Xenova's picture
Xenova HF Staff
sync 2e7068faf55e
cc4bd23 verified
Raw
History Blame
4.09 kB
{
"op": "com.microsoft.GemmaRotaryEmbedding",
"tunableSpace": { "workgroupSize": [64, 128, 256] },
"cases": [
{
"name": "gemma-rotary-f16-b1-h16-s2048-d256",
"preset": "smoke",
"vars": { "dtype": "float16" },
"inputs": {
"embT": { "shape": [1, 2048, 256], "dtype": "float32", "dist": "normal", "seed": 8100, "scale": 2 },
"qT": { "shape": [1, 16, 2048, 256], "dtype": "float16", "dist": "normal", "seed": 8101, "scale": 1 },
"qRotT": { "shape": [1, 16, 2048, 256], "dtype": "float16", "dist": "normal", "seed": 8102, "scale": 1 },
"kT": { "shape": [1, 16, 2048, 256], "dtype": "float16", "dist": "normal", "seed": 8103, "scale": 1 },
"kRotT": { "shape": [1, 16, 2048, 256], "dtype": "float16", "dist": "normal", "seed": 8104, "scale": 1 }
},
"outputs": {
"output1T": { "shape": [1, 16, 2048, 256], "dtype": "float16" },
"output2T": { "shape": [1, 16, 2048, 256], "dtype": "float16" }
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "6 * 8388608 * dtypeBytes(args.dtype) + 524288 * 4" }] }
},
{
"name": "gemma-rotary-f16-b1-h16-s1-d256",
"preset": "smoke",
"vars": { "dtype": "float16" },
"inputs": {
"embT": { "shape": [1, 1, 256], "dtype": "float32", "dist": "normal", "seed": 8110, "scale": 2 },
"qT": { "shape": [1, 16, 1, 256], "dtype": "float16", "dist": "normal", "seed": 8111, "scale": 1 },
"qRotT": { "shape": [1, 16, 1, 256], "dtype": "float16", "dist": "normal", "seed": 8112, "scale": 1 },
"kT": { "shape": [1, 16, 1, 256], "dtype": "float16", "dist": "normal", "seed": 8113, "scale": 1 },
"kRotT": { "shape": [1, 16, 1, 256], "dtype": "float16", "dist": "normal", "seed": 8114, "scale": 1 }
},
"outputs": {
"output1T": { "shape": [1, 16, 1, 256], "dtype": "float16" },
"output2T": { "shape": [1, 16, 1, 256], "dtype": "float16" }
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "6 * 4096 * dtypeBytes(args.dtype) + 256 * 4" }] }
},
{
"name": "gemma-rotary-f16-b4-h8-s512-d256",
"preset": "model",
"vars": { "dtype": "float16" },
"inputs": {
"embT": { "shape": [4, 512, 256], "dtype": "float32", "dist": "normal", "seed": 8130, "scale": 2 },
"qT": { "shape": [4, 8, 512, 256], "dtype": "float16", "dist": "normal", "seed": 8131, "scale": 1 },
"qRotT": { "shape": [4, 8, 512, 256], "dtype": "float16", "dist": "normal", "seed": 8132, "scale": 1 },
"kT": { "shape": [4, 8, 512, 256], "dtype": "float16", "dist": "normal", "seed": 8133, "scale": 1 },
"kRotT": { "shape": [4, 8, 512, 256], "dtype": "float16", "dist": "normal", "seed": 8134, "scale": 1 }
},
"outputs": {
"output1T": { "shape": [4, 8, 512, 256], "dtype": "float16" },
"output2T": { "shape": [4, 8, 512, 256], "dtype": "float16" }
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "6 * 4194304 * dtypeBytes(args.dtype) + 524288 * 4" }] }
},
{
"name": "gemma-rotary-f16-b2-h16-s512-d250-scalar",
"preset": "model",
"vars": { "dtype": "float16" },
"inputs": {
"embT": { "shape": [2, 512, 250], "dtype": "float32", "dist": "normal", "seed": 8140, "scale": 2 },
"qT": { "shape": [2, 16, 512, 250], "dtype": "float16", "dist": "normal", "seed": 8141, "scale": 1 },
"qRotT": { "shape": [2, 16, 512, 250], "dtype": "float16", "dist": "normal", "seed": 8142, "scale": 1 },
"kT": { "shape": [2, 16, 512, 250], "dtype": "float16", "dist": "normal", "seed": 8143, "scale": 1 },
"kRotT": { "shape": [2, 16, 512, 250], "dtype": "float16", "dist": "normal", "seed": 8144, "scale": 1 }
},
"outputs": {
"output1T": { "shape": [2, 16, 512, 250], "dtype": "float16" },
"output2T": { "shape": [2, 16, 512, 250], "dtype": "float16" }
},
"bench": { "metrics": [{ "type": "bandwidth", "value": "6 * 4096000 * dtypeBytes(args.dtype) + 256000 * 4" }] }
}
]
}