2677 lines
45 KiB
JSON
2677 lines
45 KiB
JSON
{
|
|
"flux": [
|
|
{
|
|
"shape": [
|
|
4608,
|
|
3072,
|
|
15360
|
|
],
|
|
"count": 38,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
4608,
|
|
12288,
|
|
3072
|
|
],
|
|
"count": 38,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
4608,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 114,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 57,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
10240,
|
|
4096
|
|
],
|
|
"count": 48,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 19,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
12288,
|
|
4096
|
|
],
|
|
"count": 24,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.qkv_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
10240
|
|
],
|
|
"count": 24,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 57,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
4096
|
|
],
|
|
"count": 24,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.o_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 19,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_add_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
3072,
|
|
4096
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
77,
|
|
3072,
|
|
768
|
|
],
|
|
"count": 12,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "clip.layers.*.mlp.fc1"
|
|
},
|
|
{
|
|
"shape": [
|
|
77,
|
|
768,
|
|
3072
|
|
],
|
|
"count": 12,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "clip.layers.*.mlp.fc2"
|
|
},
|
|
{
|
|
"shape": [
|
|
77,
|
|
2304,
|
|
768
|
|
],
|
|
"count": 12,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": "clip.layers.*.self_attn.qkv_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
64,
|
|
3072
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
3072,
|
|
64
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
77,
|
|
768,
|
|
768
|
|
],
|
|
"count": 12,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "clip.layers.*.self_attn.out_proj"
|
|
}
|
|
],
|
|
"flux2": [
|
|
{
|
|
"shape": [
|
|
4608,
|
|
55296,
|
|
6144
|
|
],
|
|
"count": 48,
|
|
"kind": "runtime_fused_qkv_mlp",
|
|
"prefix": "single_transformer_blocks.*.attn.to_qkv_mlp_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
4608,
|
|
6144,
|
|
24576
|
|
],
|
|
"count": 48,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "single_transformer_blocks.*.attn.to_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
36864,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.ff.linear_in"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
6144,
|
|
18432
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.ff.linear_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
18432,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "synthetic_packed_qkv",
|
|
"prefix": "transformer_blocks.*.attn.packed_qkv"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
12288,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "synthetic_packed_kv",
|
|
"prefix": "transformer_blocks.*.attn.packed_kv"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
6144,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_k"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
6144,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
6144,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_q"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
6144,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_v"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
36864,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.ff_context.linear_in"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
6144,
|
|
18432
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.ff_context.linear_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
18432,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "synthetic_packed_added_qkv",
|
|
"prefix": "transformer_blocks.*.attn.packed_qkv"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
12288,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "synthetic_packed_kv",
|
|
"prefix": "transformer_blocks.*.attn.packed_kv"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
6144,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.add_k_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
6144,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.add_q_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
6144,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.add_v_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
6144,
|
|
6144
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_add_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
6144,
|
|
15360
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
6144,
|
|
128
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
128,
|
|
6144
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "proj_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
36864,
|
|
6144
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
18432,
|
|
6144
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
}
|
|
],
|
|
"helios": [
|
|
{
|
|
"shape": [
|
|
11040,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 240,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
11040,
|
|
13824,
|
|
5120
|
|
],
|
|
"count": 80,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
11040,
|
|
5120,
|
|
13824
|
|
],
|
|
"count": 80,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
11040,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 80,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
8640,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 80,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
8640,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 80,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
10240,
|
|
4096
|
|
],
|
|
"count": 96,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
12288,
|
|
4096
|
|
],
|
|
"count": 48,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.qkv_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
10240
|
|
],
|
|
"count": 48,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
4096
|
|
],
|
|
"count": 48,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.o_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
4096
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
8640,
|
|
64,
|
|
5120
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
30720,
|
|
5120
|
|
],
|
|
"count": 4,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 4,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
5120,
|
|
256
|
|
],
|
|
"count": 4,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
}
|
|
],
|
|
"hunyuanvideo": [
|
|
{
|
|
"shape": [
|
|
27085,
|
|
21504,
|
|
3072
|
|
],
|
|
"count": 40,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.single_blocks.*.linear1"
|
|
},
|
|
{
|
|
"shape": [
|
|
27085,
|
|
3072,
|
|
15360
|
|
],
|
|
"count": 40,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.single_blocks.*.linear2"
|
|
},
|
|
{
|
|
"shape": [
|
|
27030,
|
|
3072,
|
|
12288
|
|
],
|
|
"count": 20,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.double_blocks.*.img_mlp.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
27030,
|
|
12288,
|
|
3072
|
|
],
|
|
"count": 20,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.double_blocks.*.img_mlp.*.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
27030,
|
|
9216,
|
|
3072
|
|
],
|
|
"count": 20,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": "Hunyuan.double_blocks.*.img_attn_qkv"
|
|
},
|
|
{
|
|
"shape": [
|
|
27030,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 20,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.double_blocks.*.img_attn_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
150,
|
|
28672,
|
|
4096
|
|
],
|
|
"count": 32,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "llama.layers.*.mlp.gate_up_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
150,
|
|
4096,
|
|
14336
|
|
],
|
|
"count": 32,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "llama.layers.*.mlp.down_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
150,
|
|
6144,
|
|
4096
|
|
],
|
|
"count": 32,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": "llama.layers.*.self_attn.qkv_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
150,
|
|
4096,
|
|
4096
|
|
],
|
|
"count": 32,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "llama.layers.*.self_attn.o_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
55,
|
|
12288,
|
|
3072
|
|
],
|
|
"count": 20,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
55,
|
|
3072,
|
|
12288
|
|
],
|
|
"count": 20,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
55,
|
|
9216,
|
|
3072
|
|
],
|
|
"count": 20,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
55,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 20,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
27030,
|
|
64,
|
|
3072
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.final_layer.linear"
|
|
},
|
|
{
|
|
"shape": [
|
|
55,
|
|
3072,
|
|
12288
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.txt_in.refiner_blocks.*.mlp.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
55,
|
|
12288,
|
|
3072
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.txt_in.refiner_blocks.*.mlp.*.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
55,
|
|
9216,
|
|
3072
|
|
],
|
|
"count": 2,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": "Hunyuan.txt_in.refiner_blocks.*.self_attn_qkv"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
18432,
|
|
3072
|
|
],
|
|
"count": 40,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
46,
|
|
3072,
|
|
768
|
|
],
|
|
"count": 12,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "clip.layers.*.mlp.fc1"
|
|
},
|
|
{
|
|
"shape": [
|
|
46,
|
|
768,
|
|
3072
|
|
],
|
|
"count": 12,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "clip.layers.*.mlp.fc2"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
9216,
|
|
3072
|
|
],
|
|
"count": 40,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
55,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.txt_in.refiner_blocks.*.self_attn_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
46,
|
|
2304,
|
|
768
|
|
],
|
|
"count": 12,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": "clip.layers.*.self_attn.qkv_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
55,
|
|
3072,
|
|
4096
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.txt_in.input_embedder"
|
|
},
|
|
{
|
|
"shape": [
|
|
46,
|
|
768,
|
|
768
|
|
],
|
|
"count": 12,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "clip.layers.*.self_attn.out_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
6144,
|
|
3072
|
|
],
|
|
"count": 3,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 3,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
3072,
|
|
4096
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.txt_in.c_embedder.*.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.txt_in.c_embedder.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.vector_in.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
3072,
|
|
256
|
|
],
|
|
"count": 3,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
3072,
|
|
768
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "Hunyuan.vector_in.*.proj"
|
|
}
|
|
],
|
|
"ltx": [
|
|
{
|
|
"shape": [
|
|
6144,
|
|
4096,
|
|
4096
|
|
],
|
|
"count": 384,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
6144,
|
|
4096,
|
|
16384
|
|
],
|
|
"count": 96,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
6144,
|
|
16384,
|
|
4096
|
|
],
|
|
"count": 96,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
6144,
|
|
4096,
|
|
4096
|
|
],
|
|
"count": 192,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
6144,
|
|
2048,
|
|
4096
|
|
],
|
|
"count": 288,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
6144,
|
|
4096,
|
|
2048
|
|
],
|
|
"count": 96,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1024,
|
|
4096,
|
|
4096
|
|
],
|
|
"count": 194,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1024,
|
|
2048,
|
|
2048
|
|
],
|
|
"count": 194,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
126,
|
|
2048,
|
|
2048
|
|
],
|
|
"count": 672,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
126,
|
|
2048,
|
|
8192
|
|
],
|
|
"count": 96,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
126,
|
|
8192,
|
|
2048
|
|
],
|
|
"count": 96,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
126,
|
|
2048,
|
|
2048
|
|
],
|
|
"count": 288,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1024,
|
|
4096,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1024,
|
|
2048,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
6144,
|
|
128,
|
|
4096
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
6144,
|
|
4096,
|
|
128
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
24576,
|
|
4096
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
4096,
|
|
4096
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
16384,
|
|
4096
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
126,
|
|
128,
|
|
2048
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
126,
|
|
2048,
|
|
128
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
12288,
|
|
2048
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
2048,
|
|
2048
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
8192,
|
|
2048
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
4096,
|
|
256
|
|
],
|
|
"count": 6,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
2048,
|
|
256
|
|
],
|
|
"count": 6,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
}
|
|
],
|
|
"mova-720p": [
|
|
{
|
|
"shape": [
|
|
44100,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 3040,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
44100,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 1760,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
44100,
|
|
13824,
|
|
5120
|
|
],
|
|
"count": 640,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
44100,
|
|
5120,
|
|
13824
|
|
],
|
|
"count": 640,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
44100,
|
|
1536,
|
|
5120
|
|
],
|
|
"count": 960,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 1280,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
2560,
|
|
4096
|
|
],
|
|
"count": 384,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
3072,
|
|
4096
|
|
],
|
|
"count": 192,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.qkv_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
1536,
|
|
1536
|
|
],
|
|
"count": 960,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
2560
|
|
],
|
|
"count": 192,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
176400,
|
|
64,
|
|
5120
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
101,
|
|
5120,
|
|
1536
|
|
],
|
|
"count": 960,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
101,
|
|
8960,
|
|
1536
|
|
],
|
|
"count": 480,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
101,
|
|
1536,
|
|
8960
|
|
],
|
|
"count": 480,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
101,
|
|
1536,
|
|
1536
|
|
],
|
|
"count": 2400,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
1024
|
|
],
|
|
"count": 192,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.o_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
101,
|
|
1536,
|
|
1536
|
|
],
|
|
"count": 1440,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
4096
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
1536,
|
|
4096
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
1536,
|
|
1536
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
30720,
|
|
5120
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
403,
|
|
128,
|
|
1536
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
9216,
|
|
1536
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
1536,
|
|
1536
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
5120,
|
|
256
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
1536,
|
|
256
|
|
],
|
|
"count": 16,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
}
|
|
],
|
|
"qwen": [
|
|
{
|
|
"shape": [
|
|
4096,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 720,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 240,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
47,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 360,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
19,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 360,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
47,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 120,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_add_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
19,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 120,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_add_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
18432,
|
|
3072
|
|
],
|
|
"count": 240,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.img_mod"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
18432,
|
|
3072
|
|
],
|
|
"count": 240,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.txt_mod"
|
|
}
|
|
],
|
|
"qwen-edit": [
|
|
{
|
|
"shape": [
|
|
8308,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 720,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
8308,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 240,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
195,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 360,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
189,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 360,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
195,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 120,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_add_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
189,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 120,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.attn.to_add_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
2,
|
|
18432,
|
|
3072
|
|
],
|
|
"count": 240,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.img_mod"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
18432,
|
|
3072
|
|
],
|
|
"count": 240,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "transformer_blocks.*.txt_mod"
|
|
}
|
|
],
|
|
"wan-i2v": [
|
|
{
|
|
"shape": [
|
|
15792,
|
|
15360,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "synthetic_packed_qkv",
|
|
"prefix": "blocks.*.attn1.packed_qkv"
|
|
},
|
|
{
|
|
"shape": [
|
|
15792,
|
|
5120,
|
|
13824
|
|
],
|
|
"count": 320,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.ffn.net.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
15792,
|
|
13824,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.ffn.net.*.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
15792,
|
|
10240,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "synthetic_packed_kv",
|
|
"prefix": "blocks.*.attn1.packed_kv"
|
|
},
|
|
{
|
|
"shape": [
|
|
15792,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_k"
|
|
},
|
|
{
|
|
"shape": [
|
|
15792,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
15792,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_q"
|
|
},
|
|
{
|
|
"shape": [
|
|
15792,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_v"
|
|
},
|
|
{
|
|
"shape": [
|
|
15792,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
15792,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_q"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
10240,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "synthetic_packed_kv",
|
|
"prefix": "blocks.*.attn2.packed_kv"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_k"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 320,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_v"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
4096
|
|
],
|
|
"count": 384,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
6144,
|
|
4096
|
|
],
|
|
"count": 192,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.qkv_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
5120
|
|
],
|
|
"count": 192,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
2048
|
|
],
|
|
"count": 192,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.o_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
4096
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
31584,
|
|
64,
|
|
5120
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "proj_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
30720,
|
|
5120
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
5120,
|
|
256
|
|
],
|
|
"count": 8,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
}
|
|
],
|
|
"wan-t2v": [
|
|
{
|
|
"shape": [
|
|
37800,
|
|
15360,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "synthetic_packed_qkv",
|
|
"prefix": "blocks.*.attn1.packed_qkv"
|
|
},
|
|
{
|
|
"shape": [
|
|
37800,
|
|
5120,
|
|
13824
|
|
],
|
|
"count": 160,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.ffn.net.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
37800,
|
|
13824,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.ffn.net.*.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
37800,
|
|
10240,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "synthetic_packed_kv",
|
|
"prefix": "blocks.*.attn1.packed_kv"
|
|
},
|
|
{
|
|
"shape": [
|
|
37800,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_k"
|
|
},
|
|
{
|
|
"shape": [
|
|
37800,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
37800,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_q"
|
|
},
|
|
{
|
|
"shape": [
|
|
37800,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_v"
|
|
},
|
|
{
|
|
"shape": [
|
|
37800,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
37800,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_q"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
10240,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "synthetic_packed_kv",
|
|
"prefix": "blocks.*.attn2.packed_kv"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
4096
|
|
],
|
|
"count": 384,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
6144,
|
|
4096
|
|
],
|
|
"count": 192,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.qkv_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_k"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 160,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_v"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
5120
|
|
],
|
|
"count": 192,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
2048
|
|
],
|
|
"count": 192,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.o_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
75600,
|
|
64,
|
|
5120
|
|
],
|
|
"count": 4,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "proj_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 4,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
5120,
|
|
4096
|
|
],
|
|
"count": 4,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
30720,
|
|
5120
|
|
],
|
|
"count": 4,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
5120,
|
|
5120
|
|
],
|
|
"count": 4,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
5120,
|
|
256
|
|
],
|
|
"count": 4,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
}
|
|
],
|
|
"wan-ti2v": [
|
|
{
|
|
"shape": [
|
|
18144,
|
|
3072,
|
|
14336
|
|
],
|
|
"count": 60,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.ffn.net.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
14336,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.ffn.net.*.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
9216,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "synthetic_packed_qkv",
|
|
"prefix": "blocks.*.attn1.packed_qkv"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
6144,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "synthetic_packed_kv",
|
|
"prefix": "blocks.*.attn1.packed_kv"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_k"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_q"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn1.to_v"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_q"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
10240,
|
|
4096
|
|
],
|
|
"count": 96,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
18432,
|
|
3072
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
12288,
|
|
4096
|
|
],
|
|
"count": 48,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.qkv_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
10240
|
|
],
|
|
"count": 48,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
6144,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "synthetic_packed_kv",
|
|
"prefix": "blocks.*.attn2.packed_kv"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
4096,
|
|
4096
|
|
],
|
|
"count": 48,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ".encoder.blocks.*.self_attn.SelfAttention.o_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_k"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 60,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "blocks.*.attn2.to_v"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
3072,
|
|
256
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
18144,
|
|
192,
|
|
3072
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "proj_out"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
3072,
|
|
4096
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "0.proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
3072,
|
|
3072
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "2"
|
|
}
|
|
],
|
|
"zimage": [
|
|
{
|
|
"shape": [
|
|
4128,
|
|
20480,
|
|
3840
|
|
],
|
|
"count": 30,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
4128,
|
|
11520,
|
|
3840
|
|
],
|
|
"count": 30,
|
|
"kind": "synthetic_packed_qkv",
|
|
"prefix": "layers.*.attention.packed_qkv"
|
|
},
|
|
{
|
|
"shape": [
|
|
4128,
|
|
3840,
|
|
10240
|
|
],
|
|
"count": 30,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
4128,
|
|
7680,
|
|
3840
|
|
],
|
|
"count": 30,
|
|
"kind": "synthetic_packed_kv",
|
|
"prefix": "layers.*.attention.packed_kv"
|
|
},
|
|
{
|
|
"shape": [
|
|
4128,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 30,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "layers.*.attention.to_k"
|
|
},
|
|
{
|
|
"shape": [
|
|
4128,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 30,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "layers.*.attention.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
4128,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 30,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "layers.*.attention.to_q"
|
|
},
|
|
{
|
|
"shape": [
|
|
4128,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 30,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "layers.*.attention.to_v"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
19456,
|
|
2560
|
|
],
|
|
"count": 36,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "qwen3.layers.*.mlp.gate_up_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
20480,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
2560,
|
|
9728
|
|
],
|
|
"count": 36,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "qwen3.layers.*.mlp.down_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
11520,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "synthetic_packed_qkv",
|
|
"prefix": "noise_refiner.*.attention.packed_qkv"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
3840,
|
|
10240
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
6144,
|
|
2560
|
|
],
|
|
"count": 36,
|
|
"kind": "runtime_fused_qkv",
|
|
"prefix": "qwen3.layers.*.self_attn.qkv_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
7680,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "synthetic_packed_kv",
|
|
"prefix": "noise_refiner.*.attention.packed_kv"
|
|
},
|
|
{
|
|
"shape": [
|
|
512,
|
|
2560,
|
|
4096
|
|
],
|
|
"count": 36,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "qwen3.layers.*.self_attn.o_proj"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "noise_refiner.*.attention.to_k"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "noise_refiner.*.attention.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "noise_refiner.*.attention.to_q"
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "noise_refiner.*.attention.to_v"
|
|
},
|
|
{
|
|
"shape": [
|
|
32,
|
|
20480,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
32,
|
|
11520,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "synthetic_packed_qkv",
|
|
"prefix": "context_refiner.*.attention.packed_qkv"
|
|
},
|
|
{
|
|
"shape": [
|
|
32,
|
|
3840,
|
|
10240
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
32,
|
|
7680,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "synthetic_packed_kv",
|
|
"prefix": "context_refiner.*.attention.packed_kv"
|
|
},
|
|
{
|
|
"shape": [
|
|
4128,
|
|
64,
|
|
3840
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
4096,
|
|
3840,
|
|
64
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
32,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "context_refiner.*.attention.to_k"
|
|
},
|
|
{
|
|
"shape": [
|
|
32,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "context_refiner.*.attention.to_out.*"
|
|
},
|
|
{
|
|
"shape": [
|
|
32,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "context_refiner.*.attention.to_q"
|
|
},
|
|
{
|
|
"shape": [
|
|
32,
|
|
3840,
|
|
3840
|
|
],
|
|
"count": 2,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": "context_refiner.*.attention.to_v"
|
|
},
|
|
{
|
|
"shape": [
|
|
32,
|
|
3840,
|
|
2560
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
15360,
|
|
256
|
|
],
|
|
"count": 32,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
3840,
|
|
256
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
256,
|
|
1024
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
},
|
|
{
|
|
"shape": [
|
|
1,
|
|
1024,
|
|
256
|
|
],
|
|
"count": 1,
|
|
"kind": "actual_runtime_linear",
|
|
"prefix": ""
|
|
}
|
|
]
|
|
}
|