@simulatte/doppler 0.1.8 → 0.2.0

package/src/rules/inference/dtype.rules.json

@@ -1,99 +0,0 @@
-{
-  "f16OrF32": [
-    { "match": { "useF16": true }, "value": "f16" },
-    { "match": {}, "value": "f32" }
-  ],
-  "dtypeFromAlias": [
-    { "match": { "dtype": "bf16" }, "value": "bf16" },
-    { "match": { "dtype": "f16" }, "value": "f16" },
-    { "match": { "dtype": "fp16" }, "value": "f16" },
-    { "match": { "dtype": "float16" }, "value": "f16" },
-    { "match": { "dtype": "f32" }, "value": "f32" },
-    { "match": { "dtype": "fp32" }, "value": "f32" },
-    { "match": { "dtype": "float32" }, "value": "f32" },
-    { "match": {}, "value": { "context": "fallback" } }
-  ],
-  "f16OrF32FromDtype": [
-    { "match": { "dtype": "f16" }, "value": "f16" },
-    { "match": {}, "value": "f32" }
-  ],
-  "f16OrF32FromDtypeAlias": [
-    { "match": { "dtype": "f16" }, "value": "f16" },
-    { "match": { "dtype": "bf16" }, "value": "f16" },
-    { "match": { "dtype": "fp16" }, "value": "f16" },
-    { "match": { "dtype": "float16" }, "value": "f16" },
-    { "match": { "dtype": "f32" }, "value": "f32" },
-    { "match": { "dtype": "fp32" }, "value": "f32" },
-    { "match": { "dtype": "float32" }, "value": "f32" },
-    { "match": {}, "value": { "context": "fallback" } }
-  ],
-  "f16OrF32FromBytes": [
-    { "match": { "bytesPerElement": { "lte": 2 } }, "value": "f16" },
-    { "match": {}, "value": "f32" }
-  ],
-  "f16OrF32FromBytesOrFallback": [
-    { "match": { "bytesPerElement": { "lt": 3 } }, "value": "f16" },
-    { "match": {}, "value": { "context": "fallback" } }
-  ],
-  "f16OrF32FromToken": [
-    { "match": { "token": "F16" }, "value": "f16" },
-    { "match": {}, "value": "f32" }
-  ],
-  "f16IfCap": [
-    { "match": { "hasF16": true }, "value": "f16" },
-    { "match": {}, "value": "f32" }
-  ],
-  "bothF16": [
-    { "match": { "aDtype": "f16", "bDtype": "f16" }, "value": "f16" },
-    { "match": {}, "value": "f32" }
-  ],
-  "diffusionActivationDtype": [
-    { "match": { "requested": "f16", "hasF16": true }, "value": "f16" },
-    { "match": {}, "value": "f32" }
-  ],
-  "f16OrFallback": [
-    { "match": { "kvDtype": "f16" }, "value": "f16" },
-    { "match": {}, "value": { "context": "fallback" } }
-  ],
-  "f16OrFallbackByFlag": [
-    { "match": { "useF16": true }, "value": "f16" },
-    { "match": {}, "value": { "context": "fallback" } }
-  ],
-  "attentionProjectionOutputDtype": [
-    { "match": { "forceF32": true }, "value": "f32" },
-    { "match": { "useF16": true }, "value": "f16" },
-    { "match": {}, "value": { "context": "fallback" } }
-  ],
-  "bytesPerElement": [
-    { "match": { "dtype": "f16" }, "value": 2 },
-    { "match": {}, "value": 4 }
-  ],
-  "dtypeFromSize": [
-    { "match": { "bytesPerElement": 2 }, "value": "f16" },
-    { "match": {}, "value": "f32" }
-  ],
-  "bytesFromDtype": [
-    { "match": { "dtype": "f16" }, "value": 2 },
-    { "match": {}, "value": 4 }
-  ],
-  "matmulDtype": [
-    { "match": { "dtype": { "in": ["f16", "bf16"] } }, "value": "f16" },
-    { "match": { "dtype": "q4k" }, "value": "q4k" },
-    { "match": {}, "value": "f32" }
-  ],
-  "logitsDtype": [
-    { "match": { "logitsDtype": "f16" }, "value": "f16" },
-    { "match": {}, "value": "f32" }
-  ],
-  "kvCacheDtype": [
-    { "match": { "forceF32": true }, "value": "f32" },
-    { "match": { "requested": "f16", "useGPU": true, "hasF16": true }, "value": "f16" },
-    { "match": { "requested": "f16" }, "value": "f32" },
-    { "match": { "requested": "f32" }, "value": "f32" },
-    { "match": {}, "value": "f32" }
-  ],
-  "lmHeadChunkWeightDtype": [
-    { "match": { "preferF16": true, "lmHeadDtype": "f16", "hasF16": true }, "value": "f16" },
-    { "match": {}, "value": "f32" }
-  ]
-}

package/src/rules/inference/execution.rules.json

@@ -1,45 +0,0 @@
-{
-  "decodeRecorderEnabled": [
-    {
-      "match": {
-        "hasDevice": true,
-        "debug": false,
-        "disableCommandBatching": false,
-        "kvLayout": { "neq": "bdpa_paged" }
-      },
-      "value": true
-    },
-    { "match": {}, "value": false }
-  ],
-  "batchDecodeEnabled": [
-    {
-      "match": {
-        "batchSize": { "gt": 1 },
-        "useGPU": true,
-        "gpuSamplingAvailable": true,
-        "disableMultiTokenDecode": false,
-        "disableCommandBatching": false,
-        "isBdpaPagedLayout": false,
-        "finitenessFallbackWindowOpen": false
-      },
-      "value": true
-    },
-    { "match": {}, "value": false }
-  ],
-  "finitenessFallbackActivationDtype": [
-    { "match": {}, "value": "f32" }
-  ],
-  "modelBatchingDefaults": [
-    {
-      "match": {
-        "presetId": { "in": ["lfm2"] }
-      },
-      "value": {
-        "batchSize": 8,
-        "stopCheckMode": "batch",
-        "readbackInterval": 8
-      }
-    },
-    { "match": {}, "value": null }
-  ]
-}

package/src/rules/inference/ffn.rules.json

@@ -1,35 +0,0 @@
-{
-  "activationOp": [
-    {
-      "match": {
-        "hiddenActivation": "gelu"
-      },
-      "value": "gelu"
-    },
-    {
-      "match": {},
-      "value": "silu"
-    }
-  ],
-  "useFusedGateUp": [
-    {
-      "match": {
-        "hasGate": true,
-        "hasUp": true,
-        "hasDown": true,
-        "hasFusedWeights": false,
-        "inputIsSupported": true,
-        "hasLoRA": false,
-        "dtypeMatches": true,
-        "dtypeSupported": true,
-        "activationDtype": "f16",
-        "f16BatchSupported": true
-      },
-      "value": true
-    },
-    {
-      "match": {},
-      "value": false
-    }
-  ]
-}

package/src/rules/inference/kernel-path.rules.json

@@ -1,92 +0,0 @@
-{
-  "autoSelect": [
-    {
-      "match": {
-        "allowCapabilityAutoSelection": true,
-        "hasSubgroups": true,
-        "kernelPathRef": "gemma3-q4k-dequant-f32a-nosubgroups"
-      },
-      "value": "gemma3-q4k-dequant-f32a-online"
-    },
-    {
-      "match": {
-        "allowCapabilityAutoSelection": true,
-        "hasSubgroups": true,
-        "kernelPathRef": "gemma3-q4k-dequant-f32a"
-      },
-      "value": "gemma3-q4k-dequant-f32a-online"
-    },
-    {
-      "match": {
-        "allowCapabilityAutoSelection": true,
-        "hasSubgroups": false,
-        "kernelPathRef": "gemma3-q4k-dequant-f32a"
-      },
-      "value": "gemma3-q4k-dequant-f32a-nosubgroups"
-    },
-    {
-      "match": {
-        "allowCapabilityAutoSelection": true,
-        "hasSubgroups": false,
-        "kernelPathRef": "gemma3-q4k-dequant-f16a-online"
-      },
-      "value": "gemma3-q4k-dequant-f32a-nosubgroups"
-    },
-    {
-      "match": {
-        "allowCapabilityAutoSelection": true,
-        "hasSubgroups": false,
-        "kernelPathRef": "gemma3-q4k-dequant-f32a-online"
-      },
-      "value": "gemma3-q4k-dequant-f32a-nosubgroups"
-    },
-    {
-      "match": {
-        "allowCapabilityAutoSelection": true,
-        "hasSubgroups": false,
-        "kernelPathRef": "lfm2-q4k-dequant-f32a-online"
-      },
-      "value": "lfm2-q4k-dequant-f32a-nosubgroups"
-    },
-    {
-      "match": {
-        "allowCapabilityAutoSelection": true,
-        "hasSubgroups": false,
-        "kernelPathRef": "gemma2-q4k-dequant-f16a"
-      },
-      "value": "gemma2-q4k-dequant-f32a-nosubgroups"
-    },
-    {
-      "match": {
-        "allowCapabilityAutoSelection": true,
-        "hasSubgroups": false,
-        "kernelPathRef": "gemma2-q4k-fused-f32a"
-      },
-      "value": "gemma2-q4k-dequant-f32a-nosubgroups"
-    },
-    { "match": {}, "value": { "context": "kernelPathRef" } }
-  ],
-  "finitenessFallback": [
-    {
-      "match": { "kernelPathId": "gemma3-f16-fused-f16a-online" },
-      "value": "gemma3-f16-fused-f32a-online"
-    },
-    {
-      "match": { "kernelPathId": "gemma3-q4k-dequant-f16a-online" },
-      "value": "gemma3-q4k-dequant-f32a-nosubgroups"
-    },
-    {
-      "match": { "kernelPathId": "lfm2-q4k-dequant-f32a-online" },
-      "value": "lfm2-q4k-dequant-f32a-nosubgroups"
-    },
-    {
-      "match": { "kernelPathId": "gemma2-f16-f16a" },
-      "value": "gemma2-f16-f32a"
-    },
-    {
-      "match": { "kernelPathId": "gemma2-q4k-dequant-f16a" },
-      "value": "gemma2-q4k-dequant-f32a-nosubgroups"
-    },
-    { "match": {}, "value": null }
-  ]
-}

package/src/rules/inference/layer-pattern.rules.json

@@ -1,16 +0,0 @@
-{
-  "patternKind": [
-    { "match": { "patternType": "alternating", "globalPattern": "even" }, "value": "alternating_even" },
-    { "match": { "patternType": "alternating", "globalPattern": "odd" }, "value": "alternating_odd" },
-    { "match": { "patternType": "every_n" }, "value": "every_n" },
-    { "match": {}, "value": null }
-  ],
-  "layerType": [
-    { "match": { "patternKind": "alternating_even", "isEven": true }, "value": "full_attention" },
-    { "match": { "patternKind": "alternating_even" }, "value": "sliding_attention" },
-    { "match": { "patternKind": "alternating_odd", "isEven": false }, "value": "full_attention" },
-    { "match": { "patternKind": "alternating_odd" }, "value": "sliding_attention" },
-    { "match": { "patternKind": "every_n", "isStride": true }, "value": "full_attention" },
-    { "match": { "patternKind": "every_n" }, "value": "sliding_attention" }
-  ]
-}

package/src/rules/inference/layer.rules.json

@@ -1,7 +0,0 @@
-{
-  "ffnMode": [
-    { "match": { "variant": "moe" }, "value": true },
-    { "match": { "variant": "auto", "canAutoMoe": true }, "value": true },
-    { "match": {}, "value": false }
-  ]
-}

package/src/rules/inference/moe.rules.json

@@ -1,48 +0,0 @@
-{
-  "biasAddCode": [
-    {
-      "match": { "logitsDtype": "f16", "biasDtype": "f16" },
-      "value": {
-        "logitsType": "f16",
-        "biasType": "f16",
-        "logitsRead": "f32(logits[idx])",
-        "biasRead": "f32(bias[e])",
-        "logitsWrite": "logits[idx] = f16(value);",
-        "enableF16": "enable f16;\n"
-      }
-    },
-    {
-      "match": { "logitsDtype": "f16" },
-      "value": {
-        "logitsType": "f16",
-        "biasType": "f32",
-        "logitsRead": "f32(logits[idx])",
-        "biasRead": "bias[e]",
-        "logitsWrite": "logits[idx] = f16(value);",
-        "enableF16": "enable f16;\n"
-      }
-    },
-    {
-      "match": { "biasDtype": "f16" },
-      "value": {
-        "logitsType": "f32",
-        "biasType": "f16",
-        "logitsRead": "logits[idx]",
-        "biasRead": "f32(bias[e])",
-        "logitsWrite": "logits[idx] = value;",
-        "enableF16": "enable f16;\n"
-      }
-    },
-    {
-      "match": {},
-      "value": {
-        "logitsType": "f32",
-        "biasType": "f32",
-        "logitsRead": "logits[idx]",
-        "biasRead": "bias[e]",
-        "logitsWrite": "logits[idx] = value;",
-        "enableF16": ""
-      }
-    }
-  ]
-}

package/src/rules/kernels/attention.rules.json

@@ -1,61 +0,0 @@
-{
-  "tier": [
-    { "match": { "canSubgroup": true }, "value": "subgroup" },
-    { "match": { "canLarge": true }, "value": "tiled_large" },
-    { "match": { "canSmall": true }, "value": "tiled_small" },
-    { "match": { "isDecode": true }, "value": "streaming" },
-    { "match": {}, "value": "streaming" }
-  ],
-  "phase": [
-    { "match": { "isDecode": true }, "value": "decode" },
-    { "match": {}, "value": "prefill" }
-  ],
-  "suffix": [
-    { "match": { "useF16": true }, "value": "_f16" },
-    { "match": { "useF16KV": true }, "value": "_f16kv" },
-    { "match": {}, "value": "" }
-  ],
-  "chunkedVariant": [
-    { "match": { "useF16": true }, "value": "decode_chunked_f16" },
-    { "match": {}, "value": "decode_chunked_f16kv" }
-  ],
-  "pagedVariant": [
-    { "match": { "useF16": true }, "value": "decode_paged_f16" },
-    { "match": {}, "value": "decode_paged_f16kv" }
-  ],
-  "optimizedVariant": [
-    { "match": { "useF16": true }, "value": "decode_online_f16" },
-    { "match": {}, "value": "decode_online_f16kv" }
-  ],
-  "tieredVariant": [
-    { "match": { "useF16": true }, "value": "decode_tiered_f16" },
-    { "match": {}, "value": "decode_tiered_f16kv" }
-  ],
-  "tieredQuantVariant": [
-    { "match": { "mode": "int4" }, "value": "decode_tiered_int4_f16kv" },
-    { "match": { "mode": "int8" }, "value": "decode_tiered_int8_f16kv" },
-    { "match": {}, "value": "decode_tiered_int8_f16kv" }
-  ],
-  "variant": [
-    { "match": { "canUseDecodeOptimized": true }, "value": { "context": "optimizedVariant" } },
-    {
-      "match": {
-        "platformId": { "in": ["apple-m2", "apple-m3", "amd-rdna3", "nvidia-rtx30", "nvidia-rtx40"] },
-        "isDecode": true,
-        "tier": "subgroup",
-        "canUseDecodeSubgroup": true
-      },
-      "value": "decode_subgroup"
-    },
-    { "match": { "isPaged": true, "isDecode": true, "canUseChunked": true }, "value": { "context": "pagedVariant" } },
-    { "match": { "tier": "subgroup", "useF16KV": true, "canUseChunked": true }, "value": { "context": "chunkedVariant" } },
-    { "match": { "tier": "subgroup", "useF16KV": true }, "value": { "template": "prefill_streaming{suffix}" } },
-    { "match": { "tier": "subgroup", "canUseDecodeSubgroup": true }, "value": "decode_subgroup" },
-    { "match": { "tier": "subgroup" }, "value": "prefill_streaming" },
-    { "match": { "tier": "tiled_large" }, "value": { "template": "{base}{suffix}" } },
-    { "match": { "tier": "tiled_small" }, "value": { "template": "prefill_small{suffix}" } },
-    { "match": { "tier": "streaming", "canUseChunked": true }, "value": { "context": "chunkedVariant" } },
-    { "match": { "tier": "streaming" }, "value": { "template": "prefill_streaming{suffix}" } },
-    { "match": {}, "value": { "template": "prefill_streaming{suffix}" } }
-  ]
-}

package/src/rules/kernels/conv2d.rules.json

@@ -1,6 +0,0 @@
-{
-  "variant": [
-    { "match": { "isF16": true }, "value": "default_f16" },
-    { "match": {}, "value": "default" }
-  ]
-}

package/src/rules/kernels/depthwise-conv2d.rules.json

@@ -1,6 +0,0 @@
-{
-  "variant": [
-    { "match": { "isF16": true }, "value": "default_f16" },
-    { "match": {}, "value": "default" }
-  ]
-}

package/src/rules/kernels/dequant.rules.json

@@ -1,58 +0,0 @@
-{
-  "variant": [
-    {
-      "match": {
-        "platformId": { "in": ["apple-m2", "apple-m3", "nvidia-rtx30", "nvidia-rtx40"] },
-        "hasSubgroups": true,
-        "wantsF16Out": true
-      },
-      "value": "subgroup_f16out"
-    },
-    {
-      "match": {
-        "platformId": { "in": ["apple-m2", "apple-m3", "nvidia-rtx30", "nvidia-rtx40"] },
-        "hasSubgroups": true
-      },
-      "value": "subgroup"
-    },
-    { "match": { "hasSubgroups": true, "wantsF16Out": true, "useVec4": true }, "value": "subgroup_vec4_f16out" },
-    { "match": { "hasSubgroups": true, "wantsF16Out": true }, "value": "subgroup_f16out" },
-    { "match": { "hasSubgroups": true, "useVec4": true }, "value": "subgroup_vec4" },
-    { "match": { "hasSubgroups": true }, "value": "subgroup" },
-    { "match": { "wantsF16Out": true, "useVec4": true }, "value": "subgroup_vec4_f16out" },
-    { "match": { "wantsF16Out": true }, "value": "subgroup_f16out" },
-    { "match": { "useVec4": true }, "value": "shared_vec4" },
-    { "match": {}, "value": "shared" }
-  ],
-  "rowwiseVariant": [
-    { "match": { "wantsF16Out": true }, "value": "f16_rowwise" },
-    { "match": {}, "value": "f32_rowwise" }
-  ],
-  "mxfp4ExpertVariant": [
-    {
-      "match": {
-        "modelType": "gpt-oss",
-        "hasF16": true,
-        "hasSubgroups": true,
-        "outputDtype": "f16",
-        "dequantTileShape": "vec4",
-        "groupSize": 32
-      },
-      "value": "mxfp4_expert_fused_f16"
-    },
-    {
-      "match": {
-        "modelType": "gpt-oss",
-        "hasF16": true,
-        "outputDtype": "f16"
-      },
-      "value": "mxfp4_expert_f16"
-    },
-    {
-      "match": { "modelType": "gpt-oss" },
-      "value": "mxfp4_expert"
-    },
-    { "match": { "outputDtype": "f16" }, "value": "mxfp4_expert_f16" },
-    { "match": {}, "value": "mxfp4_expert" }
-  ]
-}

package/src/rules/kernels/energy.rules.json

@@ -1,22 +0,0 @@
-{
-  "evalVariant": [
-    { "match": { "isF16": true }, "value": "eval_f16" },
-    { "match": {}, "value": "eval" }
-  ],
-  "updateVariant": [
-    { "match": { "isF16": true }, "value": "update_f16" },
-    { "match": {}, "value": "update" }
-  ],
-  "quintelUpdateVariant": [
-    { "match": { "isF16": true }, "value": "quintel_update_f16" },
-    { "match": {}, "value": "quintel_update" }
-  ],
-  "quintelReduceVariant": [
-    { "match": { "isF16": true }, "value": "quintel_reduce_f16" },
-    { "match": {}, "value": "quintel_reduce" }
-  ],
-  "quintelGradVariant": [
-    { "match": { "isF16": true }, "value": "quintel_grad_f16" },
-    { "match": {}, "value": "quintel_grad" }
-  ]
-}

package/src/rules/kernels/fused-ffn.rules.json

@@ -1,13 +0,0 @@
-{
-  "variant": [
-    { "match": { "isQ4K": true, "fusedAllowed": true, "hiddenAligned": true, "batchSize": { "gt": 1 } }, "value": "q4k_batched" },
-    { "match": { "isQ4K": true, "fusedAllowed": true, "hiddenAligned": true }, "value": "q4k" },
-    { "match": { "weightDtype": "f16", "hasF16": true, "useF16Input": true, "batchSize": { "gt": 1 } }, "value": "f16_native_batched" },
-    { "match": { "weightDtype": "f16", "hasF16": true, "useF16Input": true }, "value": "f16_native" },
-    { "match": { "weightDtype": "f16", "batchSize": { "gt": 1 } }, "value": "batched" },
-    { "match": { "weightDtype": "f16" }, "value": "f16" },
-    { "match": { "batchSize": { "gt": 1 } }, "value": "batched" },
-    { "match": { "useMultiOutput": true }, "value": "multi" },
-    { "match": {}, "value": "default" }
-  ]
-}

package/src/rules/kernels/fused-matmul-residual.rules.json

@@ -1,6 +0,0 @@
-{
-  "variant": [
-    { "match": { "dtype": "f16" }, "value": "f16" },
-    { "match": {}, "value": "default" }
-  ]
-}

package/src/rules/kernels/fused-matmul-rmsnorm.rules.json

@@ -1,8 +0,0 @@
-{
-  "variant": [
-    { "match": { "isSmall": true, "isF16": true }, "value": "small_f16" },
-    { "match": { "isSmall": true }, "value": "small" },
-    { "match": { "isF16": true }, "value": "medium_f16" },
-    { "match": {}, "value": "medium" }
-  ]
-}

package/src/rules/kernels/gather.rules.json

@@ -1,12 +0,0 @@
-{
-  "variant": [
-    { "match": { "useF16Input": true, "useF16Output": true, "useVec4": true }, "value": "f16_vec4_f16_out" },
-    { "match": { "useF16Input": true, "useF16Output": true }, "value": "f16_f16_out" },
-    { "match": { "useF16Input": true, "useVec4": true }, "value": "f16_vec4" },
-    { "match": { "useF16Input": true }, "value": "f16" },
-    { "match": { "useF16Output": true, "useVec4": true }, "value": "vec4_f16_out" },
-    { "match": { "useF16Output": true }, "value": "f16_out" },
-    { "match": { "useVec4": true }, "value": "vec4" },
-    { "match": {}, "value": "default" }
-  ]
-}

package/src/rules/kernels/gelu.rules.json

@@ -1,11 +0,0 @@
-{
-  "variant": [
-    { "match": { "isF16": true }, "value": "gelu_f16" },
-    { "match": {}, "value": "gelu" }
-  ],
-  "overrides": [
-    { "match": { "useRowsplit": true }, "value": { "USE_ROWSPLIT": true } },
-    { "match": { "hasGate": true }, "value": { "HAS_GATE": true } },
-    { "match": {}, "value": {} }
-  ]
-}

package/src/rules/kernels/grouped-pointwise-conv2d.rules.json

@@ -1,6 +0,0 @@
-{
-  "variant": [
-    { "match": { "isF16": true }, "value": "default_f16" },
-    { "match": {}, "value": "default" }
-  ]
-}

package/src/rules/kernels/groupnorm.rules.json

@@ -1,10 +0,0 @@
-{
-  "statsVariant": [
-    { "match": { "isF16": true }, "value": "stats_f16" },
-    { "match": {}, "value": "stats" }
-  ],
-  "applyVariant": [
-    { "match": { "isF16": true }, "value": "apply_f16" },
-    { "match": {}, "value": "apply" }
-  ]
-}

package/src/rules/kernels/kernel-validator.d.ts

@@ -1,24 +0,0 @@
-export interface KernelValidatorFs {
-  readText(path: string): Promise<string>;
-  exists(path: string): Promise<boolean>;
-  list(path: string): Promise<string[]>;
-}
-
-export interface KernelRegistryValidationResult {
-  ok: boolean;
-  errors: string[];
-  filesCount: number;
-}
-
-export interface KernelOverrideLintResult {
-  ok: boolean;
-  errors: string[];
-  checkedCount: number;
-}
-
-export declare class KernelValidator {
-  constructor(fs: KernelValidatorFs);
-
-  validateRegistry(registryPath: string, kernelDir: string): Promise<KernelRegistryValidationResult>;
-  lintWgslOverrides(kernelDir: string, whitelist?: string[]): Promise<KernelOverrideLintResult>;
-}