@simulatte/doppler 0.1.4 → 0.1.5

package/src/config/kernels/registry.json

@@ -5537,6 +5537,228 @@
         }
       }
     },
+    "depthwise_conv2d": {
+      "description": "Depthwise Conv2D (NCHW)",
+      "baseBindings": [
+        {
+          "index": 0,
+          "name": "uniforms",
+          "type": "uniform"
+        },
+        {
+          "index": 1,
+          "name": "input",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 2,
+          "name": "weight",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 3,
+          "name": "bias",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 4,
+          "name": "output",
+          "type": "storage"
+        }
+      ],
+      "baseUniforms": {
+        "size": 48,
+        "fields": [
+          {
+            "name": "channels",
+            "type": "u32",
+            "offset": 0
+          },
+          {
+            "name": "height",
+            "type": "u32",
+            "offset": 4
+          },
+          {
+            "name": "width",
+            "type": "u32",
+            "offset": 8
+          },
+          {
+            "name": "out_height",
+            "type": "u32",
+            "offset": 12
+          },
+          {
+            "name": "out_width",
+            "type": "u32",
+            "offset": 16
+          },
+          {
+            "name": "kernel_h",
+            "type": "u32",
+            "offset": 20
+          },
+          {
+            "name": "kernel_w",
+            "type": "u32",
+            "offset": 24
+          },
+          {
+            "name": "stride",
+            "type": "u32",
+            "offset": 28
+          },
+          {
+            "name": "pad",
+            "type": "u32",
+            "offset": 32
+          },
+          {
+            "name": "_pad0",
+            "type": "u32",
+            "offset": 36
+          },
+          {
+            "name": "_pad1",
+            "type": "u32",
+            "offset": 40
+          },
+          {
+            "name": "_pad2",
+            "type": "u32",
+            "offset": 44
+          }
+        ]
+      },
+      "variants": {
+        "default": {
+          "wgsl": "depthwise_conv2d.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [],
+          "outputDtype": "f32"
+        },
+        "default_f16": {
+          "wgsl": "depthwise_conv2d_f16.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [
+            "shader-f16"
+          ],
+          "outputDtype": "f16"
+        }
+      }
+    },
+    "grouped_pointwise_conv2d": {
+      "description": "Grouped Pointwise Conv2D (NCHW)",
+      "baseBindings": [
+        {
+          "index": 0,
+          "name": "uniforms",
+          "type": "uniform"
+        },
+        {
+          "index": 1,
+          "name": "input",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 2,
+          "name": "weight",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 3,
+          "name": "bias",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 4,
+          "name": "output",
+          "type": "storage"
+        }
+      ],
+      "baseUniforms": {
+        "size": 32,
+        "fields": [
+          {
+            "name": "in_channels",
+            "type": "u32",
+            "offset": 0
+          },
+          {
+            "name": "out_channels",
+            "type": "u32",
+            "offset": 4
+          },
+          {
+            "name": "height",
+            "type": "u32",
+            "offset": 8
+          },
+          {
+            "name": "width",
+            "type": "u32",
+            "offset": 12
+          },
+          {
+            "name": "groups",
+            "type": "u32",
+            "offset": 16
+          },
+          {
+            "name": "_pad0",
+            "type": "u32",
+            "offset": 20
+          },
+          {
+            "name": "_pad1",
+            "type": "u32",
+            "offset": 24
+          },
+          {
+            "name": "_pad2",
+            "type": "u32",
+            "offset": 28
+          }
+        ]
+      },
+      "variants": {
+        "default": {
+          "wgsl": "grouped_pointwise_conv2d.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [],
+          "outputDtype": "f32"
+        },
+        "default_f16": {
+          "wgsl": "grouped_pointwise_conv2d_f16.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [
+            "shader-f16"
+          ],
+          "outputDtype": "f16"
+        }
+      }
+    },
     "groupnorm_stats": {
       "description": "GroupNorm stats (mean/var) for each group",
       "baseBindings": [
@@ -6424,6 +6646,340 @@
         }
       }
     },
+    "sana_linear_attention_summary": {
+      "description": "Sana linear attention summary stage",
+      "baseBindings": [
+        {
+          "index": 0,
+          "name": "uniforms",
+          "type": "uniform"
+        },
+        {
+          "index": 1,
+          "name": "query",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 2,
+          "name": "key",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 3,
+          "name": "value",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 4,
+          "name": "summary",
+          "type": "storage"
+        }
+      ],
+      "baseUniforms": {
+        "size": 24,
+        "fields": [
+          {
+            "name": "num_heads",
+            "type": "u32",
+            "offset": 0
+          },
+          {
+            "name": "head_dim",
+            "type": "u32",
+            "offset": 4
+          },
+          {
+            "name": "num_tokens",
+            "type": "u32",
+            "offset": 8
+          },
+          {
+            "name": "hidden_size",
+            "type": "u32",
+            "offset": 12
+          },
+          {
+            "name": "_pad0",
+            "type": "u32",
+            "offset": 16
+          },
+          {
+            "name": "_pad1",
+            "type": "u32",
+            "offset": 20
+          }
+        ]
+      },
+      "variants": {
+        "default": {
+          "wgsl": "sana_linear_attention_summary.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [],
+          "outputDtype": "f32"
+        },
+        "default_f16": {
+          "wgsl": "sana_linear_attention_summary_f16.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [
+            "shader-f16"
+          ],
+          "outputDtype": "f32"
+        }
+      }
+    },
+    "sana_linear_attention_apply": {
+      "description": "Sana linear attention apply stage",
+      "baseBindings": [
+        {
+          "index": 0,
+          "name": "uniforms",
+          "type": "uniform"
+        },
+        {
+          "index": 1,
+          "name": "query",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 2,
+          "name": "summary",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 3,
+          "name": "output",
+          "type": "storage"
+        }
+      ],
+      "baseUniforms": {
+        "size": 32,
+        "fields": [
+          {
+            "name": "num_heads",
+            "type": "u32",
+            "offset": 0
+          },
+          {
+            "name": "head_dim",
+            "type": "u32",
+            "offset": 4
+          },
+          {
+            "name": "num_tokens",
+            "type": "u32",
+            "offset": 8
+          },
+          {
+            "name": "hidden_size",
+            "type": "u32",
+            "offset": 12
+          },
+          {
+            "name": "eps",
+            "type": "f32",
+            "offset": 16
+          },
+          {
+            "name": "_pad0",
+            "type": "u32",
+            "offset": 20
+          },
+          {
+            "name": "_pad1",
+            "type": "u32",
+            "offset": 24
+          },
+          {
+            "name": "_pad2",
+            "type": "u32",
+            "offset": 28
+          }
+        ]
+      },
+      "variants": {
+        "default": {
+          "wgsl": "sana_linear_attention_apply.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [],
+          "outputDtype": "f32"
+        },
+        "default_f16": {
+          "wgsl": "sana_linear_attention_apply_f16.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [
+            "shader-f16"
+          ],
+          "outputDtype": "f16"
+        }
+      }
+    },
+    "repeat_channels": {
+      "description": "Repeat channels along the NCHW channel axis",
+      "baseBindings": [
+        {
+          "index": 0,
+          "name": "uniforms",
+          "type": "uniform"
+        },
+        {
+          "index": 1,
+          "name": "input",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 2,
+          "name": "output",
+          "type": "storage"
+        }
+      ],
+      "baseUniforms": {
+        "size": 20,
+        "fields": [
+          {
+            "name": "in_channels",
+            "type": "u32",
+            "offset": 0
+          },
+          {
+            "name": "height",
+            "type": "u32",
+            "offset": 4
+          },
+          {
+            "name": "width",
+            "type": "u32",
+            "offset": 8
+          },
+          {
+            "name": "repeats",
+            "type": "u32",
+            "offset": 12
+          },
+          {
+            "name": "_pad0",
+            "type": "u32",
+            "offset": 16
+          }
+        ]
+      },
+      "variants": {
+        "default": {
+          "wgsl": "repeat_channels.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [],
+          "outputDtype": "f32"
+        },
+        "default_f16": {
+          "wgsl": "repeat_channels_f16.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [
+            "shader-f16"
+          ],
+          "outputDtype": "f16"
+        }
+      }
+    },
+    "relu": {
+      "description": "ReLU activation",
+      "baseBindings": [
+        {
+          "index": 0,
+          "name": "uniforms",
+          "type": "uniform"
+        },
+        {
+          "index": 1,
+          "name": "input",
+          "type": "read-only-storage"
+        },
+        {
+          "index": 2,
+          "name": "output",
+          "type": "storage"
+        }
+      ],
+      "baseUniforms": {
+        "size": 16,
+        "fields": [
+          {
+            "name": "size",
+            "type": "u32",
+            "offset": 0
+          },
+          {
+            "name": "_pad0",
+            "type": "u32",
+            "offset": 4
+          },
+          {
+            "name": "_pad1",
+            "type": "u32",
+            "offset": 8
+          },
+          {
+            "name": "_pad2",
+            "type": "u32",
+            "offset": 12
+          }
+        ]
+      },
+      "variants": {
+        "default": {
+          "wgsl": "relu.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [],
+          "outputDtype": "f32"
+        },
+        "default_f16": {
+          "wgsl": "relu_f16.wgsl",
+          "entryPoint": "main",
+          "workgroup": [
+            256,
+            1,
+            1
+          ],
+          "requires": [
+            "shader-f16"
+          ],
+          "outputDtype": "f16"
+        }
+      }
+    },
     "conv2d_backward_input": {
       "description": "Conv2D backward (input gradient)",
       "baseBindings": [

package/src/config/loader.js

@@ -5,6 +5,7 @@ import {
 } from './schema/index.js';
 import { createDopplerError, ERROR_CODES } from '../errors/index.js';
 import { loadJson } from '../utils/load-json.js';
+import { chooseNullish, mergeLayeredShallowObjects } from './merge-helpers.js';
 
 // Static imports keep presets bundled for browser use.
 const transformerPreset = await loadJson('./presets/models/transformer.json', import.meta.url, 'Failed to load preset');
@@ -190,18 +191,21 @@ export function resolveConfig(
   // Note: Uses nullish coalesce (??) so null values fall through to next level.
   // This means explicit null in manifest = "use preset/default".
   const presetArch = preset.architecture || {};
-  const numLayers = manifestArch.numLayers ?? presetArch.numLayers;
-  const hiddenSize = manifestArch.hiddenSize ?? presetArch.hiddenSize;
-  const intermediateSize = manifestArch.intermediateSize ?? presetArch.intermediateSize;
-  const numAttentionHeads = manifestArch.numAttentionHeads ?? presetArch.numAttentionHeads;
-  const numKeyValueHeads = manifestArch.numKeyValueHeads ?? presetArch.numKeyValueHeads ?? numAttentionHeads;
-  const headDim = manifestArch.headDim ?? presetArch.headDim ?? (
-    hiddenSize && numAttentionHeads ? hiddenSize / numAttentionHeads : undefined
+  const numLayers = chooseNullish(manifestArch.numLayers, presetArch.numLayers);
+  const hiddenSize = chooseNullish(manifestArch.hiddenSize, presetArch.hiddenSize);
+  const intermediateSize = chooseNullish(manifestArch.intermediateSize, presetArch.intermediateSize);
+  const numAttentionHeads = chooseNullish(manifestArch.numAttentionHeads, presetArch.numAttentionHeads);
+  const numKeyValueHeads = chooseNullish(
+    manifestArch.numKeyValueHeads,
+    chooseNullish(presetArch.numKeyValueHeads, numAttentionHeads)
   );
-  const vocabSize = manifestArch.vocabSize ?? presetArch.vocabSize;
-  const maxSeqLen = manifestArch.maxSeqLen ?? presetArch.maxSeqLen;
-  const ropeTheta = manifestArch.ropeTheta ?? presetArch.ropeTheta;
-  const rmsNormEps = manifestArch.rmsNormEps ?? presetArch.rmsNormEps;
+  const headDim = chooseNullish(manifestArch.headDim, chooseNullish(presetArch.headDim, (
+    hiddenSize && numAttentionHeads ? hiddenSize / numAttentionHeads : undefined
+  )));
+  const vocabSize = chooseNullish(manifestArch.vocabSize, presetArch.vocabSize);
+  const maxSeqLen = chooseNullish(manifestArch.maxSeqLen, presetArch.maxSeqLen);
+  const ropeTheta = chooseNullish(manifestArch.ropeTheta, presetArch.ropeTheta);
+  const rmsNormEps = chooseNullish(manifestArch.rmsNormEps, presetArch.rmsNormEps);
 
   const architecture = {
     numLayers,
@@ -226,44 +230,44 @@ export function resolveConfig(
   const manifestInference = extractInferenceFromConfig(manifest.config || {});
 
   const inference = {
-    attention: {
-      ...baseInference.attention,
-      ...presetInference.attention,
-      ...manifestInference.attention,
-    },
-    normalization: {
-      ...baseInference.normalization,
-      ...presetInference.normalization,
-    },
-    ffn: {
-      ...baseInference.ffn,
-      ...presetInference.ffn,
-    },
-    moe: presetInference.moe ?? baseInference.moe ?? null,
-    output: {
-      ...baseInference.output,
-      ...presetInference.output,
-      ...manifestInference.output,
-    },
-    layerPattern: presetInference.layerPattern ?? baseInference.layerPattern,
-    rope: {
-      ...baseInference.rope,
-      ...presetInference.rope,
-      ...manifestInference.rope,
-    },
-    pipeline: manifestInference.pipeline ?? presetInference.pipeline ?? baseInference.pipeline,
-    chatTemplate: {
-      ...baseInference.chatTemplate,
-      ...presetInference.chatTemplate,
-    },
-    kernelPath: presetInference.kernelPath ?? baseInference.kernelPath,
+    attention: mergeLayeredShallowObjects(
+      baseInference.attention,
+      presetInference.attention,
+      manifestInference.attention
+    ),
+    normalization: mergeLayeredShallowObjects(
+      baseInference.normalization,
+      presetInference.normalization
+    ),
+    ffn: mergeLayeredShallowObjects(
+      baseInference.ffn,
+      presetInference.ffn
+    ),
+    moe: chooseNullish(presetInference.moe, chooseNullish(baseInference.moe, null)),
+    output: mergeLayeredShallowObjects(
+      baseInference.output,
+      presetInference.output,
+      manifestInference.output
+    ),
+    layerPattern: chooseNullish(presetInference.layerPattern, baseInference.layerPattern),
+    rope: mergeLayeredShallowObjects(
+      baseInference.rope,
+      presetInference.rope,
+      manifestInference.rope
+    ),
+    pipeline: chooseNullish(manifestInference.pipeline, chooseNullish(presetInference.pipeline, baseInference.pipeline)),
+    chatTemplate: mergeLayeredShallowObjects(
+      baseInference.chatTemplate,
+      presetInference.chatTemplate
+    ),
+    kernelPath: chooseNullish(presetInference.kernelPath, baseInference.kernelPath),
   };
 
   // Merge tokenizer config
-  const tokenizer = {
-    ...preset.tokenizer,
-    ...extractTokenizerFromManifest(manifest),
-  };
+  const tokenizer = mergeLayeredShallowObjects(
+    preset.tokenizer,
+    extractTokenizerFromManifest(manifest)
+  );
 
   // Sampling defaults
   const sampling = preset.sampling ?? {

package/src/config/merge-contract-check.d.ts

@@ -0,0 +1,16 @@
+export interface MergeContractCheckResult {
+  id: string;
+  ok: boolean;
+  detail: string;
+  mode: 'actual' | 'modeled';
+}
+
+export interface MergeContractArtifact {
+  schemaVersion: 1;
+  source: 'doppler';
+  ok: boolean;
+  checks: MergeContractCheckResult[];
+  errors: string[];
+}
+
+export declare function buildMergeContractArtifact(): MergeContractArtifact;