@simulatte/doppler 0.1.6 → 0.1.8

package/src/config/param-validator.js

@@ -2,6 +2,7 @@ import { log } from '../debug/index.js';
 import { PARAM_CATEGORIES, CategoryRules } from './param-categories.js';
 import { TOOLING_INTENTS, TOOLING_DIAGNOSTICS } from './schema/tooling.schema.js';
 import { validateEcosystemConfig } from './schema/ecosystem.schema.js';
+import { isPlainObject } from '../utils/plain-object.js';
 
 export function validateCallTimeOptions(options) {
   if (!options) return;
@@ -33,7 +34,23 @@ export function validateCallTimeOptions(options) {
 }
 
 export function validateRuntimeOverrides(overrides) {
+  if (!isPlainObject(overrides)) {
+    throw new Error('DopplerConfigError: runtime overrides must be an object when provided.');
+  }
+
+  assertRequiredRuntimeOverrideNotNull(overrides, 'shared');
+  assertRequiredRuntimeOverrideNotNull(overrides, 'loading');
+  assertRequiredRuntimeOverrideNotNull(overrides, 'inference');
+  assertRequiredRuntimeOverrideNotNull(overrides, 'emulation');
+  assertRequiredRuntimeOverrideNotNull(overrides?.inference, 'batching', 'runtime.inference');
+  assertRequiredRuntimeOverrideNotNull(overrides?.inference, 'compute', 'runtime.inference');
+  assertRequiredRuntimeOverrideNotNull(overrides?.inference, 'generation', 'runtime.inference');
+  assertRequiredRuntimeOverrideNotNull(overrides?.inference, 'kernelPathPolicy', 'runtime.inference');
+
   const modelOverrides = overrides?.inference?.modelOverrides;
+  if (modelOverrides !== undefined && modelOverrides !== null && !isPlainObject(modelOverrides)) {
+    throw new Error('DopplerConfigError: runtime.inference.modelOverrides must be an object when provided.');
+  }
   if (!modelOverrides) return;
 
   const params = flattenObject(modelOverrides);
@@ -214,6 +231,15 @@ function validateKernelPathPolicy(label, value) {
   if (!value || typeof value !== 'object' || Array.isArray(value)) {
     throw new Error(`DopplerConfigError: ${label} must be an object.`);
   }
+  if (
+    value.sourceScope !== undefined
+    && value.allowSources !== undefined
+    && !arraysEqual(value.sourceScope, value.allowSources)
+  ) {
+    throw new Error(
+      `DopplerConfigError: ${label}.sourceScope and ${label}.allowSources must match exactly when both are provided.`
+    );
+  }
   if (value.mode !== 'locked' && value.mode !== 'capability-aware') {
     throw new Error(`DopplerConfigError: ${label}.mode must be "locked" or "capability-aware".`);
   }
@@ -224,12 +250,31 @@ function validateKernelPathPolicy(label, value) {
   if (value.onIncompatible !== 'error' && value.onIncompatible !== 'remap') {
     throw new Error(`DopplerConfigError: ${label}.onIncompatible must be "error" or "remap".`);
   }
-  const validSources = new Set(['model', 'manifest', 'config', 'runtime', 'execution-v0']);
+  const validSources = new Set(['model', 'manifest', 'config', 'execution-v0']);
   for (const source of sourceScope) {
     if (!validSources.has(source)) {
       throw new Error(
-        `DopplerConfigError: ${label}.sourceScope entries must be model|manifest|config|runtime|execution-v0.`
+        `DopplerConfigError: ${label}.sourceScope entries must be model|manifest|config|execution-v0.`
       );
     }
   }
 }
+
+function assertRequiredRuntimeOverrideNotNull(container, key, prefix = 'runtime') {
+  if (!isPlainObject(container) || !Object.prototype.hasOwnProperty.call(container, key)) {
+    return;
+  }
+  if (container[key] === null) {
+    throw new Error(`DopplerConfigError: ${prefix}.${key} must not be null.`);
+  }
+}
+
+function arraysEqual(left, right) {
+  if (!Array.isArray(left) || !Array.isArray(right)) {
+    return false;
+  }
+  if (left.length !== right.length) {
+    return false;
+  }
+  return left.every((value, index) => value === right[index]);
+}

package/src/config/presets/kernel-paths/{gemma2-q4k-dequant-f32a.json → gemma2-q4k-dequant-f32a-nosubgroups.json}

@@ -1,7 +1,7 @@
 {
-  "id": "gemma2-q4k-dequant-f32a",
-  "name": "Gemma 2 Q4K Dequant (F32 activations)",
-  "description": "Q4K weights dequantized to F16 with F32 activations. Non-fused compatibility path with no subgroup requirement.",
+  "id": "gemma2-q4k-dequant-f32a-nosubgroups",
+  "name": "Gemma 2 Q4K Dequant (F32 activations, no subgroups)",
+  "description": "Q4K weights dequantized to F16 with F32 activations. Non-fused path with no subgroup requirement that still requires shader-f16 kernels.",
   "activationDtype": "f32",
 
   "decode": {

package/src/config/presets/kernel-paths/gemma3-f16-fused-f32a-online-streamingprefill.json

@@ -0,0 +1,223 @@
+{
+    "id": "gemma3-f16-fused-f32a-online-streamingprefill",
+    "name": "Gemma 3 F16 (F32 activations, online, streaming prefill)",
+    "description": "F16 weights with F32 activations, online decode attention, and streaming prefill attention for Gemma 3 stability-sensitive runs.",
+    "activationDtype": "f32",
+    "kvDtype": "f16",
+    "decode": {
+        "steps": [
+            {
+                "op": "input_norm",
+                "kernel": "rmsnorm.wgsl",
+                "entry": "main"
+            },
+            {
+                "op": "q_proj",
+                "kernel": "matmul_gemv_subgroup.wgsl",
+                "entry": "main_vec4",
+                "weights": "layer.{L}.self_attn.q_proj"
+            },
+            {
+                "op": "k_proj",
+                "kernel": "matmul_gemv_subgroup.wgsl",
+                "entry": "main_vec4",
+                "weights": "layer.{L}.self_attn.k_proj"
+            },
+            {
+                "op": "v_proj",
+                "kernel": "matmul_gemv_subgroup.wgsl",
+                "entry": "main_vec4",
+                "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_gemv_subgroup.wgsl",
+                "entry": "main_vec4",
+                "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_gemv_subgroup.wgsl",
+                "entry": "main_vec4",
+                "weights": "layer.{L}.mlp.gate_proj"
+            },
+            {
+                "op": "up_proj",
+                "kernel": "matmul_gemv_subgroup.wgsl",
+                "entry": "main_vec4",
+                "weights": "layer.{L}.mlp.up_proj"
+            },
+            {
+                "op": "activation",
+                "kernel": "gelu.wgsl",
+                "entry": "main",
+                "constants": {
+                    "HAS_GATE": true
+                }
+            },
+            {
+                "op": "down_proj",
+                "kernel": "matmul_gemv_subgroup.wgsl",
+                "entry": "main_vec4",
+                "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_streaming_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.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_tiled.wgsl",
+            "entry": "main",
+            "weights": "lm_head"
+        }
+    ],
+    "sampling": [
+        {
+            "op": "sample",
+            "kernel": "sample.wgsl",
+            "entry": "sample_single_pass"
+        }
+    ]
+}

package/src/config/presets/kernel-paths/{gemma3-q4k-dequant-f32a.json → gemma3-q4k-dequant-f32a-nosubgroups.json}

@@ -1,7 +1,7 @@
 {
-  "id": "gemma3-q4k-dequant-f32a",
-  "name": "Gemma 3 Q4K Dequant (F32 activations)",
-  "description": "Q4K weights dequantized to F16 with F32 activations for improved stability on small Gemma 3 checkpoints.",
+  "id": "gemma3-q4k-dequant-f32a-nosubgroups",
+  "name": "Gemma 3 Q4K Dequant (F32 activations, no subgroups)",
+  "description": "Q4K weights dequantized to F16 with F32 activations. Subgroup-free non-online path for Gemma 3 that still requires shader-f16 kernels.",
   "activationDtype": "f32",
   "kvDtype": "f16",
 

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

@@ -16,11 +16,18 @@
       "notes": "Default Gemma 2 Q4K dequant path for baseline F16 activation workflows."
     },
     {
-      "id": "gemma2-q4k-dequant-f32a",
-      "file": "gemma2-q4k-dequant-f32a.json",
+      "id": "gemma2-q4k-dequant-f32a-nosubgroups",
+      "file": "gemma2-q4k-dequant-f32a-nosubgroups.json",
       "status": "canonical",
-      "statusReason": "compatibility",
-      "notes": "Subgroup-free Gemma 2 Q4K dequant path with F32 activations."
+      "statusReason": "subgroup-free",
+      "notes": "Subgroup-free Gemma 2 Q4K dequant path with F32 activations. Still requires shader-f16 kernels."
+    },
+    {
+      "id": "gemma2-q4k-dequant-f32a",
+      "aliasOf": "gemma2-q4k-dequant-f32a-nosubgroups",
+      "status": "legacy",
+      "statusReason": "compatibility-alias",
+      "notes": "Legacy alias for gemma2-q4k-dequant-f32a-nosubgroups."
     },
     {
       "id": "gemma2-f16-f16a",
@@ -50,6 +57,13 @@
       "statusReason": "benchmark-probe",
       "notes": "Experimental fused F32 pipeline variant for fast decode."
     },
+    {
+      "id": "gemma3-f16-fused-f32a-online-streamingprefill",
+      "file": "gemma3-f16-fused-f32a-online-streamingprefill.json",
+      "status": "experimental",
+      "statusReason": "stability-probe",
+      "notes": "Gemma 3 F16/F32 online path with streaming prefill attention instead of the small-tile prefill kernel."
+    },
     {
       "id": "gemma3-q4k-dequant-f16a-online",
       "file": "gemma3-q4k-dequant-f16a-online.json",
@@ -58,11 +72,18 @@
       "notes": "Experimental Gemma 3 Q4K path using online decode attention on subgroup-capable GPUs."
     },
     {
-      "id": "gemma3-q4k-dequant-f32a",
-      "file": "gemma3-q4k-dequant-f32a.json",
+      "id": "gemma3-q4k-dequant-f32a-nosubgroups",
+      "file": "gemma3-q4k-dequant-f32a-nosubgroups.json",
       "status": "canonical",
-      "statusReason": "accuracy-first",
-      "notes": "Gemma 3 Q4K dequant path with F32 activation for quality-first runs."
+      "statusReason": "subgroup-free",
+      "notes": "Subgroup-free Gemma 3 Q4K dequant path with F32 activations. Still requires shader-f16 kernels."
+    },
+    {
+      "id": "gemma3-q4k-dequant-f32a",
+      "aliasOf": "gemma3-q4k-dequant-f32a-nosubgroups",
+      "status": "legacy",
+      "statusReason": "compatibility-alias",
+      "notes": "Legacy alias for gemma3-q4k-dequant-f32a-nosubgroups."
     },
     {
       "id": "gemma3-q4k-dequant-f32a-online",
@@ -71,6 +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",
@@ -78,6 +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
@@ -40,8 +41,8 @@
         "f32": "gemma2-f16-f32a"
       },
       "q4k": {
-        "f16": "gemma2-q4k-dequant-f32a",
-        "f32": "gemma2-q4k-dequant-f32a"
+        "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/experiments/bench/gemma3-bench-q4k.json

@@ -6,7 +6,7 @@
   "owner": "doppler-core",
   "createdAtUtc": "2026-02-25T00:00:00Z",
   "extends": "modes/bench",
-  "model": "gemma-3-1b-it-wq4k-ef16-hf16",
+  "model": "gemma-3-1b-it-q4k-ehf16-af32",
   "runtime": {
     "shared": {
       "benchmark": {

package/src/config/presets/runtime/experiments/debug/gemma3-debug-q4k.json

@@ -7,7 +7,7 @@
   "owner": "doppler-core",
   "createdAtUtc": "2026-02-25T00:00:00Z",
   "extends": "default",
-  "model": "gemma-3-1b-it-wq4k-ef16-hf16",
+  "model": "gemma-3-1b-it-q4k-ehf16-af32",
   "runtime": {
     "shared": {
       "tooling": {

package/src/config/presets/runtime/experiments/verify/gemma3-verify.json

@@ -6,7 +6,7 @@
   "owner": "doppler-core",
   "createdAtUtc": "2026-02-25T00:00:00Z",
   "extends": "modes/bench",
-  "model": "gemma-3-1b-it-wf16-ef16-hf16",
+  "model": "gemma-3-1b-it-f16-af32",
   "runtime": {
     "shared": {
       "tooling": {