File size: 13,093 Bytes
b88d0d9 | 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 | {
"op": "com.microsoft.MatMulNBitsQkv",
"tunableSpace": { "TILE_N": [4, 8, 16], "LANES": [4, 8, 16] },
"cases": [
{
"name": "qkv-q4-decode-k2048-nq2048-nkv512",
"preset": "smoke",
"vars": { "dtype": "float32" },
"attrs": { "K": 2048, "Nq": 2048, "Nkv": 512, "block_size": 32 },
"inputs": {
"aT": { "shape": [1, 2048], "dtype": "float32", "dist": "normal", "seed": 9101, "scale": 1 },
"normScaleT": { "shape": [2048], "dtype": "float32", "dist": "normal", "seed": 9102, "scale": 1 },
"qBT": { "shape": [2048, 64, 16], "dtype": "uint8", "dist": "uniform", "seed": 9103, "scale": 255 },
"qScalesT": { "shape": [2048, 64], "dtype": "float32", "dist": "normal", "seed": 9104, "scale": 0.05 },
"kBT": { "shape": [512, 64, 16], "dtype": "uint8", "dist": "uniform", "seed": 9105, "scale": 255 },
"kScalesT": { "shape": [512, 64], "dtype": "float32", "dist": "normal", "seed": 9106, "scale": 0.05 },
"vBT": { "shape": [512, 64, 16], "dtype": "uint8", "dist": "uniform", "seed": 9107, "scale": 255 },
"vScalesT": { "shape": [512, 64], "dtype": "float32", "dist": "normal", "seed": 9108, "scale": 0.05 }
},
"outputs": {
"qT": { "shape": [1, 2048], "dtype": "float32" },
"kT": { "shape": [1, 512], "dtype": "float32" },
"vT": { "shape": [1, 512], "dtype": "float32" }
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "(2048 + 512 + 512) * 64 * 16 * 4" }] }
},
{
"name": "qkv-q4-llama-decode-k4096-nq4096-nkv4096",
"preset": "model",
"provenance": {
"notes": "Llama class defaults (hidden_size 4096, num_attention_heads 32, no GQA so num_key_value_heads matches) at a decode step. Bytes are GPU storage, not the packed blob: this backend gives a uint8 tensor one byte per u32 element, so a 4-bit code costs 2 bytes of traffic and the projection moves four times its on-disk size. Counting the blob understated the traffic 4x and reported this kernel at a tenth of the bandwidth rail when it runs at about half."
},
"vars": { "dtype": "float32", "tokens": 1, "K": 4096, "Nq": 4096, "Nkv": 4096 },
"attrs": { "K": 4096, "Nq": 4096, "Nkv": 4096, "block_size": 32 },
"inputs": {
"aT": { "shape": [1, 4096], "dtype": "float32", "dist": "normal", "seed": 7100, "scale": 0.5 },
"normScaleT": {
"shape": [4096],
"dtype": "float32",
"dist": "uniform",
"seed": 7101,
"scale": 0.1,
"offset": 1
},
"qBT": { "shape": [4096, 128, 16], "dtype": "uint8", "dist": "uniform", "seed": 7102, "min": 0, "max": 255 },
"qScalesT": {
"shape": [4096, 128],
"dtype": "float32",
"dist": "uniform",
"seed": 7103,
"min": 0.005,
"max": 0.05
},
"kBT": { "shape": [4096, 128, 16], "dtype": "uint8", "dist": "uniform", "seed": 7104, "min": 0, "max": 255 },
"kScalesT": {
"shape": [4096, 128],
"dtype": "float32",
"dist": "uniform",
"seed": 7105,
"min": 0.005,
"max": 0.05
},
"vBT": { "shape": [4096, 128, 16], "dtype": "uint8", "dist": "uniform", "seed": 7106, "min": 0, "max": 255 },
"vScalesT": {
"shape": [4096, 128],
"dtype": "float32",
"dist": "uniform",
"seed": 7107,
"min": 0.005,
"max": 0.05
}
},
"outputs": {
"qT": { "shape": [1, 4096], "dtype": "float32" },
"kT": { "shape": [1, 4096], "dtype": "float32" },
"vT": { "shape": [1, 4096], "dtype": "float32" }
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "(args.Nq + 2 * args.Nkv) * args.K * 2" }] }
},
{
"name": "qkv-q4-mistral-decode-k4096-nq4096-nkv1024",
"preset": "model",
"provenance": {
"notes": "Mistral class defaults (hidden_size 4096, 32 query heads, 8 key/value heads, head_dim 128) -- the GQA case, where K and V projections are a quarter of Q. Bytes are GPU storage, not the packed blob: this backend gives a uint8 tensor one byte per u32 element, so a 4-bit code costs 2 bytes of traffic and the projection moves four times its on-disk size. Counting the blob understated the traffic 4x and reported this kernel at a tenth of the bandwidth rail when it runs at about half."
},
"vars": { "dtype": "float32", "tokens": 1, "K": 4096, "Nq": 4096, "Nkv": 1024 },
"attrs": { "K": 4096, "Nq": 4096, "Nkv": 1024, "block_size": 32 },
"inputs": {
"aT": { "shape": [1, 4096], "dtype": "float32", "dist": "normal", "seed": 7200, "scale": 0.5 },
"normScaleT": {
"shape": [4096],
"dtype": "float32",
"dist": "uniform",
"seed": 7201,
"scale": 0.1,
"offset": 1
},
"qBT": { "shape": [4096, 128, 16], "dtype": "uint8", "dist": "uniform", "seed": 7202, "min": 0, "max": 255 },
"qScalesT": {
"shape": [4096, 128],
"dtype": "float32",
"dist": "uniform",
"seed": 7203,
"min": 0.005,
"max": 0.05
},
"kBT": { "shape": [1024, 128, 16], "dtype": "uint8", "dist": "uniform", "seed": 7204, "min": 0, "max": 255 },
"kScalesT": {
"shape": [1024, 128],
"dtype": "float32",
"dist": "uniform",
"seed": 7205,
"min": 0.005,
"max": 0.05
},
"vBT": { "shape": [1024, 128, 16], "dtype": "uint8", "dist": "uniform", "seed": 7206, "min": 0, "max": 255 },
"vScalesT": {
"shape": [1024, 128],
"dtype": "float32",
"dist": "uniform",
"seed": 7207,
"min": 0.005,
"max": 0.05
}
},
"outputs": {
"qT": { "shape": [1, 4096], "dtype": "float32" },
"kT": { "shape": [1, 1024], "dtype": "float32" },
"vT": { "shape": [1, 1024], "dtype": "float32" }
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "(args.Nq + 2 * args.Nkv) * args.K * 2" }] }
},
{
"name": "qkv-q4-gemma3-decode-k2304-nq2048-nkv1024",
"preset": "model",
"provenance": {
"notes": "Gemma3 class defaults (hidden_size 2304, 8 query heads, 4 key/value heads, head_dim 256); the projection is wider than hidden, so K does not equal Nq. Bytes are GPU storage, not the packed blob: this backend gives a uint8 tensor one byte per u32 element, so a 4-bit code costs 2 bytes of traffic and the projection moves four times its on-disk size. Counting the blob understated the traffic 4x and reported this kernel at a tenth of the bandwidth rail when it runs at about half."
},
"vars": { "dtype": "float32", "tokens": 1, "K": 2304, "Nq": 2048, "Nkv": 1024 },
"attrs": { "K": 2304, "Nq": 2048, "Nkv": 1024, "block_size": 32 },
"inputs": {
"aT": { "shape": [1, 2304], "dtype": "float32", "dist": "normal", "seed": 7300, "scale": 0.5 },
"normScaleT": {
"shape": [2304],
"dtype": "float32",
"dist": "uniform",
"seed": 7301,
"scale": 0.1,
"offset": 1
},
"qBT": { "shape": [2048, 72, 16], "dtype": "uint8", "dist": "uniform", "seed": 7302, "min": 0, "max": 255 },
"qScalesT": {
"shape": [2048, 72],
"dtype": "float32",
"dist": "uniform",
"seed": 7303,
"min": 0.005,
"max": 0.05
},
"kBT": { "shape": [1024, 72, 16], "dtype": "uint8", "dist": "uniform", "seed": 7304, "min": 0, "max": 255 },
"kScalesT": {
"shape": [1024, 72],
"dtype": "float32",
"dist": "uniform",
"seed": 7305,
"min": 0.005,
"max": 0.05
},
"vBT": { "shape": [1024, 72, 16], "dtype": "uint8", "dist": "uniform", "seed": 7306, "min": 0, "max": 255 },
"vScalesT": {
"shape": [1024, 72],
"dtype": "float32",
"dist": "uniform",
"seed": 7307,
"min": 0.005,
"max": 0.05
}
},
"outputs": {
"qT": { "shape": [1, 2048], "dtype": "float32" },
"kT": { "shape": [1, 1024], "dtype": "float32" },
"vT": { "shape": [1, 1024], "dtype": "float32" }
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "(args.Nq + 2 * args.Nkv) * args.K * 2" }] }
},
{
"name": "qkv-q4-phi3-decode-k3072-nq3072-nkv3072",
"preset": "model",
"provenance": {
"notes": "Phi-3 class defaults (hidden_size 3072, num_attention_heads 32, no GQA). Bytes are GPU storage, not the packed blob: this backend gives a uint8 tensor one byte per u32 element, so a 4-bit code costs 2 bytes of traffic and the projection moves four times its on-disk size. Counting the blob understated the traffic 4x and reported this kernel at a tenth of the bandwidth rail when it runs at about half."
},
"vars": { "dtype": "float32", "tokens": 1, "K": 3072, "Nq": 3072, "Nkv": 3072 },
"attrs": { "K": 3072, "Nq": 3072, "Nkv": 3072, "block_size": 32 },
"inputs": {
"aT": { "shape": [1, 3072], "dtype": "float32", "dist": "normal", "seed": 7400, "scale": 0.5 },
"normScaleT": {
"shape": [3072],
"dtype": "float32",
"dist": "uniform",
"seed": 7401,
"scale": 0.1,
"offset": 1
},
"qBT": { "shape": [3072, 96, 16], "dtype": "uint8", "dist": "uniform", "seed": 7402, "min": 0, "max": 255 },
"qScalesT": {
"shape": [3072, 96],
"dtype": "float32",
"dist": "uniform",
"seed": 7403,
"min": 0.005,
"max": 0.05
},
"kBT": { "shape": [3072, 96, 16], "dtype": "uint8", "dist": "uniform", "seed": 7404, "min": 0, "max": 255 },
"kScalesT": {
"shape": [3072, 96],
"dtype": "float32",
"dist": "uniform",
"seed": 7405,
"min": 0.005,
"max": 0.05
},
"vBT": { "shape": [3072, 96, 16], "dtype": "uint8", "dist": "uniform", "seed": 7406, "min": 0, "max": 255 },
"vScalesT": {
"shape": [3072, 96],
"dtype": "float32",
"dist": "uniform",
"seed": 7407,
"min": 0.005,
"max": 0.05
}
},
"outputs": {
"qT": { "shape": [1, 3072], "dtype": "float32" },
"kT": { "shape": [1, 3072], "dtype": "float32" },
"vT": { "shape": [1, 3072], "dtype": "float32" }
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "(args.Nq + 2 * args.Nkv) * args.K * 2" }] }
},
{
"name": "qkv-q4-llama-spec8-k4096-nq4096-nkv4096",
"preset": "model",
"provenance": {
"notes": "Llama class defaults with 8 rows, the shape speculative decoding verifies in one pass. Bytes are GPU storage, not the packed blob: this backend gives a uint8 tensor one byte per u32 element, so a 4-bit code costs 2 bytes of traffic and the projection moves four times its on-disk size. Counting the blob understated the traffic 4x and reported this kernel at a tenth of the bandwidth rail when it runs at about half."
},
"vars": { "dtype": "float32", "tokens": 8, "K": 4096, "Nq": 4096, "Nkv": 4096 },
"attrs": { "K": 4096, "Nq": 4096, "Nkv": 4096, "block_size": 32 },
"inputs": {
"aT": { "shape": [8, 4096], "dtype": "float32", "dist": "normal", "seed": 7500, "scale": 0.5 },
"normScaleT": {
"shape": [4096],
"dtype": "float32",
"dist": "uniform",
"seed": 7501,
"scale": 0.1,
"offset": 1
},
"qBT": { "shape": [4096, 128, 16], "dtype": "uint8", "dist": "uniform", "seed": 7502, "min": 0, "max": 255 },
"qScalesT": {
"shape": [4096, 128],
"dtype": "float32",
"dist": "uniform",
"seed": 7503,
"min": 0.005,
"max": 0.05
},
"kBT": { "shape": [4096, 128, 16], "dtype": "uint8", "dist": "uniform", "seed": 7504, "min": 0, "max": 255 },
"kScalesT": {
"shape": [4096, 128],
"dtype": "float32",
"dist": "uniform",
"seed": 7505,
"min": 0.005,
"max": 0.05
},
"vBT": { "shape": [4096, 128, 16], "dtype": "uint8", "dist": "uniform", "seed": 7506, "min": 0, "max": 255 },
"vScalesT": {
"shape": [4096, 128],
"dtype": "float32",
"dist": "uniform",
"seed": 7507,
"min": 0.005,
"max": 0.05
}
},
"outputs": {
"qT": { "shape": [8, 4096], "dtype": "float32" },
"kT": { "shape": [8, 4096], "dtype": "float32" },
"vT": { "shape": [8, 4096], "dtype": "float32" }
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "(args.Nq + 2 * args.Nkv) * args.K * 2" }] }
}
]
}
|