Xenova's picture
Xenova HF Staff
sync 2e7068faf55e
fe90e05 verified
Raw
History Blame
2.49 kB
{
"op": "com.microsoft.GemmFastGelu",
"cases": [
{
"name": "gemmfastgelu-bert-base-b8-s384-h768-i3072",
"preset": "smoke",
"vars": { "dtype": "float32" },
"inputs": {
"X": { "shape": [8, 384, 768], "dtype": "float32", "dist": "normal", "seed": 5200, "scale": 1 },
"W": { "shape": [768, 3072], "dtype": "float32", "dist": "normal", "seed": 5201, "scale": 1 },
"bias": { "shape": [3072], "dtype": "float32", "dist": "normal", "seed": 5202, "scale": 1 }
},
"outputs": { "Y": { "shape": [8, 384, 3072], "dtype": "float32" } },
"bench": { "metrics": [{ "type": "gflops", "value": "2 * 3072 * 768 * 3072" }] }
},
{
"name": "gemmfastgelu-f16-bert-base-b8-s384-h768-i3072",
"preset": "smoke",
"vars": { "dtype": "float16" },
"inputs": {
"X": { "shape": [8, 384, 768], "dtype": "float16", "dist": "normal", "seed": 5210, "scale": 1 },
"W": { "shape": [768, 3072], "dtype": "float16", "dist": "normal", "seed": 5211, "scale": 1 },
"bias": { "shape": [3072], "dtype": "float16", "dist": "normal", "seed": 5212, "scale": 1 }
},
"outputs": { "Y": { "shape": [8, 384, 3072], "dtype": "float16" } },
"bench": { "metrics": [{ "type": "gflops", "value": "2 * 3072 * 768 * 3072" }] }
},
{
"name": "gemmfastgelu-bert-large-b4-s512-h1024-i4096",
"preset": "model",
"vars": { "dtype": "float32" },
"inputs": {
"X": { "shape": [4, 512, 1024], "dtype": "float32", "dist": "normal", "seed": 5220, "scale": 1 },
"W": { "shape": [1024, 4096], "dtype": "float32", "dist": "normal", "seed": 5221, "scale": 1 },
"bias": { "shape": [4096], "dtype": "float32", "dist": "normal", "seed": 5222, "scale": 1 }
},
"outputs": { "Y": { "shape": [4, 512, 4096], "dtype": "float32" } },
"bench": { "metrics": [{ "type": "gflops", "value": "2 * 2048 * 1024 * 4096" }] }
},
{
"name": "gemmfastgelu-nobias-b8-s384-h768-i3072",
"preset": "model",
"vars": { "dtype": "float32" },
"inputs": {
"X": { "shape": [8, 384, 768], "dtype": "float32", "dist": "normal", "seed": 5230, "scale": 1 },
"W": { "shape": [768, 3072], "dtype": "float32", "dist": "normal", "seed": 5231, "scale": 1 }
},
"outputs": { "Y": { "shape": [8, 384, 3072], "dtype": "float32" } },
"bench": { "metrics": [{ "type": "gflops", "value": "2 * 3072 * 768 * 3072" }] }
}
]
}