ardy-web-onnx / text-encoder /encoder.json
sprited's picture
Text encoder (Llama-3 LLM2Vec, INT4 MatMulNBits) in three parts for onnxruntime-web
3a3fe24 verified
Raw History Blame Contribute Delete
1.57 kB
{
"seq": 64,
"llm_dim": 4096,
"cut_layers": [
8,
20
],
"parts": [
{
"name": "part_a",
"size_bytes": 2032271360,
"inputs": [
{
"name": "attention_mask",
"dtype": "int64",
"shape": [
1,
64
]
},
{
"name": "input_ids",
"dtype": "int64",
"shape": [
1,
64
]
}
],
"outputs": [
{
"name": "/model/layers.7/Add_1_output_0",
"dtype": "float16",
"shape": [
1,
64,
4096
]
}
]
},
{
"name": "part_b",
"size_bytes": 1472397312,
"inputs": [
{
"name": "/model/layers.7/Add_1_output_0",
"dtype": "float16",
"shape": [
1,
64,
4096
]
},
{
"name": "attention_mask",
"dtype": "int64",
"shape": [
1,
64
]
}
],
"outputs": [
{
"name": "/model/layers.19/Add_1_output_0",
"dtype": "float16",
"shape": [
1,
64,
4096
]
}
]
},
{
"name": "part_c",
"size_bytes": 1472405504,
"inputs": [
{
"name": "/model/layers.19/Add_1_output_0",
"dtype": "float16",
"shape": [
1,
64,
4096
]
},
{
"name": "attention_mask",
"dtype": "int64",
"shape": [
1,
64
]
},
{
"name": "embed_mask",
"dtype": "int64",
"shape": [
1,
64
]
}
],
"outputs": [
{
"name": "text_embedding",
"dtype": "float",
"shape": [
1,
4096
]
}
]
}
]
}