@simulatte/doppler 0.1.3 → 0.1.5

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

@@ -1,6 +1,8 @@
 import { DEFAULT_ENTRY } from './schema/kernel-path.schema.js';
 import { KERNEL_CONFIGS } from '../gpu/kernels/utils.js';
+import { selectByRules } from '../gpu/kernels/rule-matcher.js';
 import { loadJson } from '../utils/load-json.js';
+import { buildKernelPathContractArtifact } from './kernel-path-contract-check.js';
 
 // =============================================================================
 // Built-in Kernel Paths (imported at build time)
@@ -77,6 +79,11 @@ const KERNEL_PATH_REGISTRY_INDEX = new Map(
 );
 
 const KERNEL_PATH_REGISTRY = Object.create(null);
+const KERNEL_PATH_RULES = await loadJson(
+  '../rules/inference/kernel-path.rules.json',
+  import.meta.url,
+  'Failed to load kernel path rules'
+);
 
 const resolveKernelPathConfig = (id, chain = new Set()) => {
   if (KERNEL_PATH_REGISTRY[id] !== undefined) {
@@ -117,6 +124,42 @@ for (const entry of KERNEL_PATH_REGISTRY_ENTRIES) {
   resolveKernelPathConfig(entry.id);
 }
 
+const KERNEL_PATH_FINITENESS_FALLBACK_MAPPINGS = KERNEL_PATH_REGISTRY_ENTRIES
+  .map((entry) => {
+    const fallbackKernelPathId = selectByRules(
+      Array.isArray(KERNEL_PATH_RULES?.finitenessFallback) ? KERNEL_PATH_RULES.finitenessFallback : [],
+      { kernelPathId: entry.id }
+    );
+    if (typeof fallbackKernelPathId !== 'string' || fallbackKernelPathId.length === 0) {
+      return null;
+    }
+    return {
+      primaryKernelPathId: entry.id,
+      fallbackKernelPathId,
+      primaryActivationDtype: KERNEL_PATH_REGISTRY[entry.id]?.activationDtype ?? null,
+      fallbackActivationDtype: KERNEL_PATH_REGISTRY[fallbackKernelPathId]?.activationDtype ?? null,
+    };
+  })
+  .filter(Boolean);
+
+const KERNEL_PATH_CONTRACT_ARTIFACT = buildKernelPathContractArtifact(
+  {
+    registryId: 'builtin-kernel-paths',
+    entries: KERNEL_PATH_REGISTRY_ENTRIES,
+    fallbackMappings: KERNEL_PATH_FINITENESS_FALLBACK_MAPPINGS,
+    fallbackRules: Array.isArray(KERNEL_PATH_RULES?.finitenessFallback)
+      ? KERNEL_PATH_RULES.finitenessFallback
+      : [],
+    autoSelectRules: Array.isArray(KERNEL_PATH_RULES?.autoSelect)
+      ? KERNEL_PATH_RULES.autoSelect
+      : [],
+  }
+);
+
+if (!KERNEL_PATH_CONTRACT_ARTIFACT.ok) {
+  throw new Error(KERNEL_PATH_CONTRACT_ARTIFACT.errors[0]);
+}
+
 // =============================================================================
 // Public API
 // =============================================================================
@@ -129,6 +172,17 @@ export function listKernelPaths() {
   return Object.keys(KERNEL_PATH_REGISTRY);
 }
 
+export function getKernelPathContractArtifact() {
+  return {
+    schemaVersion: KERNEL_PATH_CONTRACT_ARTIFACT.schemaVersion,
+    source: KERNEL_PATH_CONTRACT_ARTIFACT.source,
+    ok: KERNEL_PATH_CONTRACT_ARTIFACT.ok,
+    checks: KERNEL_PATH_CONTRACT_ARTIFACT.checks.map((entry) => ({ ...entry })),
+    errors: [...KERNEL_PATH_CONTRACT_ARTIFACT.errors],
+    stats: { ...KERNEL_PATH_CONTRACT_ARTIFACT.stats },
+  };
+}
+
 export function resolveKernelPath(ref) {
   if (typeof ref === 'string') {
     const path = getKernelPath(ref);

package/src/config/kernels/kernel-ref-digests.js

@@ -57,6 +57,8 @@ export const KERNEL_REF_CONTENT_DIGESTS = Object.freeze({
   "conv2d_f16.wgsl#main": "aa139e9f0270873acbc1c4b3cbacff4d224cae7247b520ec129a4f068eb6ed59",
   "conv2d.wgsl#main": "484a676692d2b8097daeefe42e2296a1f8b3ef11abfd7b41df6cdcdf16b7a8fd",
   "cross_entropy_loss.wgsl#main": "5a48087bdec94184432c90ce5b345e1eadbdfcb13b9793ecee8052bc7392239c",
+  "depthwise_conv2d_f16.wgsl#main": "d5d8d195b1449e39715340af4a0759da4b44b54f6a3cfbdfa6abe743b0f1d002",
+  "depthwise_conv2d.wgsl#main": "e5da160f505e18508619b78ba30f9bde0c84689a166df06cb59ef0e6591c6faf",
   "dequant_f16_out_vec4.wgsl#main_vec4": "61c20e6c71c1c8421b4ec202dbd26292a6300587bd44c314f2a6c6d9d9442c3a",
   "dequant_f16_out.wgsl#main": "94d61843d56f9a3bbc6b7c2b95dc6ecbba3f6a262b2c4086a076f69a8c38ccae",
   "dequant_f16_rowwise.wgsl#main": "f5bf7cef950b52d65cee6121dbaa176244d3221045b3b6386b3be47f23ce17dc",
@@ -116,6 +118,8 @@ export const KERNEL_REF_CONTENT_DIGESTS = Object.freeze({
   "gelu.wgsl#main": "a9007ea08aaff98f9be08f1e0490a6bcf252883eac5513de876ab9ce918865e6",
   "gptoss_mxfp4_expert_fused.wgsl#main_expert": "3159e8cd81da13f909cf905e6d35307fefe1dcbbdf1b2b8e8ff0ce923bd71180",
   "gptoss_router_topk.wgsl#softmax_topk": "86e4ea709c0c0084d09c6a4cd07710dc14f380e03f91b8ed9ec871b310be49f1",
+  "grouped_pointwise_conv2d_f16.wgsl#main": "11bcaefc5929b2e3c1ba338ebea6a28d2cac26553be8b00f51bfddbabf513be7",
+  "grouped_pointwise_conv2d.wgsl#main": "c0d5cdec0743b4ee337a8df95bda442e617c1678e3d1b6e20ec692d500ede50d",
   "groupnorm_apply_f16.wgsl#main": "cfd850b87944ac1c03ba7bd98136db556dadd8a70611e351d82d297299a7cd02",
   "groupnorm_apply.wgsl#main": "b09b8f2f57dcdfa1a0366daa30d3910feb134204652c711d2ba564e566b5a334",
   "groupnorm_stats_f16.wgsl#main": "fb76f78ce668ea8459110335698fe4b09a2425fc71deed3bab67efd7641c3199",
@@ -153,6 +157,10 @@ export const KERNEL_REF_CONTENT_DIGESTS = Object.freeze({
   "moe_offsets.wgsl#build_offsets": "3ea004145fa234659408cdeb0d4d802adff1037c9c5c03af146b3734cc69dd27",
   "pixel_shuffle_f16.wgsl#main": "57903a9c19cecc56371b2198402745127115680d266c3ce609201be9119aa359",
   "pixel_shuffle.wgsl#main": "845b88700b1b46d18cde6f2ec11bb89512c90d7e148763e74ce2a4173fd99b21",
+  "relu_f16.wgsl#main": "fc6134aabe43081b42ce8507d8f374092d0f2e03316aa42c25dd50229dc0ee40",
+  "relu.wgsl#main": "ca2c9bfa0acb9ece3b7e67de5209e00e553602b3917d23aca10338c1e6f01e27",
+  "repeat_channels_f16.wgsl#main": "e7e4d9164752e782d482db40256d0d86d96f784aa7debdb72faf3261b9bdd737",
+  "repeat_channels.wgsl#main": "ad0e34925c8c1173b9f0d92fa6e3808d039f82b3d9ad943b0a75b213ee1776e5",
   "residual_f16_vec4.wgsl#add_vec4": "30e9226fb6636e2f01e65b1dc8e93c8e849a87acec6215342fc114996da1ed41",
   "residual_f16.wgsl#main": "d392433f3065d1caf68b033219f4ffacf022dc1f90fc3cf3fd620e4ba49f3219",
   "residual_vec4.wgsl#add_vec4": "ef011d1683e62887db712da563e783d12fdc80c152955661137d2dca612d7d6a",
@@ -186,6 +194,10 @@ export const KERNEL_REF_CONTENT_DIGESTS = Object.freeze({
   "sample.wgsl#find_topk_phase2": "940b216e605d22096da5aca65950a8030866fc5a39e7fdf484d69a832de1b63a",
   "sample.wgsl#sample_single_pass": "4412357e84113ee2f1bc0dc8bf89e314c2ab482c89c14ca016ea9949d16a9d0c",
   "sample.wgsl#softmax_and_sample": "7172c60e76430fbe130e530e3564b569b45eccf193987b32d6f52bd6bbcc9f08",
+  "sana_linear_attention_apply_f16.wgsl#main": "e47366b94d40c4388e631b5bf93f8d61ef4e52cc65ffcd3b08d9d170616bb138",
+  "sana_linear_attention_apply.wgsl#main": "59cad7974c644fd910af776ad85a9a2c43c00492d4d1152fdc8373ecbb8bba18",
+  "sana_linear_attention_summary_f16.wgsl#main": "e3c040bb6469d37fc78eb22c1cc3e0456301607e461bbcdf5365a583c5d260d2",
+  "sana_linear_attention_summary.wgsl#main": "20c7ecdbcd1c73c0f9937c3cdac07b4b6edfe8618bf6f66281806343fd41b122",
   "scale.wgsl#main": "44ec481452b586307957163e3d65c9d02561d3f2f3db633f906f5488b1ea1ca4",
   "scale.wgsl#main_inplace": "020824c7118a59c461ce81f1c2cd01b7c2a3f1aab326392b7d48d4448a0c2ed1",
   "scatter_add_dynamic_f16_weights.wgsl#scatter_add_dynamic": "42799e745bc445b199b1cbc384bc12bb9372ed1599af3260a803cefc8dd35497",

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": [