ai.onnx.SwiGLU / build /webgpu /bench.json
Xenova's picture
Xenova HF Staff
sync 6fdf6301e2bb
e34da4c verified
Raw History Blame
2.28 kB
{
"tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
"cases": [
{
"name": "swiglu-f32-64x257",
"preset": "smoke",
"vars": { "dtype": "float32", "count": 16448 },
"inputs": {
"a": { "shape": [64, 257], "dtype": "float32", "dist": "normal", "seed": 9111, "scale": 2 },
"b": { "shape": [64, 257], "dtype": "float32", "dist": "normal", "seed": 9112, "scale": 2 }
},
"outputs": { "y": { "shape": [64, 257], "dtype": "float32" } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3" }] }
},
{
"name": "swiglu-f32-4m",
"preset": "smoke",
"vars": { "dtype": "float32", "count": 4194304 },
"inputs": {
"a": { "shape": [4096, 1024], "dtype": "float32", "dist": "normal", "seed": 9101, "scale": 2 },
"b": { "shape": [4096, 1024], "dtype": "float32", "dist": "normal", "seed": 9102, "scale": 2 }
},
"outputs": { "y": { "shape": [4096, 1024], "dtype": "float32" } },
"bench": {
"primary": true,
"metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3" }]
}
},
{
"name": "swiglu-f16-4m",
"preset": "smoke",
"vars": { "dtype": "float16", "count": 4194304 },
"inputs": {
"a": { "shape": [4096, 1024], "dtype": "float16", "dist": "normal", "seed": 9103, "scale": 2 },
"b": { "shape": [4096, 1024], "dtype": "float16", "dist": "normal", "seed": 9104, "scale": 2 }
},
"outputs": { "y": { "shape": [4096, 1024], "dtype": "float16" } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3" }] }
},
{
"name": "swiglu-f32-odd-scalar-x4",
"preset": "edge",
"vars": { "dtype": "float32", "count": 4190208 },
"inputs": {
"a": { "shape": [4096, 1023], "dtype": "float32", "dist": "normal", "seed": 9105, "scale": 2 },
"b": { "shape": [4096, 1023], "dtype": "float32", "dist": "normal", "seed": 9106, "scale": 2 }
},
"outputs": { "y": { "shape": [4096, 1023], "dtype": "float32" } },
"bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 3" }] }
}
]
}