File size: 18,456 Bytes
38770a7 8c41d43 38770a7 8c41d43 38770a7 8c41d43 38770a7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 | {
"domain": "com.microsoft",
"name": "MatMulBnb4",
"sinceVersion": 1,
"description": "Computes `A @ dequant(B)^T` where `B` uses bitsandbytes 4-bit quantization: `quant_type = 0` selects FP4 and `quant_type = 1` selects NF4. Supports rank-2 float16/float32 `A`, `transB = 1`, and `training_mode = 0`; rank-1 and rank-3-or-higher `A`, bfloat16, `transB = 0`, and training are not implemented. `B` is the flattened `[N, K]` weight, two codes per byte with the even flat index in the high nibble. Each code indexes a fixed 16-entry codebook, and the value is `codebook[code] * absmax[flat_index / block_size]`.",
"inputs": [
{ "role": "A", "dtype": "T1", "rank": 2, "description": "Float input matrix of shape `(M, K)`, not quantized." },
{
"role": "B",
"dtype": "T2",
"rank": 1,
"description": "The `[N, K]` weight, flattened and quantized to 4 bits, stored as `(N * K + 1) / 2` bytes; the ONNX type is uint8 (this WebGPU implementation reads one widened u32 per stored byte)."
},
{
"role": "absmax",
"dtype": "T1",
"rank": 1,
"description": "Per-block absolute-maximum dequantization scales of shape `((N * K + block_size - 1) / block_size)`, same dtype as A."
}
],
"outputs": [
{
"role": "Y",
"dtype": "T1",
"rank": 2,
"shape": "[dim(shapes.A, 0), attrs.N]",
"description": "Result of `A` multiplied by the dequantized, transposed weight matrix, with shape `(M, N)` and the same dtype as `A`."
}
],
"attributes": { "training_mode": 0, "transB": 1 },
"attributeConstraints": {
"K": { "required": true },
"N": { "required": true },
"block_size": { "required": true },
"quant_type": { "required": true, "values": [0, 1] },
"training_mode": { "values": [0] },
"transB": { "values": [1] }
},
"attributeDescriptions": {
"K": "Input feature count (the shared dimension).",
"N": "Output feature count.",
"block_size": "Number of weights sharing one absmax scale; a power of two, at least 16.",
"quant_type": "Codebook selector: 0 = FP4, 1 = NF4.",
"training_mode": "Whether training outputs are requested. This inference-only implementation supports the standard default value 0.",
"transB": "Whether the quantized weight is stored transposed. This implementation supports the standard default value 1."
},
"typeConstraints": { "T1": ["float32", "float16"], "T2": ["uint8"] },
"args": {
"aT": { "kind": "tensor", "semantic": "A", "role": "input" },
"bT": { "kind": "tensor", "semantic": "B", "role": "input" },
"absmaxT": { "kind": "tensor", "semantic": "absmax", "role": "input" },
"yT": { "kind": "tensor", "semantic": "Y", "role": "output" }
},
"tunables": {
"WORKGROUP_SIZE": 64,
"TILE_MIN_M": 16,
"PORTABLE_TILE_K": 16,
"SGMAT_TILE_ROWS": 64,
"SGMAT_TALL_TILE_ROWS": 128,
"SGMAT_TALL_MIN_M": 128,
"SGMAT_TILE_COLS": 64,
"SGMAT_TILE_K": 32
},
"bindingSets": {
"main": [
{
"name": "a",
"arg": "aT",
"semantic": "A",
"buffer": { "type": "read-only-storage" },
"elementType": "$aScalar"
},
{ "name": "b", "arg": "bT", "semantic": "B", "buffer": { "type": "read-only-storage" }, "elementType": "u32" },
{
"name": "absmax",
"arg": "absmaxT",
"semantic": "absmax",
"buffer": { "type": "read-only-storage" },
"elementType": "$absmaxScalar",
"length": "$ABSMAX_LEN"
},
{ "name": "y", "arg": "yT", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "$aScalar" },
{
"name": "params",
"semantic": "kernel.params",
"buffer": { "type": "uniform" },
"struct": {
"name": "Params",
"fields": [
{ "name": "rows", "type": "u32", "value": "dim(shapes.A, 0)" },
{ "name": "K", "type": "u32", "value": "attrs.K" },
{ "name": "N", "type": "u32", "value": "attrs.N" },
{ "name": "blockSize", "type": "u32", "value": "attrs.block_size" }
]
}
}
],
"gemv": [
{
"name": "a",
"arg": "aT",
"semantic": "A",
"buffer": { "type": "read-only-storage" },
"elementType": "$aScalar"
},
{ "name": "b", "arg": "bT", "semantic": "B", "buffer": { "type": "read-only-storage" }, "elementType": "u32" },
{
"name": "absmax",
"arg": "absmaxT",
"semantic": "absmax",
"buffer": { "type": "read-only-storage" },
"elementType": "$absmaxScalar",
"length": "$ABSMAX_LEN"
},
{ "name": "y", "arg": "yT", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "$aScalar" },
{
"name": "params",
"semantic": "kernel.params",
"buffer": { "type": "uniform" },
"struct": {
"name": "Params",
"fields": [
{ "name": "K", "type": "u32", "value": "attrs.K" },
{ "name": "N", "type": "u32", "value": "attrs.N" },
{ "name": "blockSize", "type": "u32", "value": "attrs.block_size" }
]
}
}
],
"sgmat": [
{
"name": "a",
"arg": "aT",
"semantic": "A",
"buffer": { "type": "read-only-storage" },
"elementType": "$aScalar"
},
{ "name": "b", "arg": "bT", "semantic": "B", "buffer": { "type": "read-only-storage" }, "elementType": "u32" },
{
"name": "absmax",
"arg": "absmaxT",
"semantic": "absmax",
"buffer": { "type": "read-only-storage" },
"elementType": "$absmaxScalar",
"length": "$ABSMAX_LEN"
},
{ "name": "y", "arg": "yT", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "$aScalar" }
],
"castA": [
{
"name": "x",
"arg": "aT",
"semantic": "A",
"buffer": { "type": "read-only-storage" },
"elementType": "$srcScalar"
},
{ "name": "y", "semantic": "aF32", "buffer": { "type": "storage" }, "elementType": "f32" },
{
"name": "params",
"semantic": "kernel.params",
"buffer": { "type": "uniform" },
"struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.A)" }] }
}
],
"sgmatWiden": [
{ "name": "a", "semantic": "aF32", "buffer": { "type": "read-only-storage" }, "elementType": "f32" },
{ "name": "b", "arg": "bT", "semantic": "B", "buffer": { "type": "read-only-storage" }, "elementType": "u32" },
{
"name": "absmax",
"arg": "absmaxT",
"semantic": "absmax",
"buffer": { "type": "read-only-storage" },
"elementType": "$absmaxScalar",
"length": "$ABSMAX_LEN"
},
{ "name": "y", "semantic": "yF32", "buffer": { "type": "storage" }, "elementType": "f32" }
],
"castY": [
{ "name": "x", "semantic": "yF32", "buffer": { "type": "read-only-storage" }, "elementType": "f32" },
{ "name": "y", "arg": "yT", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "$srcScalar" },
{
"name": "params",
"semantic": "kernel.params",
"buffer": { "type": "uniform" },
"struct": {
"name": "Params",
"fields": [{ "name": "count", "type": "u32", "value": "dim(shapes.A, 0) * attrs.N" }]
}
}
]
},
"derive": {
"deviceWorkgroupCap": "min(device.limits.maxComputeInvocationsPerWorkgroup, device.limits.maxComputeWorkgroupSizeX)",
"wave32Adapter": "has(device.adapterInfo, \"subgroupMinSize\") and has(device.adapterInfo, \"subgroupMaxSize\") and device.adapterInfo.subgroupMinSize == 32 and device.adapterInfo.subgroupMaxSize == 32",
"canPinSubgroupSize32": "device.features.has(\"subgroups\") and device.features.has(\"subgroup-size-control\") and has(device.adapterInfo, \"subgroupMinSize\") and has(device.adapterInfo, \"subgroupMaxSize\") and device.adapterInfo.subgroupMinSize <= 32 and device.adapterInfo.subgroupMaxSize >= 32",
"pinSubgroupSize32": "canPinSubgroupSize32 and not wave32Adapter",
"wave32Effective": "wave32Adapter or pinSubgroupSize32",
"packedBytesExpected": "ceilDiv(attrs.N * attrs.K, 2)",
"absmaxCountExpected": "ceilDiv(attrs.N * attrs.K, attrs.block_size)",
"aFloatOk": "(tensorDtypes.A == \"float32\" or tensorDtypes.A == \"float16\") and f16Ok(tensorDtypes.A)",
"portableWorkgroupSize": "min(tunables.WORKGROUP_SIZE, deviceWorkgroupCap)",
"commonShapeValid": "ranks.A == 2 and ranks.B == 1 and ranks.absmax == 1 and ranks.Y == 2 and aFloatOk and tensorDtypes.B == \"uint8\" and tensorDtypes.absmax == tensorDtypes.A and tensorDtypes.Y == tensorDtypes.A and attrs.K > 0 and attrs.N > 0 and attrs.block_size >= 16 and pow2ceil(attrs.block_size) == attrs.block_size and dim(shapes.A, 1) == attrs.K and dim(shapes.B, 0) == packedBytesExpected and dim(shapes.absmax, 0) == absmaxCountExpected and dim(shapes.Y, 0) == dim(shapes.A, 0) and dim(shapes.Y, 1) == attrs.N",
"gemvShapeValid": "commonShapeValid and dim(shapes.A, 0) == 1",
"portableWorkgroupFits": "portableWorkgroupSize > 0 and portableWorkgroupSize * 16 <= device.limits.maxComputeWorkgroupStorageSize",
"portableTileKValid": "tunables.PORTABLE_TILE_K >= 8 and tunables.PORTABLE_TILE_K % 8 == 0",
"tileEligible": "commonShapeValid and dim(shapes.A, 0) >= tunables.TILE_MIN_M and portableTileKValid",
"tileWorkgroupStorageBytes": "8 * 64 * tunables.PORTABLE_TILE_K",
"tileWorkgroupFits": "16 <= device.limits.maxComputeWorkgroupSizeX and 8 <= device.limits.maxComputeWorkgroupSizeY and 128 <= device.limits.maxComputeInvocationsPerWorkgroup and tileWorkgroupStorageBytes <= device.limits.maxComputeWorkgroupStorageSize and ceilDiv(attrs.N, 64) <= device.limits.maxComputeWorkgroupsPerDimension and ceilDiv(dim(shapes.A, 0), 64) <= device.limits.maxComputeWorkgroupsPerDimension",
"sgmatMatrixSize": "8",
"sgmatTileRows": "tunables.SGMAT_TALL_TILE_ROWS if dim(shapes.A, 0) >= tunables.SGMAT_TALL_MIN_M else tunables.SGMAT_TILE_ROWS",
"sgmatRowSubtiles": "4",
"sgmatSubRows": "sgmatTileRows / sgmatRowSubtiles",
"sgmatSubCols": "4 * sgmatMatrixSize",
"sgmatLoadWidth": "sgmatMatrixSize",
"sgmatColSubtiles": "tunables.SGMAT_TILE_COLS / sgmatSubCols",
"sgmatNumSubgroups": "sgmatRowSubtiles * sgmatColSubtiles",
"sgmatSubgroupSize": "device.adapterInfo.subgroupMinSize if has(device.adapterInfo, \"subgroupMinSize\") else 1",
"sgmatWorkgroupSize": "sgmatNumSubgroups * sgmatSubgroupSize",
"sgmatBLoadsPerRow": "tunables.SGMAT_TILE_K / sgmatLoadWidth",
"sgmatWorkgroupStorageBytes": "4 * tunables.SGMAT_TILE_COLS * tunables.SGMAT_TILE_K",
"sgmatDispatchN": "ceilDiv(attrs.N, tunables.SGMAT_TILE_COLS)",
"sgmatDispatchM": "ceilDiv(dim(shapes.A, 0), sgmatTileRows)",
"sgmatWorkgroupFits": "sgmatWorkgroupSize <= deviceWorkgroupCap and sgmatWorkgroupStorageBytes <= device.limits.maxComputeWorkgroupStorageSize and sgmatDispatchN <= device.limits.maxComputeWorkgroupsPerDimension and sgmatDispatchM <= device.limits.maxComputeWorkgroupsPerDimension",
"sgmatWidenBytes": "numel(shapes.A) * 4",
"sgmatWidenOutBytes": "dim(shapes.A, 0) * attrs.N * 4",
"sgmatWidenFits": "sgmatWidenBytes <= device.limits.maxStorageBufferBindingSize and sgmatWidenBytes <= device.limits.maxBufferSize and sgmatWidenOutBytes <= device.limits.maxStorageBufferBindingSize and sgmatWidenOutBytes <= device.limits.maxBufferSize"
},
"constants": { "quantType": "attrs.quant_type", "ABSMAX_LEN": "ceilDiv(attrs.N * attrs.K, attrs.block_size)" },
"variants": [
{
"id": "gemv",
"priority": 20,
"when": ["gemvShapeValid", "portableWorkgroupFits"],
"constants": {
"workgroupSize": "portableWorkgroupSize",
"aScalar": "\"f16\" if tensorDtypes.A == \"float16\" else \"f32\"",
"usesF16": "tensorDtypes.A == \"float16\"",
"absmaxScalar": "\"f16\" if tensorDtypes.absmax == \"float16\" else \"f32\""
},
"passes": [
{
"id": "main",
"shader": "matmul-bnb4-gemv.wgsl.jinja",
"bindings": "gemv",
"dispatch": { "workgroups": "attrs.N" }
}
]
},
{
"id": "sgmat",
"priority": 15,
"requires": {
"features": ["subgroups", "chromium-experimental-subgroup-matrix"],
"limits": { "maxComputeWorkgroupStorageSize": 8192 },
"subgroupMatrixConfigs": [{ "componentType": "f32", "resultComponentType": "f32", "M": 8, "N": 8, "K": 8 }]
},
"when": ["commonShapeValid", "tensorDtypes.A == \"float32\"", "dim(shapes.A, 0) >= sgmatTileRows", "dim(shapes.A, 0) % sgmatTileRows == 0", "attrs.N % tunables.SGMAT_TILE_COLS == 0", "attrs.K % tunables.SGMAT_TILE_K == 0", "attrs.K % attrs.block_size == 0", "attrs.block_size % sgmatLoadWidth == 0", "wave32Effective", "sgmatWorkgroupFits"],
"constants": {
"K": "attrs.K",
"N": "attrs.N",
"blockSize": "attrs.block_size",
"tileRows": "sgmatTileRows",
"tileCols": "tunables.SGMAT_TILE_COLS",
"tileK": "tunables.SGMAT_TILE_K",
"subRows": "sgmatSubRows",
"subCols": "sgmatSubCols",
"matrixSize": "sgmatMatrixSize",
"rowMatrices": "sgmatSubRows / sgmatMatrixSize",
"colMatrices": "sgmatSubCols / sgmatMatrixSize",
"loadWidth": "sgmatLoadWidth",
"rowSubtiles": "sgmatRowSubtiles",
"workgroupSize": "sgmatWorkgroupSize",
"bLoadsPerRow": "sgmatBLoadsPerRow",
"aScalar": "\"f32\"",
"usesF16": false,
"absmaxScalar": "\"f16\" if tensorDtypes.absmax == \"float16\" else \"f32\""
},
"passes": [
{
"id": "main",
"shader": "matmul-bnb4-sgmat.wgsl.jinja",
"bindings": "sgmat",
"dispatch": { "x": "sgmatDispatchN", "y": "sgmatDispatchM" }
}
]
},
{
"id": "sgmat_widened",
"priority": 16,
"requires": {
"features": ["subgroups", "chromium-experimental-subgroup-matrix"],
"limits": { "maxComputeWorkgroupStorageSize": 8192 },
"subgroupMatrixConfigs": [{ "componentType": "f32", "resultComponentType": "f32", "M": 8, "N": 8, "K": 8 }]
},
"when": ["commonShapeValid", "tensorDtypes.A == \"float16\"", "dim(shapes.A, 0) >= sgmatTileRows", "dim(shapes.A, 0) % sgmatTileRows == 0", "attrs.N % tunables.SGMAT_TILE_COLS == 0", "attrs.K % tunables.SGMAT_TILE_K == 0", "attrs.K % attrs.block_size == 0", "attrs.block_size % sgmatLoadWidth == 0", "wave32Effective", "sgmatWorkgroupFits", "sgmatWidenFits"],
"constants": {
"K": "attrs.K",
"N": "attrs.N",
"blockSize": "attrs.block_size",
"tileRows": "sgmatTileRows",
"tileCols": "tunables.SGMAT_TILE_COLS",
"tileK": "tunables.SGMAT_TILE_K",
"subRows": "sgmatSubRows",
"subCols": "sgmatSubCols",
"matrixSize": "sgmatMatrixSize",
"rowMatrices": "sgmatSubRows / sgmatMatrixSize",
"colMatrices": "sgmatSubCols / sgmatMatrixSize",
"loadWidth": "sgmatLoadWidth",
"rowSubtiles": "sgmatRowSubtiles",
"workgroupSize": "sgmatWorkgroupSize",
"bLoadsPerRow": "sgmatBLoadsPerRow",
"aScalar": "\"f32\"",
"usesF16": true,
"absmaxScalar": "\"f16\" if tensorDtypes.absmax == \"float16\" else \"f32\"",
"srcScalar": "\"f16\"",
"outScalar": "\"f32\"",
"wrapNarrowInt": false,
"wrapSigned": false
},
"passes": [
{
"id": "widen_a",
"name": "MatMulBnb4.WidenActivations",
"shader": "cast-scalar-x4.wgsl.jinja",
"bindings": "castA",
"dispatch": { "threads": "ceilDiv(numel(shapes.A), 4)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
},
{
"id": "main",
"name": "MatMulBnb4.SubgroupMatrixWidened",
"shader": "matmul-bnb4-sgmat.wgsl.jinja",
"bindings": "sgmatWiden",
"dispatch": { "x": "sgmatDispatchN", "y": "sgmatDispatchM" }
},
{
"id": "narrow_y",
"name": "MatMulBnb4.NarrowOutput",
"shader": "cast-scalar-x4.wgsl.jinja",
"bindings": "castY",
"constants": { "outScalar": "\"f16\"" },
"dispatch": {
"threads": "ceilDiv(dim(shapes.A, 0) * attrs.N, 4)",
"workgroupSize": "tunables.WORKGROUP_SIZE"
}
}
],
"description": "Runs the subgroup-matrix tier for a float16 request by widening the activations to float32 either side of the multiply. The matrix units this operator uses accumulate in float32 from float32 operands; the device's float16 configuration returns a float16 result, so feeding them float16 directly would drop the accumulator's precision.",
"intermediates": [
{ "id": "aF32", "dtype": "float32", "shape": "[numel(shapes.A)]" },
{ "id": "yF32", "dtype": "float32", "shape": "[dim(shapes.A, 0) * attrs.N]" }
]
},
{
"id": "tiled",
"priority": 10,
"when": ["tileEligible", "tileWorkgroupFits"],
"constants": {
"tileK": "tunables.PORTABLE_TILE_K",
"aScalar": "\"f16\" if tensorDtypes.A == \"float16\" else \"f32\"",
"usesF16": "tensorDtypes.A == \"float16\"",
"absmaxScalar": "\"f16\" if tensorDtypes.absmax == \"float16\" else \"f32\""
},
"passes": [
{
"id": "main",
"shader": "matmul-bnb4-tiled.wgsl.jinja",
"bindings": "main",
"dispatch": { "x": "ceilDiv(attrs.N, 64)", "y": "ceilDiv(dim(shapes.A, 0), 64)" }
}
]
},
{
"id": "scalar",
"priority": 0,
"when": ["commonShapeValid", "portableWorkgroupSize > 0"],
"constants": {
"workgroupSize": "portableWorkgroupSize",
"aScalar": "\"f16\" if tensorDtypes.A == \"float16\" else \"f32\"",
"usesF16": "tensorDtypes.A == \"float16\"",
"absmaxScalar": "\"f16\" if tensorDtypes.absmax == \"float16\" else \"f32\""
},
"passes": [
{
"id": "main",
"shader": "matmul-bnb4.wgsl.jinja",
"bindings": "main",
"dispatch": { "threads": "numel(shapes.Y)", "workgroupSize": "constants.workgroupSize" }
}
]
}
]
}
|