@simulatte/doppler 0.1.7 → 0.1.9

package/src/config/platforms/loader.js

@@ -9,6 +9,8 @@ const platformCache = new Map();
 
 let platformsBaseUrl = null;
 
+const DEFAULT_PREFER_UNIFIED_MEMORY = false;
+
 const PLATFORM_FILES = [
   'apple-m3',
   'apple-m2',
@@ -131,7 +133,7 @@ export function getMemoryHints() {
 }
 
 export function prefersUnifiedMemory() {
-  return getMemoryHints()?.preferUnifiedMemory ?? false;
+  return getMemoryHints()?.preferUnifiedMemory ?? DEFAULT_PREFER_UNIFIED_MEMORY;
 }
 
 export function getBufferAlignment() {

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

@@ -8,19 +8,19 @@
   "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": "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_decode_chunked_f16kv.wgsl", "entry": "main" },
-      { "op": "o_proj",       "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
+      { "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.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": "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.wgsl", "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
+      { "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" }
     ]
   },
@@ -28,19 +28,19 @@
   "prefill": {
     "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": "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.wgsl", "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
+      { "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.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": "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.wgsl", "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
+      { "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" }
     ]
   },
@@ -51,8 +51,8 @@
 
   "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.wgsl", "entry": "main",  "weights": "lm_head" }
+    { "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": [

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

@@ -28,19 +28,19 @@
   "prefill": {
     "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": "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.wgsl",               "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
+      { "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.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": "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.wgsl",               "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
+      { "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" }
     ]
   },
@@ -52,7 +52,7 @@
   "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" }
+    { "op": "lm_head_prefill", "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "lm_head" }
   ],
 
   "sampling": [

package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f32a-small-attn.json

@@ -0,0 +1,61 @@
+{
+  "id": "gemma3-q4k-dequant-f32a-small-attn",
+  "name": "Gemma 3 Q4K Dequant (F32 activations, small-attn prefill)",
+  "description": "Q4K dequantized to F16 with F32 activations. Same as gemma3-q4k-dequant-f32a-online but uses attention_small_f16kv.wgsl for prefill (diagnostic variant).",
+  "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_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_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-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 +92,20 @@
       "statusReason": "default",
       "notes": "Gemma 3 Q4K dequant default: subgroup GEMV + online attention + tuned lm_head multicol, F32 activations."
     },
+    {
+      "id": "gemma3-q4k-dequant-f32a-small-attn",
+      "file": "gemma3-q4k-dequant-f32a-small-attn.json",
+      "status": "experimental",
+      "statusReason": "diagnostic-probe",
+      "notes": "Diagnostic: same as gemma3-q4k-dequant-f32a-online but uses attention_small_f16kv.wgsl for prefill to isolate streaming attention bug."
+    },
+    {
+      "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 +113,13 @@
       "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": {
@@ -32,7 +34,8 @@
     },
     "rope": {
       "ropeTheta": 1000000,
-      "ropeLocalTheta": 10000
+      "ropeLocalTheta": 10000,
+      "ropeLocalScalingFactor": 1.0
     },
     "chatTemplate": {
       "type": "gemma",

package/src/config/presets/models/gemma4.json

@@ -0,0 +1,61 @@
+{
+  "id": "gemma4",
+  "name": "Gemma 4",
+  "extends": "gemma3",
+  "modelType": "mixtral",
+
+  "inference": {
+    "attention": {
+      "slidingWindow": 1024
+    },
+    "rope": {
+      "ropeTheta": 1000000,
+      "ropeLocalTheta": 10000,
+      "ropeScalingType": "yarn",
+      "ropeScalingFactor": 8.0,
+      "yarnBetaFast": 4.0,
+      "yarnBetaSlow": 1.0,
+      "yarnOriginalMaxPos": 32768
+    },
+    "moe": {
+      "kernelProfileId": "mixtral-moe-v1",
+      "numExperts": 8,
+      "topK": 2,
+      "numSharedExperts": 0,
+      "routerDtype": "f32",
+      "supportedActivationDtypes": ["f16", "f32"],
+      "preferredActivationDtype": "f32",
+      "tensorPattern": "mixtral"
+    },
+    "kernelPaths": {
+      "q4k": {
+        "default": "gemma3-q4k-dequant-f32a-online",
+        "f16": "gemma3-q4k-dequant-f16a-online",
+        "f16a": "gemma3-q4k-dequant-f16a-online",
+        "f32": "gemma3-q4k-dequant-f32a-online"
+      }
+    }
+  },
+
+  "tensorPatterns": {
+    "ffn": {
+      "gate": ["layers.{layer}.block_sparse_moe.experts.{expert}.w1.weight"],
+      "up": ["layers.{layer}.block_sparse_moe.experts.{expert}.w3.weight"],
+      "down": ["layers.{layer}.block_sparse_moe.experts.{expert}.w2.weight"]
+    }
+  },
+
+  "detection": {
+    "architecturePatterns": [
+      "gemma4",
+      "Gemma4ForCausalLM",
+      "Gemma4ForConditionalGeneration",
+      "gemma-4"
+    ],
+    "modelTypePatterns": [
+      "gemma4",
+      "gemma4_text",
+      "gemma4_moe"
+    ]
+  }
+}

package/src/config/presets/models/granite-docling.json

@@ -0,0 +1,70 @@
+{
+  "id": "granite-docling",
+  "name": "Granite-Docling (Document OCR VLM)",
+  "extends": "transformer",
+  "modelType": "ocr",
+
+  "_notes": "Stabilized successor to SmolDocling-256M-preview. Requires full multimodal pipeline: SigLIP vision encoder, Idefics3-style image-token merge, pixel-shuffle connector, SmolLM2 decoder, DocTags output parsing. This preset covers the decoder config only — vision encoder and connector are separate pipeline stages not yet implemented in Doppler.",
+
+  "architecture": {
+    "headDim": 64,
+    "ropeTheta": 10000,
+    "visionEncoder": {
+      "type": "siglip_b16",
+      "patchSize": 16,
+      "imageSize": 512,
+      "hiddenSize": 768,
+      "numLayers": 12,
+      "numHeads": 12,
+      "parameterCount": 93000000,
+      "_note": "SigLIP base patch-16/512 backbone. Requires dedicated vision encoder pipeline in Doppler."
+    },
+    "connector": {
+      "type": "mlp_pixel_shuffle",
+      "downsampleFactor": 2,
+      "_note": "Idefics3/SmolVLM-style projection. Maps vision tokens to decoder embedding space."
+    }
+  },
+
+  "inference": {
+    "attention": {
+      "queryKeyNorm": false,
+      "causal": true
+    },
+    "normalization": {
+      "rmsNormWeightOffset": false,
+      "rmsNormEps": 1e-5
+    },
+    "ffn": {
+      "activation": "silu"
+    },
+    "output": {
+      "scaleEmbeddings": false,
+      "tieWordEmbeddings": true
+    },
+    "chatTemplate": {
+      "enabled": false
+    },
+    "kernelPaths": {
+      "q4k": {
+        "f16": "granite-docling-q4k-dequant-f32a",
+        "f32": "granite-docling-q4k-dequant-f32a"
+      },
+      "f16": {
+        "f16": "granite-docling-f16-f32a",
+        "f32": "granite-docling-f16-f32a"
+      }
+    }
+  },
+
+  "tokenizer": {
+    "bosToken": "<|endoftext|>",
+    "eosTokens": ["<|endoftext|>", "<|im_end|>"],
+    "addBosToken": true
+  },
+
+  "detection": {
+    "architecturePatterns": ["granite-docling", "GraniteDocling", "smoldocling", "SmolDocling", "SmolVLM"],
+    "modelTypePatterns": ["granite-docling", "smoldocling", "smolvlm"]
+  }
+}

package/src/config/presets/models/lfm2.json

@@ -39,11 +39,16 @@
       "period": null,
       "offset": null,
       "layerTypes": null
+    },
+    "chatTemplate": {
+      "type": "chatml",
+      "enabled": true
     }
   },
 
   "tokenizer": {
-    "addBosToken": false,
+    "bosTokenId": 1,
+    "addBosToken": true,
     "addEosToken": false
   },
 

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/models/qwen3_vl.json

@@ -0,0 +1,40 @@
+{
+  "id": "qwen3_vl",
+  "name": "Qwen 3 VL",
+  "extends": "qwen3",
+
+  "architecture": {
+    "ropeTheta": 5000000
+  },
+
+  "inference": {
+    "normalization": {
+      "rmsNormWeightOffset": false
+    },
+    "rope": {
+      "ropeTheta": 5000000,
+      "mropeInterleaved": true,
+      "mropeSection": [24, 20, 20],
+      "partialRotaryFactor": null
+    }
+  },
+
+  "vision": {
+    "patchSize": 16,
+    "spatialMergeSize": 2,
+    "temporalPatchSize": 2,
+    "eps": 1e-6,
+    "minPixels": 3136,
+    "maxPixels": 1003520,
+    "projectorType": "spatial_merge",
+    "normalization": {
+      "mean": [0.48145466, 0.4578275, 0.40821073],
+      "std": [0.26862954, 0.26130258, 0.27577711]
+    }
+  },
+
+  "detection": {
+    "architecturePatterns": ["qwen3_vl", "Qwen3VLForConditionalGeneration"],
+    "modelTypePatterns": ["qwen3_vl"]
+  }
+}