immich-app/ml-models
1
0
Fork 0
mirror of https://github.com/immich-app/ml-models.git synced 2026-09-30 21:27:43 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
df0ea33c8a696ead9fc416a83e8bc1d0c06e4229
ml-models/ci/graphs/ViT-B-16-SigLIP-384__webli/visual.txt
T
Mert df0ea33c8a feat: usable as library, model optimizations (#56) 
* feat: torch-free export core, shared graph/IR helpers, and CLI scaffolding

* feat: fused InsightFace face exporter

* feat: RKNN export path with per-SoC compilation

* feat: single-partition CoreML CLIP export and runtime rewrite API

* feat: refactor to passes, ep-specific fixes and optimizations

* feat: check command diffing committed graph renderings, wired into CI

* chore: graph and rewrite-plan renderings for the catalog

* use pokedex large
2026-08-04 17:46:25 -04:00

626 lines
38 KiB
Plaintext
Raw Blame History

<
ir_version: 10,
opset_import: ["" : 23],
producer_name: "pytorch"
>
main_graph (uint8[batch,384,384,3] image) => (float[batch,768] image_embedding)
<
float[batch,576,768] add_1007
float[batch,576,768] add_1068
float[batch,576,768] add_107
float[batch,576,768] add_1097
float[batch,1,768] add_1195
float[batch,576,768] add_13
float[batch,576,768] add_168
float[batch,576,768] add_197
float[batch,576,768] add_258
float[batch,576,768] add_287
float[batch,576,768] add_348
float[batch,576,768] add_377
float[batch,576,768] add_438
float[batch,576,768] add_467
float[batch,576,768] add_528
float[batch,576,768] add_557
float[batch,576,768] add_618
float[batch,576,768] add_647
float[batch,576,768] add_708
float[batch,576,768] add_737
float[batch,576,768] add_78
float[batch,576,768] add_798
float[batch,576,768] add_827
float[batch,576,768] add_888
float[batch,576,768] add_917
float[batch,576,768] add_978
float[batch,1] clamp_min
float[batch,768,24,24] conv2d
float[batch,576,3072] gelu
float[batch,576,3072] gelu_1
float[batch,576,3072] gelu_10
float[batch,576,3072] gelu_11
float[batch,1,3072] gelu_12
float[batch,576,3072] gelu_2
float[batch,576,3072] gelu_3
float[batch,576,3072] gelu_4
float[batch,576,3072] gelu_5
float[batch,576,3072] gelu_6
float[batch,576,3072] gelu_7
float[batch,576,3072] gelu_8
float[batch,576,3072] gelu_9
float[batch,3,384,384] image_chw
float[batch] image_ez
float[batch] image_ez_r
float[batch,1,1,1] image_ez_s
float[batch,384,384,3] image_f32
float[batch,576,768] layer_norm
float[batch,576,768] layer_norm_1
float[batch,576,768] layer_norm_10
float[batch,576,768] layer_norm_11
float[batch,576,768] layer_norm_12
float[batch,576,768] layer_norm_13
float[batch,576,768] layer_norm_14
float[batch,576,768] layer_norm_15
float[batch,576,768] layer_norm_16
float[batch,576,768] layer_norm_17
float[batch,576,768] layer_norm_18
float[batch,576,768] layer_norm_19
float[batch,576,768] layer_norm_2
float[batch,576,768] layer_norm_20
float[batch,576,768] layer_norm_21
float[batch,576,768] layer_norm_22
float[batch,576,768] layer_norm_23
float[batch,576,768] layer_norm_24
float[batch,1,768] layer_norm_25
float[batch,576,768] layer_norm_3
float[batch,576,768] layer_norm_4
float[batch,576,768] layer_norm_5
float[batch,576,768] layer_norm_6
float[batch,576,768] layer_norm_7
float[batch,576,768] layer_norm_8
float[batch,576,768] layer_norm_9
float[batch,1] linalg_vector_norm
float[batch,576,2304] linear
float[batch,576,768] linear_1
float[batch,576,3072] linear_10
float[batch,576,768] linear_11
float[batch,576,2304] linear_12
float[batch,576,768] linear_13
float[batch,576,3072] linear_14
float[batch,576,768] linear_15
float[batch,576,2304] linear_16
float[batch,576,768] linear_17
float[batch,576,3072] linear_18
float[batch,576,768] linear_19
float[batch,576,3072] linear_2
float[batch,576,2304] linear_20
float[batch,576,768] linear_21
float[batch,576,3072] linear_22
float[batch,576,768] linear_23
float[batch,576,2304] linear_24
float[batch,576,768] linear_25
float[batch,576,3072] linear_26
float[batch,576,768] linear_27
float[batch,576,2304] linear_28
float[batch,576,768] linear_29
float[batch,576,768] linear_3
float[batch,576,3072] linear_30
float[batch,576,768] linear_31
float[batch,576,2304] linear_32
float[batch,576,768] linear_33
float[batch,576,3072] linear_34
float[batch,576,768] linear_35
float[batch,576,2304] linear_36
float[batch,576,768] linear_37
float[batch,576,3072] linear_38
float[batch,576,768] linear_39
float[batch,576,2304] linear_4
float[batch,576,2304] linear_40
float[batch,576,768] linear_41
float[batch,576,3072] linear_42
float[batch,576,768] linear_43
float[batch,576,2304] linear_44
float[batch,576,768] linear_45
float[batch,576,3072] linear_46
float[batch,576,768] linear_47
float[batch,576,1536] linear_49
float[batch,576,768] linear_5
float[batch,1,768] linear_50
float[batch,1,3072] linear_51
float[batch,1,768] linear_52
float[batch,576,3072] linear_6
float[batch,576,768] linear_7
float[batch,576,2304] linear_8
float[batch,576,768] linear_9
float[batch,576,768] node_scaled_dot_product_attention_10_k
float[batch,576,768] node_scaled_dot_product_attention_10_q
float[batch,576,768] node_scaled_dot_product_attention_10_v
float[batch,576,768] node_scaled_dot_product_attention_11_k
float[batch,576,768] node_scaled_dot_product_attention_11_q
float[batch,576,768] node_scaled_dot_product_attention_11_v
float[batch,576,768] node_scaled_dot_product_attention_12_k
float[batch,1,768] node_scaled_dot_product_attention_12_q
float[batch,1,1] node_scaled_dot_product_attention_12_q_col_out
float[batch,576,768] node_scaled_dot_product_attention_12_v
float[batch,576,768] node_scaled_dot_product_attention_1_k
float[batch,576,768] node_scaled_dot_product_attention_1_q
float[batch,576,768] node_scaled_dot_product_attention_1_v
float[batch,576,768] node_scaled_dot_product_attention_2_k
float[batch,576,768] node_scaled_dot_product_attention_2_q
float[batch,576,768] node_scaled_dot_product_attention_2_v
float[batch,576,768] node_scaled_dot_product_attention_3_k
float[batch,576,768] node_scaled_dot_product_attention_3_q
float[batch,576,768] node_scaled_dot_product_attention_3_v
float[batch,576,768] node_scaled_dot_product_attention_4_k
float[batch,576,768] node_scaled_dot_product_attention_4_q
float[batch,576,768] node_scaled_dot_product_attention_4_v
float[batch,576,768] node_scaled_dot_product_attention_5_k
float[batch,576,768] node_scaled_dot_product_attention_5_q
float[batch,576,768] node_scaled_dot_product_attention_5_v
float[batch,576,768] node_scaled_dot_product_attention_6_k
float[batch,576,768] node_scaled_dot_product_attention_6_q
float[batch,576,768] node_scaled_dot_product_attention_6_v
float[batch,576,768] node_scaled_dot_product_attention_7_k
float[batch,576,768] node_scaled_dot_product_attention_7_q
float[batch,576,768] node_scaled_dot_product_attention_7_v
float[batch,576,768] node_scaled_dot_product_attention_8_k
float[batch,576,768] node_scaled_dot_product_attention_8_q
float[batch,576,768] node_scaled_dot_product_attention_8_v
float[batch,576,768] node_scaled_dot_product_attention_9_k
float[batch,576,768] node_scaled_dot_product_attention_9_q
float[batch,576,768] node_scaled_dot_product_attention_9_v
float[batch,576,768] node_scaled_dot_product_attention_k
float[batch,576,768] node_scaled_dot_product_attention_q
float[batch,576,768] node_scaled_dot_product_attention_v
float[batch,576,768] scaled_dot_product_attention
float[batch,576,768] scaled_dot_product_attention_1
float[batch,576,768] scaled_dot_product_attention_10
float[batch,576,768] scaled_dot_product_attention_11
float[batch,1,768] scaled_dot_product_attention_12
float[batch,576,768] scaled_dot_product_attention_2
float[batch,576,768] scaled_dot_product_attention_3
float[batch,576,768] scaled_dot_product_attention_4
float[batch,576,768] scaled_dot_product_attention_5
float[batch,576,768] scaled_dot_product_attention_6
float[batch,576,768] scaled_dot_product_attention_7
float[batch,576,768] scaled_dot_product_attention_8
float[batch,576,768] scaled_dot_product_attention_9
float[batch,768] select
float[batch,576,768] transpose
float[batch,576,768] val_100
float[batch,576,3072] val_101
float[batch,576,768] val_102
float[batch,576,1536] val_103
float[batch,1,768] val_104
float[batch,1,3072] val_105
float[batch,1,768] val_106
float[batch,576,2304] val_55
float[batch,576,768] val_56
float[batch,576,3072] val_57
float[batch,576,768] val_58
float[batch,576,2304] val_59
float[batch,576,768] val_60
float[batch,576,3072] val_61
float[batch,576,768] val_62
float[batch,576,2304] val_63
float[batch,576,768] val_64
float[batch,576,3072] val_65
float[batch,576,768] val_66
float[batch,576,2304] val_67
float[batch,576,768] val_68
float[batch,576,3072] val_69
float[batch,576,768] val_70
float[batch,576,2304] val_71
float[batch,576,768] val_72
float[batch,576,3072] val_73
float[batch,576,768] val_74
float[batch,576,2304] val_75
float[batch,576,768] val_76
float[batch,576,3072] val_77
float[batch,576,768] val_78
float[batch,576,2304] val_79
float[batch,576,768] val_80
float[batch,576,3072] val_81
float[batch,576,768] val_82
float[batch,576,2304] val_83
float[batch,576,768] val_84
float[batch,576,3072] val_85
float[batch,576,768] val_86
float[batch,576,2304] val_87
float[batch,576,768] val_88
float[batch,576,3072] val_89
float[batch,576,768] val_90
float[batch,576,2304] val_91
float[batch,576,768] val_92
float[batch,576,3072] val_93
float[batch,576,768] val_94
float[batch,576,2304] val_95
float[batch,576,768] val_96
float[batch,576,3072] val_97
float[batch,576,768] val_98
float[batch,576,2304] val_99
float[batch,768,576] view
>
{
[pre_cast] image_f32 = Cast <to: int = 1> (image)
[pre_nhwc_to_nchw] image_chw = Transpose <perm: ints = [0, 3, 1, 2]> (image_f32)
[node_conv2d] conv2d = Conv <auto_pad: string = "NOTSET", dilations: ints = [1, 1], group: int = 1, pads: ints = [0, 0, 0, 0], strides: ints = [16, 16]> (image_chw, "visual.trunk.patch_embed.proj.weight", "visual.trunk.patch_embed.proj.bias")
view = Reshape <allowzero: int = 1> (conv2d, view_target)
[node_transpose] transpose = Transpose <perm: ints = [0, 2, 1]> (view)
add_13 = Add (transpose, "visual.trunk.pos_embed")
[node_layer_norm] layer_norm = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_13, "visual.trunk.blocks.0.norm1.weight", "visual.trunk.blocks.0.norm1.bias")
val_55 = MatMul (layer_norm, val_3)
[node_linear] linear = Add (val_55, "visual.trunk.blocks.0.attn.qkv.bias")
[node_scaled_dot_product_attention_qkv_split] node_scaled_dot_product_attention_q, node_scaled_dot_product_attention_k, node_scaled_dot_product_attention_v = Split <axis: int = -1> (linear, attn3d_split_3x768)
[node_scaled_dot_product_attention] scaled_dot_product_attention = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_q, node_scaled_dot_product_attention_k, node_scaled_dot_product_attention_v)
val_56 = MatMul (scaled_dot_product_attention, val_4)
linear_1 = Add (val_56, "visual.trunk.blocks.0.attn.proj.bias")
add_78 = Add (add_13, linear_1)
layer_norm_1 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_78, "visual.trunk.blocks.0.norm2.weight", "visual.trunk.blocks.0.norm2.bias")
val_57 = MatMul (layer_norm_1, val_5)
linear_2 = Add (val_57, "visual.trunk.blocks.0.mlp.fc1.bias")
[node_gelu] gelu = Gelu <approximate: string = "none"> (linear_2)
val_58 = MatMul (gelu, val_6)
linear_3 = Add (val_58, "visual.trunk.blocks.0.mlp.fc2.bias")
add_107 = Add (add_78, linear_3)
layer_norm_2 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_107, "visual.trunk.blocks.1.norm1.weight", "visual.trunk.blocks.1.norm1.bias")
val_59 = MatMul (layer_norm_2, val_7)
linear_4 = Add (val_59, "visual.trunk.blocks.1.attn.qkv.bias")
[node_scaled_dot_product_attention_1_qkv_split] node_scaled_dot_product_attention_1_q, node_scaled_dot_product_attention_1_k, node_scaled_dot_product_attention_1_v = Split <axis: int = -1> (linear_4, attn3d_split_3x768)
scaled_dot_product_attention_1 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_1_q, node_scaled_dot_product_attention_1_k, node_scaled_dot_product_attention_1_v)
val_60 = MatMul (scaled_dot_product_attention_1, val_8)
linear_5 = Add (val_60, "visual.trunk.blocks.1.attn.proj.bias")
add_168 = Add (add_107, linear_5)
layer_norm_3 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_168, "visual.trunk.blocks.1.norm2.weight", "visual.trunk.blocks.1.norm2.bias")
val_61 = MatMul (layer_norm_3, val_9)
linear_6 = Add (val_61, "visual.trunk.blocks.1.mlp.fc1.bias")
gelu_1 = Gelu <approximate: string = "none"> (linear_6)
val_62 = MatMul (gelu_1, val_10)
linear_7 = Add (val_62, "visual.trunk.blocks.1.mlp.fc2.bias")
add_197 = Add (add_168, linear_7)
layer_norm_4 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_197, "visual.trunk.blocks.2.norm1.weight", "visual.trunk.blocks.2.norm1.bias")
val_63 = MatMul (layer_norm_4, val_11)
linear_8 = Add (val_63, "visual.trunk.blocks.2.attn.qkv.bias")
[node_scaled_dot_product_attention_2_qkv_split] node_scaled_dot_product_attention_2_q, node_scaled_dot_product_attention_2_k, node_scaled_dot_product_attention_2_v = Split <axis: int = -1> (linear_8, attn3d_split_3x768)
scaled_dot_product_attention_2 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_2_q, node_scaled_dot_product_attention_2_k, node_scaled_dot_product_attention_2_v)
val_64 = MatMul (scaled_dot_product_attention_2, val_12)
linear_9 = Add (val_64, "visual.trunk.blocks.2.attn.proj.bias")
add_258 = Add (add_197, linear_9)
layer_norm_5 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_258, "visual.trunk.blocks.2.norm2.weight", "visual.trunk.blocks.2.norm2.bias")
val_65 = MatMul (layer_norm_5, val_13)
linear_10 = Add (val_65, "visual.trunk.blocks.2.mlp.fc1.bias")
gelu_2 = Gelu <approximate: string = "none"> (linear_10)
val_66 = MatMul (gelu_2, val_14)
linear_11 = Add (val_66, "visual.trunk.blocks.2.mlp.fc2.bias")
add_287 = Add (add_258, linear_11)
layer_norm_6 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_287, "visual.trunk.blocks.3.norm1.weight", "visual.trunk.blocks.3.norm1.bias")
val_67 = MatMul (layer_norm_6, val_15)
linear_12 = Add (val_67, "visual.trunk.blocks.3.attn.qkv.bias")
[node_scaled_dot_product_attention_3_qkv_split] node_scaled_dot_product_attention_3_q, node_scaled_dot_product_attention_3_k, node_scaled_dot_product_attention_3_v = Split <axis: int = -1> (linear_12, attn3d_split_3x768)
scaled_dot_product_attention_3 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_3_q, node_scaled_dot_product_attention_3_k, node_scaled_dot_product_attention_3_v)
val_68 = MatMul (scaled_dot_product_attention_3, val_16)
linear_13 = Add (val_68, "visual.trunk.blocks.3.attn.proj.bias")
add_348 = Add (add_287, linear_13)
layer_norm_7 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_348, "visual.trunk.blocks.3.norm2.weight", "visual.trunk.blocks.3.norm2.bias")
val_69 = MatMul (layer_norm_7, val_17)
linear_14 = Add (val_69, "visual.trunk.blocks.3.mlp.fc1.bias")
gelu_3 = Gelu <approximate: string = "none"> (linear_14)
val_70 = MatMul (gelu_3, val_18)
linear_15 = Add (val_70, "visual.trunk.blocks.3.mlp.fc2.bias")
add_377 = Add (add_348, linear_15)
layer_norm_8 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_377, "visual.trunk.blocks.4.norm1.weight", "visual.trunk.blocks.4.norm1.bias")
val_71 = MatMul (layer_norm_8, val_19)
linear_16 = Add (val_71, "visual.trunk.blocks.4.attn.qkv.bias")
[node_scaled_dot_product_attention_4_qkv_split] node_scaled_dot_product_attention_4_q, node_scaled_dot_product_attention_4_k, node_scaled_dot_product_attention_4_v = Split <axis: int = -1> (linear_16, attn3d_split_3x768)
scaled_dot_product_attention_4 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_4_q, node_scaled_dot_product_attention_4_k, node_scaled_dot_product_attention_4_v)
val_72 = MatMul (scaled_dot_product_attention_4, val_20)
linear_17 = Add (val_72, "visual.trunk.blocks.4.attn.proj.bias")
add_438 = Add (add_377, linear_17)
layer_norm_9 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_438, "visual.trunk.blocks.4.norm2.weight", "visual.trunk.blocks.4.norm2.bias")
val_73 = MatMul (layer_norm_9, val_21)
linear_18 = Add (val_73, "visual.trunk.blocks.4.mlp.fc1.bias")
gelu_4 = Gelu <approximate: string = "none"> (linear_18)
val_74 = MatMul (gelu_4, val_22)
linear_19 = Add (val_74, "visual.trunk.blocks.4.mlp.fc2.bias")
add_467 = Add (add_438, linear_19)
layer_norm_10 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_467, "visual.trunk.blocks.5.norm1.weight", "visual.trunk.blocks.5.norm1.bias")
val_75 = MatMul (layer_norm_10, val_23)
linear_20 = Add (val_75, "visual.trunk.blocks.5.attn.qkv.bias")
[node_scaled_dot_product_attention_5_qkv_split] node_scaled_dot_product_attention_5_q, node_scaled_dot_product_attention_5_k, node_scaled_dot_product_attention_5_v = Split <axis: int = -1> (linear_20, attn3d_split_3x768)
scaled_dot_product_attention_5 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_5_q, node_scaled_dot_product_attention_5_k, node_scaled_dot_product_attention_5_v)
val_76 = MatMul (scaled_dot_product_attention_5, val_24)
linear_21 = Add (val_76, "visual.trunk.blocks.5.attn.proj.bias")
add_528 = Add (add_467, linear_21)
layer_norm_11 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_528, "visual.trunk.blocks.5.norm2.weight", "visual.trunk.blocks.5.norm2.bias")
val_77 = MatMul (layer_norm_11, val_25)
linear_22 = Add (val_77, "visual.trunk.blocks.5.mlp.fc1.bias")
gelu_5 = Gelu <approximate: string = "none"> (linear_22)
val_78 = MatMul (gelu_5, val_26)
linear_23 = Add (val_78, "visual.trunk.blocks.5.mlp.fc2.bias")
add_557 = Add (add_528, linear_23)
layer_norm_12 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_557, "visual.trunk.blocks.6.norm1.weight", "visual.trunk.blocks.6.norm1.bias")
val_79 = MatMul (layer_norm_12, val_27)
linear_24 = Add (val_79, "visual.trunk.blocks.6.attn.qkv.bias")
[node_scaled_dot_product_attention_6_qkv_split] node_scaled_dot_product_attention_6_q, node_scaled_dot_product_attention_6_k, node_scaled_dot_product_attention_6_v = Split <axis: int = -1> (linear_24, attn3d_split_3x768)
scaled_dot_product_attention_6 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_6_q, node_scaled_dot_product_attention_6_k, node_scaled_dot_product_attention_6_v)
val_80 = MatMul (scaled_dot_product_attention_6, val_28)
linear_25 = Add (val_80, "visual.trunk.blocks.6.attn.proj.bias")
add_618 = Add (add_557, linear_25)
layer_norm_13 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_618, "visual.trunk.blocks.6.norm2.weight", "visual.trunk.blocks.6.norm2.bias")
val_81 = MatMul (layer_norm_13, val_29)
linear_26 = Add (val_81, "visual.trunk.blocks.6.mlp.fc1.bias")
gelu_6 = Gelu <approximate: string = "none"> (linear_26)
val_82 = MatMul (gelu_6, val_30)
linear_27 = Add (val_82, "visual.trunk.blocks.6.mlp.fc2.bias")
add_647 = Add (add_618, linear_27)
layer_norm_14 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_647, "visual.trunk.blocks.7.norm1.weight", "visual.trunk.blocks.7.norm1.bias")
val_83 = MatMul (layer_norm_14, val_31)
linear_28 = Add (val_83, "visual.trunk.blocks.7.attn.qkv.bias")
[node_scaled_dot_product_attention_7_qkv_split] node_scaled_dot_product_attention_7_q, node_scaled_dot_product_attention_7_k, node_scaled_dot_product_attention_7_v = Split <axis: int = -1> (linear_28, attn3d_split_3x768)
scaled_dot_product_attention_7 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_7_q, node_scaled_dot_product_attention_7_k, node_scaled_dot_product_attention_7_v)
val_84 = MatMul (scaled_dot_product_attention_7, val_32)
linear_29 = Add (val_84, "visual.trunk.blocks.7.attn.proj.bias")
add_708 = Add (add_647, linear_29)
layer_norm_15 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_708, "visual.trunk.blocks.7.norm2.weight", "visual.trunk.blocks.7.norm2.bias")
val_85 = MatMul (layer_norm_15, val_33)
linear_30 = Add (val_85, "visual.trunk.blocks.7.mlp.fc1.bias")
gelu_7 = Gelu <approximate: string = "none"> (linear_30)
val_86 = MatMul (gelu_7, val_34)
linear_31 = Add (val_86, "visual.trunk.blocks.7.mlp.fc2.bias")
add_737 = Add (add_708, linear_31)
layer_norm_16 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_737, "visual.trunk.blocks.8.norm1.weight", "visual.trunk.blocks.8.norm1.bias")
val_87 = MatMul (layer_norm_16, val_35)
linear_32 = Add (val_87, "visual.trunk.blocks.8.attn.qkv.bias")
[node_scaled_dot_product_attention_8_qkv_split] node_scaled_dot_product_attention_8_q, node_scaled_dot_product_attention_8_k, node_scaled_dot_product_attention_8_v = Split <axis: int = -1> (linear_32, attn3d_split_3x768)
scaled_dot_product_attention_8 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_8_q, node_scaled_dot_product_attention_8_k, node_scaled_dot_product_attention_8_v)
val_88 = MatMul (scaled_dot_product_attention_8, val_36)
linear_33 = Add (val_88, "visual.trunk.blocks.8.attn.proj.bias")
add_798 = Add (add_737, linear_33)
layer_norm_17 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_798, "visual.trunk.blocks.8.norm2.weight", "visual.trunk.blocks.8.norm2.bias")
val_89 = MatMul (layer_norm_17, val_37)
linear_34 = Add (val_89, "visual.trunk.blocks.8.mlp.fc1.bias")
gelu_8 = Gelu <approximate: string = "none"> (linear_34)
val_90 = MatMul (gelu_8, val_38)
linear_35 = Add (val_90, "visual.trunk.blocks.8.mlp.fc2.bias")
add_827 = Add (add_798, linear_35)
layer_norm_18 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_827, "visual.trunk.blocks.9.norm1.weight", "visual.trunk.blocks.9.norm1.bias")
val_91 = MatMul (layer_norm_18, val_39)
linear_36 = Add (val_91, "visual.trunk.blocks.9.attn.qkv.bias")
[node_scaled_dot_product_attention_9_qkv_split] node_scaled_dot_product_attention_9_q, node_scaled_dot_product_attention_9_k, node_scaled_dot_product_attention_9_v = Split <axis: int = -1> (linear_36, attn3d_split_3x768)
scaled_dot_product_attention_9 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_9_q, node_scaled_dot_product_attention_9_k, node_scaled_dot_product_attention_9_v)
val_92 = MatMul (scaled_dot_product_attention_9, val_40)
linear_37 = Add (val_92, "visual.trunk.blocks.9.attn.proj.bias")
add_888 = Add (add_827, linear_37)
layer_norm_19 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_888, "visual.trunk.blocks.9.norm2.weight", "visual.trunk.blocks.9.norm2.bias")
val_93 = MatMul (layer_norm_19, val_41)
linear_38 = Add (val_93, "visual.trunk.blocks.9.mlp.fc1.bias")
gelu_9 = Gelu <approximate: string = "none"> (linear_38)
val_94 = MatMul (gelu_9, val_42)
linear_39 = Add (val_94, "visual.trunk.blocks.9.mlp.fc2.bias")
add_917 = Add (add_888, linear_39)
layer_norm_20 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_917, "visual.trunk.blocks.10.norm1.weight", "visual.trunk.blocks.10.norm1.bias")
val_95 = MatMul (layer_norm_20, val_43)
linear_40 = Add (val_95, "visual.trunk.blocks.10.attn.qkv.bias")
[node_scaled_dot_product_attention_10_qkv_split] node_scaled_dot_product_attention_10_q, node_scaled_dot_product_attention_10_k, node_scaled_dot_product_attention_10_v = Split <axis: int = -1> (linear_40, attn3d_split_3x768)
scaled_dot_product_attention_10 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_10_q, node_scaled_dot_product_attention_10_k, node_scaled_dot_product_attention_10_v)
val_96 = MatMul (scaled_dot_product_attention_10, val_44)
linear_41 = Add (val_96, "visual.trunk.blocks.10.attn.proj.bias")
add_978 = Add (add_917, linear_41)
layer_norm_21 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_978, "visual.trunk.blocks.10.norm2.weight", "visual.trunk.blocks.10.norm2.bias")
val_97 = MatMul (layer_norm_21, val_45)
linear_42 = Add (val_97, "visual.trunk.blocks.10.mlp.fc1.bias")
gelu_10 = Gelu <approximate: string = "none"> (linear_42)
val_98 = MatMul (gelu_10, val_46)
linear_43 = Add (val_98, "visual.trunk.blocks.10.mlp.fc2.bias")
add_1007 = Add (add_978, linear_43)
layer_norm_22 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_1007, "visual.trunk.blocks.11.norm1.weight", "visual.trunk.blocks.11.norm1.bias")
val_99 = MatMul (layer_norm_22, val_47)
linear_44 = Add (val_99, "visual.trunk.blocks.11.attn.qkv.bias")
[node_scaled_dot_product_attention_11_qkv_split] node_scaled_dot_product_attention_11_q, node_scaled_dot_product_attention_11_k, node_scaled_dot_product_attention_11_v = Split <axis: int = -1> (linear_44, attn3d_split_3x768)
scaled_dot_product_attention_11 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_11_q, node_scaled_dot_product_attention_11_k, node_scaled_dot_product_attention_11_v)
val_100 = MatMul (scaled_dot_product_attention_11, val_48)
linear_45 = Add (val_100, "visual.trunk.blocks.11.attn.proj.bias")
add_1068 = Add (add_1007, linear_45)
layer_norm_23 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_1068, "visual.trunk.blocks.11.norm2.weight", "visual.trunk.blocks.11.norm2.bias")
val_101 = MatMul (layer_norm_23, val_49)
linear_46 = Add (val_101, "visual.trunk.blocks.11.mlp.fc1.bias")
gelu_11 = Gelu <approximate: string = "none"> (linear_46)
val_102 = MatMul (gelu_11, val_50)
linear_47 = Add (val_102, "visual.trunk.blocks.11.mlp.fc2.bias")
add_1097 = Add (add_1068, linear_47)
layer_norm_24 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (add_1097, "visual.trunk.norm.weight", "visual.trunk.norm.bias")
[ez_slice] image_ez_s = Slice (image_f32, image_ez_starts, image_ez_ends, image_ez_axes)
[ez_flatten] image_ez_r = Reshape (image_ez_s, val_0)
[ez_mul] image_ez = Mul (image_ez_r, image_ez_zero)
val_103 = MatMul (layer_norm_24, val_51)
linear_49 = Add (val_103, "visual.trunk.attn_pool.kv.bias")
[node_scaled_dot_product_attention_12_qkv_split] node_scaled_dot_product_attention_12_k, node_scaled_dot_product_attention_12_v = Split <axis: int = -1> (linear_49, attn3d_split_2x768)
[node_scaled_dot_product_attention_12_q_col] node_scaled_dot_product_attention_12_q_col_out = Unsqueeze (image_ez, node_scaled_dot_product_attention_12_q_col_axes)
[node_scaled_dot_product_attention_12_q_bcast] node_scaled_dot_product_attention_12_q = Add (node_scaled_dot_product_attention_12_q3, node_scaled_dot_product_attention_12_q_col_out)
scaled_dot_product_attention_12 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, softcap: float = 0> (node_scaled_dot_product_attention_12_q, node_scaled_dot_product_attention_12_k, node_scaled_dot_product_attention_12_v)
val_104 = MatMul (scaled_dot_product_attention_12, val_52)
linear_50 = Add (val_104, "visual.trunk.attn_pool.proj.bias")
layer_norm_25 = LayerNormalization <axis: int = -1, epsilon: float = 1e-06, stash_type: int = 1> (linear_50, "visual.trunk.attn_pool.norm.weight", "visual.trunk.attn_pool.norm.bias")
val_105 = MatMul (layer_norm_25, val_53)
linear_51 = Add (val_105, "visual.trunk.attn_pool.mlp.fc1.bias")
gelu_12 = Gelu <approximate: string = "none"> (linear_51)
val_106 = MatMul (gelu_12, val_54)
linear_52 = Add (val_106, "visual.trunk.attn_pool.mlp.fc2.bias")
add_1195 = Add (linear_50, linear_52)
select = Squeeze (add_1195, val_2)
[node_linalg_vector_norm] linalg_vector_norm = ReduceL2 <keepdims: int = 1, noop_with_empty_axes: int = 0> (select, val_0)
[node_clamp_min] clamp_min = Clip (linalg_vector_norm, val_1)
[node_div] image_embedding = Div (select, clamp_min)
}
weights:
attn3d_split_2x768 INT64[2] 6b8b40015c71
attn3d_split_3x768 INT64[3] eca50b2daf34
image_ez_axes INT64[3] e2e2033ae7e1
image_ez_ends INT64[3] 605390e5a369
image_ez_starts INT64[3] 9d908ecfb6b2
image_ez_zero FLOAT[] df3f619804a9
node_scaled_dot_product_attention_12_q3 FLOAT[1,1,768] dffc0d264ff1
node_scaled_dot_product_attention_12_q_col_axes INT64[2] 0c730b69905c
val_0 INT64[1] 12a3ae445661
val_1 FLOAT[] 6708d9be4956
val_10 FLOAT[3072,768] 32cd85868898
val_11 FLOAT[768,2304] 0c95f36ac6d8
val_12 FLOAT[768,768] dc0376d94bd3
val_13 FLOAT[768,3072] d32dcfd11534
val_14 FLOAT[3072,768] ec29e8c9b643
val_15 FLOAT[768,2304] 807c2301ee50
val_16 FLOAT[768,768] 188e9408f9d1
val_17 FLOAT[768,3072] ba83e1632c83
val_18 FLOAT[3072,768] ce69886f3a25
val_19 FLOAT[768,2304] f80dfd5f872e
val_2 INT64[1] 7c9fa136d441
val_20 FLOAT[768,768] 300dc013ea62
val_21 FLOAT[768,3072] 8d41909dffcc
val_22 FLOAT[3072,768] 40f90beee17e
val_23 FLOAT[768,2304] e10a2f2b9e15
val_24 FLOAT[768,768] f56e07886a8e
val_25 FLOAT[768,3072] 961b5e11a6c5
val_26 FLOAT[3072,768] 93bf5664b0b4
val_27 FLOAT[768,2304] 1978daad9423
val_28 FLOAT[768,768] f3ea6414c3e8
val_29 FLOAT[768,3072] 71619d740a6b
val_3 FLOAT[768,2304] 72ec385c7c4b
val_30 FLOAT[3072,768] 3aea33a66e61
val_31 FLOAT[768,2304] 52e79e2b56a9
val_32 FLOAT[768,768] 33505e27cf19
val_33 FLOAT[768,3072] 422c73ab1e1f
val_34 FLOAT[3072,768] 188e76ccc2ae
val_35 FLOAT[768,2304] 48497931487f
val_36 FLOAT[768,768] 5bec50b866d7
val_37 FLOAT[768,3072] f5017e94bcdd
val_38 FLOAT[3072,768] 1d962a7c6db9
val_39 FLOAT[768,2304] 666ec36c4f2f
val_4 FLOAT[768,768] de718acfb152
val_40 FLOAT[768,768] 37e9f2efb165
val_41 FLOAT[768,3072] 7a23d8b465d8
val_42 FLOAT[3072,768] 43c8c05ceb08
val_43 FLOAT[768,2304] 24740324f5b8
val_44 FLOAT[768,768] 5ea9b10dad3e
val_45 FLOAT[768,3072] 3b8aa8a5871f
val_46 FLOAT[3072,768] 6918855e7f0c
val_47 FLOAT[768,2304] e21fd24e37c6
val_48 FLOAT[768,768] 596e5623a177
val_49 FLOAT[768,3072] c4756558133d
val_5 FLOAT[768,3072] c9dad04cda7e
val_50 FLOAT[3072,768] 00b4b147ba1c
val_51 FLOAT[768,1536] d9a820c601e1
val_52 FLOAT[768,768] 2850e68de8ab
val_53 FLOAT[768,3072] 630e2665c2b3
val_54 FLOAT[3072,768] bd2bbab71426
val_6 FLOAT[3072,768] b950b44d8157
val_7 FLOAT[768,2304] ddbfcdb6ded7
val_8 FLOAT[768,768] d3126cb5cc9e
val_9 FLOAT[768,3072] 326343a14a92
view_target INT64[3] 8d30e545cb18
visual.trunk.attn_pool.kv.bias FLOAT[1536] 5ea0f83309db
visual.trunk.attn_pool.mlp.fc1.bias FLOAT[3072] 82abe8523fc5
visual.trunk.attn_pool.mlp.fc2.bias FLOAT[768] f36c88ccdd1c
visual.trunk.attn_pool.norm.bias FLOAT[768] c8e6a0807c68
visual.trunk.attn_pool.norm.weight FLOAT[768] e835a27353ba
visual.trunk.attn_pool.proj.bias FLOAT[768] 4d52d9b8b89f
visual.trunk.blocks.0.attn.proj.bias FLOAT[768] fb32c6be05e3
visual.trunk.blocks.0.attn.qkv.bias FLOAT[2304] 55e049eacf53
visual.trunk.blocks.0.mlp.fc1.bias FLOAT[3072] b9b570ce2dea
visual.trunk.blocks.0.mlp.fc2.bias FLOAT[768] fd77992c093c
visual.trunk.blocks.0.norm1.bias FLOAT[768] 7379ec576226
visual.trunk.blocks.0.norm1.weight FLOAT[768] 9e4c9a1b13b4
visual.trunk.blocks.0.norm2.bias FLOAT[768] f08a93181100
visual.trunk.blocks.0.norm2.weight FLOAT[768] b8011a1d95f1
visual.trunk.blocks.1.attn.proj.bias FLOAT[768] 1cbc4b64b3e8
visual.trunk.blocks.1.attn.qkv.bias FLOAT[2304] 0b8c21d1ff53
visual.trunk.blocks.1.mlp.fc1.bias FLOAT[3072] a2adbf3d4bbb
visual.trunk.blocks.1.mlp.fc2.bias FLOAT[768] a959e5bc92b9
visual.trunk.blocks.1.norm1.bias FLOAT[768] 56ce557ea4bf
visual.trunk.blocks.1.norm1.weight FLOAT[768] 4ddc2b23debf
visual.trunk.blocks.1.norm2.bias FLOAT[768] 7194924db9f7
visual.trunk.blocks.1.norm2.weight FLOAT[768] a8abd52f9560
visual.trunk.blocks.10.attn.proj.bias FLOAT[768] 0f20cecf99fe
visual.trunk.blocks.10.attn.qkv.bias FLOAT[2304] 19e6fe9db881
visual.trunk.blocks.10.mlp.fc1.bias FLOAT[3072] f1afe1d3b7b1
visual.trunk.blocks.10.mlp.fc2.bias FLOAT[768] b8d433a102d7
visual.trunk.blocks.10.norm1.bias FLOAT[768] 8e627d2fa6de
visual.trunk.blocks.10.norm1.weight FLOAT[768] 3d73b0510054
visual.trunk.blocks.10.norm2.bias FLOAT[768] cbdd4a408d1e
visual.trunk.blocks.10.norm2.weight FLOAT[768] 43bfc6ab20d1
visual.trunk.blocks.11.attn.proj.bias FLOAT[768] 44482ba61475
visual.trunk.blocks.11.attn.qkv.bias FLOAT[2304] e059ac53b0fc
visual.trunk.blocks.11.mlp.fc1.bias FLOAT[3072] 7e19edd90652
visual.trunk.blocks.11.mlp.fc2.bias FLOAT[768] b39fc33767ba
visual.trunk.blocks.11.norm1.bias FLOAT[768] a271ce503b17
visual.trunk.blocks.11.norm1.weight FLOAT[768] d98f5440e39a
visual.trunk.blocks.11.norm2.bias FLOAT[768] 35f33989da53
visual.trunk.blocks.11.norm2.weight FLOAT[768] 988b9e3b7e42
visual.trunk.blocks.2.attn.proj.bias FLOAT[768] fcbb3274615d
visual.trunk.blocks.2.attn.qkv.bias FLOAT[2304] 28e262d42ccb
visual.trunk.blocks.2.mlp.fc1.bias FLOAT[3072] 38f84a2c14f5
visual.trunk.blocks.2.mlp.fc2.bias FLOAT[768] 7f2f7e1c60a6
visual.trunk.blocks.2.norm1.bias FLOAT[768] 535579aeb4c9
visual.trunk.blocks.2.norm1.weight FLOAT[768] 2384c7e1e9d9
visual.trunk.blocks.2.norm2.bias FLOAT[768] 92d37fb6398d
visual.trunk.blocks.2.norm2.weight FLOAT[768] 09e1542afd4f
visual.trunk.blocks.3.attn.proj.bias FLOAT[768] 7011ffd49dbe
visual.trunk.blocks.3.attn.qkv.bias FLOAT[2304] 85e3c91f6f45
visual.trunk.blocks.3.mlp.fc1.bias FLOAT[3072] 655f94f08630
visual.trunk.blocks.3.mlp.fc2.bias FLOAT[768] 40062e1d2dfc
visual.trunk.blocks.3.norm1.bias FLOAT[768] 5163b075cc33
visual.trunk.blocks.3.norm1.weight FLOAT[768] b6303a1bf974
visual.trunk.blocks.3.norm2.bias FLOAT[768] b89747ae9c38
visual.trunk.blocks.3.norm2.weight FLOAT[768] 53330c999e0c
visual.trunk.blocks.4.attn.proj.bias FLOAT[768] 9ffd9b3ea44b
visual.trunk.blocks.4.attn.qkv.bias FLOAT[2304] 7783677dfe2e
visual.trunk.blocks.4.mlp.fc1.bias FLOAT[3072] 3d45a716fe0a
visual.trunk.blocks.4.mlp.fc2.bias FLOAT[768] b593d876de9d
visual.trunk.blocks.4.norm1.bias FLOAT[768] ecb8263f1456
visual.trunk.blocks.4.norm1.weight FLOAT[768] b57a681c5c74
visual.trunk.blocks.4.norm2.bias FLOAT[768] f2a9edc8f8fa
visual.trunk.blocks.4.norm2.weight FLOAT[768] 2f18b6c7d914
visual.trunk.blocks.5.attn.proj.bias FLOAT[768] c91ad71190de
visual.trunk.blocks.5.attn.qkv.bias FLOAT[2304] 8244e316061c
visual.trunk.blocks.5.mlp.fc1.bias FLOAT[3072] a48a6a380283
visual.trunk.blocks.5.mlp.fc2.bias FLOAT[768] 3e38a87eb086
visual.trunk.blocks.5.norm1.bias FLOAT[768] a2ee084997fa
visual.trunk.blocks.5.norm1.weight FLOAT[768] 59287c46aa60
visual.trunk.blocks.5.norm2.bias FLOAT[768] 4babc45cb3fc
visual.trunk.blocks.5.norm2.weight FLOAT[768] 212beef92293
visual.trunk.blocks.6.attn.proj.bias FLOAT[768] 896920c19ea9
visual.trunk.blocks.6.attn.qkv.bias FLOAT[2304] 92c34260cd0b
visual.trunk.blocks.6.mlp.fc1.bias FLOAT[3072] c76a297dd8a2
visual.trunk.blocks.6.mlp.fc2.bias FLOAT[768] ea8ee09f2755
visual.trunk.blocks.6.norm1.bias FLOAT[768] 2db5744a1bed
visual.trunk.blocks.6.norm1.weight FLOAT[768] 8b4b086907da
visual.trunk.blocks.6.norm2.bias FLOAT[768] b4a1ec75fa58
visual.trunk.blocks.6.norm2.weight FLOAT[768] 7ad41d612a60
visual.trunk.blocks.7.attn.proj.bias FLOAT[768] 801aa993b1fd
visual.trunk.blocks.7.attn.qkv.bias FLOAT[2304] 56149f5cebb8
visual.trunk.blocks.7.mlp.fc1.bias FLOAT[3072] ab0cc1552be0
visual.trunk.blocks.7.mlp.fc2.bias FLOAT[768] 1bf8a4aad1bc
visual.trunk.blocks.7.norm1.bias FLOAT[768] 250aa82d3076
visual.trunk.blocks.7.norm1.weight FLOAT[768] 3fc4c9edbbde
visual.trunk.blocks.7.norm2.bias FLOAT[768] 19305c0b6e66
visual.trunk.blocks.7.norm2.weight FLOAT[768] c477fdc172b6
visual.trunk.blocks.8.attn.proj.bias FLOAT[768] fc01102ca7e2
visual.trunk.blocks.8.attn.qkv.bias FLOAT[2304] d7b14a7a9c9d
visual.trunk.blocks.8.mlp.fc1.bias FLOAT[3072] 1b417b9a2220
visual.trunk.blocks.8.mlp.fc2.bias FLOAT[768] e18b0d0221b2
visual.trunk.blocks.8.norm1.bias FLOAT[768] d4592155ea6a
visual.trunk.blocks.8.norm1.weight FLOAT[768] c8daa76e003e
visual.trunk.blocks.8.norm2.bias FLOAT[768] 7e1cf2d8d005
visual.trunk.blocks.8.norm2.weight FLOAT[768] 15fef7298d14
visual.trunk.blocks.9.attn.proj.bias FLOAT[768] 5ca6136dedf3
visual.trunk.blocks.9.attn.qkv.bias FLOAT[2304] a0dd0ebd8544
visual.trunk.blocks.9.mlp.fc1.bias FLOAT[3072] b91a525b1353
visual.trunk.blocks.9.mlp.fc2.bias FLOAT[768] eec0491754ac
visual.trunk.blocks.9.norm1.bias FLOAT[768] e9d6b6770f4c
visual.trunk.blocks.9.norm1.weight FLOAT[768] 0728df499e34
visual.trunk.blocks.9.norm2.bias FLOAT[768] dd570703020f
visual.trunk.blocks.9.norm2.weight FLOAT[768] 1007bbaba35d
visual.trunk.norm.bias FLOAT[768] 0996d37acedf
visual.trunk.norm.weight FLOAT[768] e0220bdaf431
visual.trunk.patch_embed.proj.bias FLOAT[768] f242e94ef0bb
visual.trunk.patch_embed.proj.weight FLOAT[768,3,16,16] 6720fd78d7a5
visual.trunk.pos_embed FLOAT[1,576,768] 8f050e0de1a7
Reference in New Issue View Git Blame Copy Permalink