": "main" },
+      { "op": "attention",     "kernel": "attention_small_f16kv.wgsl",           "entry": "main" },
+      { "op": "o_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
+      { "op": "attn_residual", "kernel": "residual.wgsl",                       "entry": "main" },
+      { "op": "post_attn_norm","kernel": "rmsnorm.wgsl",                        "entry": "main" },
+      { "op": "gate_proj",     "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.gate_proj" },
+      { "op": "up_proj",       "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.up_proj" },
+      { "op": "activation",    "kernel": "gelu.wgsl",                           "entry": "main", "constants": { "HAS_GATE": true } },
+      { "op": "down_proj",     "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
+      { "op": "ffn_residual",  "kernel": "residual.wgsl",                       "entry": "main" }
+    ]
+  },
+
+  "preLayer": [
+    { "op": "embed",           "kernel": "gather_f16.wgsl",                     "entry": "main",  "weights": "embed_tokens" }
+  ],
+
+  "postLayer": [
+    { "op": "final_norm",      "kernel": "rmsnorm.wgsl",                        "entry": "main" },
+    { "op": "lm_head",         "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "lm_head" },
+    { "op": "lm_head_prefill", "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "lm_head" }
+  ],
+
+  "sampling": [
+    { "op": "sample",          "kernel": "sample.wgsl",                         "entry": "sample_single_pass" }
+  ]
+}

package/src/config/presets/kernel-paths/registry.json

@@ -92,6 +92,13 @@
       "statusReason": "default",
       "notes": "Gemma 3 Q4K dequant default: subgroup GEMV + online attention + tuned lm_head multicol, F32 activations."
     },
+    {
+      "id": "gemma3-q4k-dequant-f32w-f32a-online",
+      "file": "gemma3-q4k-dequant-f32w-f32a-online.json",
+      "status": "experimental",
+      "statusReason": "accuracy-probe",
+      "notes": "Gemma 3 Q4K dequant path that keeps matmul weights in F32 and runs F32 matmul kernels for numeric-sensitivity debugging."
+    },
     {
       "id": "lfm2-q4k-dequant-f32a-online",
       "file": "lfm2-q4k-dequant-f32a-online.json",
@@ -99,6 +106,13 @@
       "statusReason": "default",
       "notes": "LFM2 Q4K default: subgroup GEMV decode with tiled fast-prefill path and F32 activations."
     },
+    {
+      "id": "lfm2-q4k-dequant-f32a-nosubgroups",
+      "file": "lfm2-q4k-dequant-f32a-nosubgroups.json",
+      "status": "canonical",
+      "statusReason": "subgroup-free",
+      "notes": "Subgroup-free LFM2 Q4K dequant path with F32 activations and tiled prefill. Still requires shader-f16 kernels."
+    },
     {
       "id": "embeddinggemma-f16-f32a",
       "file": "embeddinggemma-f16-f32a.json",

package/src/config/presets/models/gemma2.json

@@ -9,6 +9,7 @@
 
   "inference": {
     "attention": {
+      "queryPreAttnScalar": 256,
       "slidingWindow": 4096,
       "attnLogitSoftcapping": 50.0,
       "queryKeyNorm": false
@@ -40,7 +41,7 @@
         "f32": "gemma2-f16-f32a"
       },
       "q4k": {
-        "f16": "gemma2-q4k-dequant-f32a-nosubgroups",
+        "f16": "gemma2-q4k-dequant-f16a",
         "f32": "gemma2-q4k-dequant-f32a-nosubgroups"
       }
     }

package/src/config/presets/models/gemma3.json

@@ -8,7 +8,9 @@
   },
   "inference": {
     "attention": {
+      "queryPreAttnScalar": 256,
       "attnLogitSoftcapping": null,
+      "slidingWindow": 512,
       "queryKeyNorm": true
     },
     "normalization": {

package/src/config/presets/models/qwen3.json

@@ -10,7 +10,8 @@
   "inference": {
     "attention": {
       "slidingWindow": null,
-      "queryKeyNorm": true
+      "queryKeyNorm": true,
+      "attentionOutputGate": true
     },
     "output": {
       "scaleEmbeddings": false
@@ -39,8 +40,8 @@
   },
 
   "detection": {
-    "architecturePatterns": ["qwen3", "qwen3_5", "Qwen3ForCausalLM", "Qwen3_5ForCausalLM", "Qwen2ForCausalLM"],
-    "modelTypePatterns": ["qwen3_5", "qwen3_5_text", "qwen2"],
+    "architecturePatterns": ["qwen3", "Qwen3ForCausalLM", "Qwen2ForCausalLM"],
+    "modelTypePatterns": ["qwen3", "qwen2"],
     "configPatterns": {
       "model_type": "qwen2"
     }

package/src/config/presets/models/qwen3_5.json

@@ -0,0 +1,16 @@
+{
+  "id": "qwen3_5",
+  "name": "Qwen 3.5",
+  "extends": "qwen3",
+
+  "inference": {
+    "normalization": {
+      "rmsNormWeightOffset": true
+    }
+  },
+
+  "detection": {
+    "architecturePatterns": ["qwen3_5", "Qwen3_5ForCausalLM", "Qwen3_5ForConditionalGeneration"],
+    "modelTypePatterns": ["qwen3_5", "qwen3_5_text"]
+  }
+}

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]
+          }
+        ]
+      }
+    }
+  }
+}