{ "format": "tensorviz.graph", "version": 3, "exportedAt": "2026-09-17T06:51:52.437778+00:00", "status": "static", "warningCount": 0, "document": { "schemaVersion": 2, "semantic": { "graph_id": "snapshot", "name": "TinyDiT", "revision": 0, "nodes": [ { "node_id": "input", "kind": "input", "op": "Input", "params": { "shape": "1,2,4,4" }, "in_ports": [], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "input", "name": "Input", "displayName": "Input", "description": "Model input node", "icon": "ArrowRight", "category": "io", "parameters": [ { "name": "shape", "type": "string", "defaultValue": "1,3,224,224", "required": true, "description": "Input tensor shape (comma-separated, e.g., 1,3,224,224)" }, { "name": "name", "type": "string", "defaultValue": "input", "required": false, "description": "Name for this input" } ] } } }, { "node_id": "patchify", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 2, "out_channels": 8, "kernel_size": 2, "stride": 2 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "conv2d", "name": "Conv2d", "displayName": "Conv2D", "description": "2D Convolutional Layer", "icon": "Grid3X3", "category": "conv", "parameters": [ { "name": "in_channels", "type": "number", "defaultValue": 3, "required": true, "description": "Number of input channels", "min": 1 }, { "name": "out_channels", "type": "number", "defaultValue": 64, "required": true, "description": "Number of output channels", "min": 1 }, { "name": "kernel_size", "type": "number", "defaultValue": 3, "required": true, "description": "Size of the convolving kernel", "min": 1 }, { "name": "stride", "type": "number", "defaultValue": 1, "required": false, "description": "Stride of the convolution", "min": 1 }, { "name": "padding", "type": "number", "defaultValue": 0, "required": false, "description": "Zero-padding added to both sides", "min": 0 }, { "name": "dilation", "type": "number", "defaultValue": 1, "required": false, "description": "Spacing between kernel elements", "min": 1 }, { "name": "groups", "type": "number", "defaultValue": 1, "required": false, "description": "Number of blocked connections from input to output channels", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, adds a learnable bias" } ] } } }, { "node_id": "flatten@TinyDiT#0", "kind": "layer", "op": "Flatten", "params": { "start_dim": 2 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "flatten", "name": "Flatten", "displayName": "Flatten", "description": "Flattens input tensor contiguously", "icon": "Maximize2", "category": "tensor", "parameters": [ { "name": "start_dim", "type": "number", "defaultValue": 1, "required": false, "description": "First dim to flatten", "min": 0 }, { "name": "end_dim", "type": "number", "defaultValue": -1, "required": false, "description": "Last dim to flatten" } ] } } }, { "node_id": "time_embedding", "kind": "module", "op": "Module", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "module", "name": "Module", "displayName": "Module", "description": "Custom module group (nn.Module)", "icon": "Package", "category": "module", "parameters": [ { "name": "name", "type": "string", "defaultValue": "CustomModule", "required": true, "description": "Module class name" }, { "name": "description", "type": "string", "defaultValue": "", "required": false, "description": "Module description" } ] }, "moduleInstanceId": "time_embedding", "moduleTemplateId": "model::TimestepEmbedding", "isGroup": true, "moduleMembers": [ "time_embedding.up", "mul@time_embedding#0", "python@time_embedding#0", "mul@time_embedding#1", "cat@time_embedding#0", "time_embedding.silu", "time_embedding.down" ], "moduleLayerCount": 7 } }, { "node_id": "class_embedding", "kind": "layer", "op": "Embedding", "params": { "num_embeddings": 4, "embedding_dim": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "embedding", "name": "Embedding", "displayName": "Embedding", "description": "Lookup table for dense vector embeddings", "icon": "Hash", "category": "embedding", "parameters": [ { "name": "num_embeddings", "type": "number", "defaultValue": 10000, "required": true, "description": "Size of the dictionary of embeddings", "min": 1 }, { "name": "embedding_dim", "type": "number", "defaultValue": 128, "required": true, "description": "Size of each embedding vector", "min": 1 }, { "name": "padding_idx", "type": "number", "defaultValue": 0, "required": false, "description": "Pad output with zeros at this index", "min": 0 }, { "name": "max_norm", "type": "number", "defaultValue": 0, "required": false, "description": "If > 0, re-normalizes embeddings with norm > max_norm", "min": 0 }, { "name": "sparse", "type": "boolean", "defaultValue": false, "required": false, "description": "If True, gradient w.r.t. weight is sparse" } ] } } }, { "node_id": "add@TinyDiT#0", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] } } }, { "node_id": "block", "kind": "module", "op": "Module", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "module", "name": "Module", "displayName": "Module", "description": "Custom module group (nn.Module)", "icon": "Package", "category": "module", "parameters": [ { "name": "name", "type": "string", "defaultValue": "CustomModule", "required": true, "description": "Module class name" }, { "name": "description", "type": "string", "defaultValue": "", "required": false, "description": "Module description" } ] }, "moduleInstanceId": "block", "moduleTemplateId": "model::AdaLNZeroBlock", "isGroup": true, "moduleMembers": [ "block.silu", "block.modulation", "python@block#0", "block.attention_norm", "mul@block#0", "add@block#0", "block.attention", "mul@block#1", "add@block#1", "block.ff_norm", "mul@block#2", "add@block#2", "block.up", "block.gelu", "block.down", "mul@block#3", "add@block#3" ], "moduleLayerCount": 25 } }, { "node_id": "decoder", "kind": "module", "op": "Module", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "module", "name": "Module", "displayName": "Module", "description": "Custom module group (nn.Module)", "icon": "Package", "category": "module", "parameters": [ { "name": "name", "type": "string", "defaultValue": "CustomModule", "required": true, "description": "Module class name" }, { "name": "description", "type": "string", "defaultValue": "", "required": false, "description": "Module description" } ] }, "moduleInstanceId": "decoder", "moduleTemplateId": "model::PatchDecoder", "isGroup": true, "moduleMembers": [ "silu@decoder#0", "decoder.norm", "decoder.modulation", "python@decoder#0", "mul@decoder#0", "add@decoder#0", "decoder.project", "python@decoder#1" ], "moduleLayerCount": 8 } }, { "node_id": "output", "kind": "output", "op": "Output", "params": {}, "in_ports": [ "x" ], "out_ports": [], "metadata": { "layerType": { "id": "output", "name": "Output", "displayName": "Output", "description": "Model output node", "icon": "ArrowLeft", "category": "io", "parameters": [ { "name": "name", "type": "string", "defaultValue": "output", "required": false, "description": "Name for this output" } ] } } }, { "node_id": "time_embedding.up", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "time_embedding" } }, { "node_id": "mul@time_embedding#0", "kind": "layer", "op": "torch.mul", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "mul", "name": "torch.mul", "displayName": "torch.mul", "description": "PyTorch mul operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "time_embedding" } }, { "node_id": "python@time_embedding#0", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "time_embedding" } }, { "node_id": "mul@time_embedding#1", "kind": "layer", "op": "torch.mul", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "mul", "name": "torch.mul", "displayName": "torch.mul", "description": "PyTorch mul operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "time_embedding" } }, { "node_id": "cat@time_embedding#0", "kind": "layer", "op": "torch.cat", "params": { "dim": -1 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "cat", "name": "torch.cat", "displayName": "torch.cat", "description": "PyTorch cat operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "time_embedding" } }, { "node_id": "time_embedding.silu", "kind": "layer", "op": "SiLU", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "silu", "name": "SiLU", "displayName": "SiLU", "description": "Sigmoid Linear Unit (Swish)", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "inplace", "type": "boolean", "defaultValue": false, "required": false, "description": "Can optionally do the operation in-place" } ] }, "parentId": "time_embedding" } }, { "node_id": "time_embedding.down", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "time_embedding" } }, { "node_id": "block.silu", "kind": "layer", "op": "SiLU", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "silu", "name": "SiLU", "displayName": "SiLU", "description": "Sigmoid Linear Unit (Swish)", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "inplace", "type": "boolean", "defaultValue": false, "required": false, "description": "Can optionally do the operation in-place" } ] }, "parentId": "block" } }, { "node_id": "block.modulation", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 48 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "block" } }, { "node_id": "python@block#0", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "block" } }, { "node_id": "block.attention_norm", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 8, "elementwise_affine": false }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "layernorm", "name": "LayerNorm", "displayName": "LayerNorm", "description": "Layer Normalization", "icon": "Layers", "category": "norm", "parameters": [ { "name": "normalized_shape", "type": "string", "defaultValue": "512", "required": true, "description": "Input shape for normalization (comma-separated for multi-dim, e.g., 512)" }, { "name": "eps", "type": "number", "defaultValue": 0.00001, "required": false, "description": "Value for numerical stability", "min": 0 }, { "name": "elementwise_affine", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, learnable affine parameters" } ] }, "parentId": "block" } }, { "node_id": "mul@block#0", "kind": "layer", "op": "torch.mul", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "mul", "name": "torch.mul", "displayName": "torch.mul", "description": "PyTorch mul operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "block" } }, { "node_id": "add@block#0", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "block" } }, { "node_id": "block.attention", "kind": "module", "op": "Module", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "module", "name": "Module", "displayName": "Module", "description": "Custom module group (nn.Module)", "icon": "Package", "category": "module", "parameters": [ { "name": "name", "type": "string", "defaultValue": "CustomModule", "required": true, "description": "Module class name" }, { "name": "description", "type": "string", "defaultValue": "", "required": false, "description": "Module description" } ] }, "parentId": "block", "moduleInstanceId": "block.attention", "moduleTemplateId": "model::SelfAttention", "isGroup": true, "moduleMembers": [ "block.attention.query", "python@block.attention#0", "block.attention.key", "python@block.attention#1", "div@block.attention#0", "block.attention.value", "python@block.attention#2", "block.attention.softmax", "block.attention.project" ], "moduleLayerCount": 9 } }, { "node_id": "mul@block#1", "kind": "layer", "op": "torch.mul", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "mul", "name": "torch.mul", "displayName": "torch.mul", "description": "PyTorch mul operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "block" } }, { "node_id": "add@block#1", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "block" } }, { "node_id": "block.ff_norm", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 8, "elementwise_affine": false }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "layernorm", "name": "LayerNorm", "displayName": "LayerNorm", "description": "Layer Normalization", "icon": "Layers", "category": "norm", "parameters": [ { "name": "normalized_shape", "type": "string", "defaultValue": "512", "required": true, "description": "Input shape for normalization (comma-separated for multi-dim, e.g., 512)" }, { "name": "eps", "type": "number", "defaultValue": 0.00001, "required": false, "description": "Value for numerical stability", "min": 0 }, { "name": "elementwise_affine", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, learnable affine parameters" } ] }, "parentId": "block" } }, { "node_id": "mul@block#2", "kind": "layer", "op": "torch.mul", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "mul", "name": "torch.mul", "displayName": "torch.mul", "description": "PyTorch mul operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "block" } }, { "node_id": "add@block#2", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "block" } }, { "node_id": "block.up", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 32 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "block" } }, { "node_id": "block.gelu", "kind": "layer", "op": "GELU", "params": { "approximate": "tanh" }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "gelu", "name": "GELU", "displayName": "GELU", "description": "Gaussian Error Linear Unit", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "approximate", "type": "select", "defaultValue": "none", "required": false, "description": "The gelu approximation algorithm to use: 'none' or 'tanh'", "options": [ "none", "tanh" ] } ] }, "parentId": "block" } }, { "node_id": "block.down", "kind": "layer", "op": "Linear", "params": { "in_features": 32, "out_features": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "block" } }, { "node_id": "mul@block#3", "kind": "layer", "op": "torch.mul", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "mul", "name": "torch.mul", "displayName": "torch.mul", "description": "PyTorch mul operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "block" } }, { "node_id": "add@block#3", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "block" } }, { "node_id": "block.attention.query", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "block.attention" } }, { "node_id": "python@block.attention#0", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "block.attention" } }, { "node_id": "block.attention.key", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "block.attention" } }, { "node_id": "python@block.attention#1", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "block.attention" } }, { "node_id": "div@block.attention#0", "kind": "layer", "op": "torch.div", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "div", "name": "torch.div", "displayName": "torch.div", "description": "PyTorch div operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "block.attention" } }, { "node_id": "block.attention.value", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "block.attention" } }, { "node_id": "python@block.attention#2", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "block.attention" } }, { "node_id": "block.attention.softmax", "kind": "layer", "op": "Softmax", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "softmax", "name": "Softmax", "displayName": "Softmax", "description": "Softmax over a dimension", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "dim", "type": "number", "defaultValue": 1, "required": true, "description": "Dimension along which Softmax is computed" } ] }, "parentId": "block.attention" } }, { "node_id": "block.attention.project", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "block.attention" } }, { "node_id": "silu@decoder#0", "kind": "layer", "op": "SiLU", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "silu", "name": "SiLU", "displayName": "SiLU", "description": "Sigmoid Linear Unit (Swish)", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "inplace", "type": "boolean", "defaultValue": false, "required": false, "description": "Can optionally do the operation in-place" } ] }, "parentId": "decoder" } }, { "node_id": "decoder.norm", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 8, "elementwise_affine": false }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "layernorm", "name": "LayerNorm", "displayName": "LayerNorm", "description": "Layer Normalization", "icon": "Layers", "category": "norm", "parameters": [ { "name": "normalized_shape", "type": "string", "defaultValue": "512", "required": true, "description": "Input shape for normalization (comma-separated for multi-dim, e.g., 512)" }, { "name": "eps", "type": "number", "defaultValue": 0.00001, "required": false, "description": "Value for numerical stability", "min": 0 }, { "name": "elementwise_affine", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, learnable affine parameters" } ] }, "parentId": "decoder" } }, { "node_id": "decoder.modulation", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 16 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "decoder" } }, { "node_id": "python@decoder#0", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "decoder" } }, { "node_id": "mul@decoder#0", "kind": "layer", "op": "torch.mul", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "mul", "name": "torch.mul", "displayName": "torch.mul", "description": "PyTorch mul operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "decoder" } }, { "node_id": "add@decoder#0", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "decoder" } }, { "node_id": "decoder.project", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 16 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "decoder" } }, { "node_id": "python@decoder#1", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "decoder" } } ], "edges": [ { "edge_id": "edge-1", "source_node_id": "time_embedding.up", "target_node_id": "mul@time_embedding#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-2", "source_node_id": "mul@time_embedding#0", "target_node_id": "python@time_embedding#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-3", "source_node_id": "python@time_embedding#0", "target_node_id": "mul@time_embedding#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-4", "source_node_id": "mul@time_embedding#1", "target_node_id": "cat@time_embedding#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-5", "source_node_id": "cat@time_embedding#0", "target_node_id": "time_embedding.up", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-6", "source_node_id": "time_embedding.up", "target_node_id": "time_embedding.silu", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-7", "source_node_id": "time_embedding.silu", "target_node_id": "time_embedding.down", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-8", "source_node_id": "block.attention.query", "target_node_id": "python@block.attention#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-9", "source_node_id": "block.attention.key", "target_node_id": "python@block.attention#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-10", "source_node_id": "block.attention.value", "target_node_id": "python@block.attention#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-11", "source_node_id": "python@block.attention#0", "target_node_id": "div@block.attention#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-12", "source_node_id": "python@block.attention#1", "target_node_id": "div@block.attention#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-13", "source_node_id": "div@block.attention#0", "target_node_id": "block.attention.softmax", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-14", "source_node_id": "block.attention.softmax", "target_node_id": "block.attention.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-15", "source_node_id": "python@block.attention#2", "target_node_id": "block.attention.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-16", "source_node_id": "block.silu", "target_node_id": "block.modulation", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-17", "source_node_id": "block.modulation", "target_node_id": "python@block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-18", "source_node_id": "block.attention_norm", "target_node_id": "mul@block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-19", "source_node_id": "python@block#0", "target_node_id": "mul@block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-20", "source_node_id": "mul@block#0", "target_node_id": "add@block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-21", "source_node_id": "python@block#0", "target_node_id": "add@block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-22", "source_node_id": "add@block#0", "target_node_id": "block.attention.query", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-23", "source_node_id": "add@block#0", "target_node_id": "python@block.attention#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-24", "source_node_id": "add@block#0", "target_node_id": "block.attention.key", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-25", "source_node_id": "add@block#0", "target_node_id": "python@block.attention#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-26", "source_node_id": "add@block#0", "target_node_id": "block.attention.value", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-27", "source_node_id": "add@block#0", "target_node_id": "python@block.attention#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-28", "source_node_id": "add@block#0", "target_node_id": "block.attention.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-29", "source_node_id": "add@block#0", "target_node_id": "block.attention", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-30", "source_node_id": "python@block#0", "target_node_id": "mul@block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-31", "source_node_id": "block.attention.project", "target_node_id": "mul@block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-32", "source_node_id": "block.attention", "target_node_id": "mul@block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-33", "source_node_id": "mul@block#1", "target_node_id": "add@block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-34", "source_node_id": "add@block#1", "target_node_id": "block.ff_norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-35", "source_node_id": "block.ff_norm", "target_node_id": "mul@block#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-36", "source_node_id": "python@block#0", "target_node_id": "mul@block#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-37", "source_node_id": "mul@block#2", "target_node_id": "add@block#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-38", "source_node_id": "python@block#0", "target_node_id": "add@block#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-39", "source_node_id": "add@block#2", "target_node_id": "block.up", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-40", "source_node_id": "block.up", "target_node_id": "block.gelu", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-41", "source_node_id": "block.gelu", "target_node_id": "block.down", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-42", "source_node_id": "python@block#0", "target_node_id": "mul@block#3", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-43", "source_node_id": "block.down", "target_node_id": "mul@block#3", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-44", "source_node_id": "add@block#1", "target_node_id": "add@block#3", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-45", "source_node_id": "mul@block#3", "target_node_id": "add@block#3", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-46", "source_node_id": "silu@decoder#0", "target_node_id": "decoder.modulation", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-47", "source_node_id": "decoder.modulation", "target_node_id": "python@decoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-48", "source_node_id": "decoder.norm", "target_node_id": "mul@decoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-49", "source_node_id": "python@decoder#0", "target_node_id": "mul@decoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-50", "source_node_id": "mul@decoder#0", "target_node_id": "add@decoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-51", "source_node_id": "python@decoder#0", "target_node_id": "add@decoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-52", "source_node_id": "add@decoder#0", "target_node_id": "decoder.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-53", "source_node_id": "decoder.project", "target_node_id": "python@decoder#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-54", "source_node_id": "input", "target_node_id": "patchify", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-55", "source_node_id": "patchify", "target_node_id": "flatten@TinyDiT#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-56", "source_node_id": "input", "target_node_id": "mul@time_embedding#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-57", "source_node_id": "input", "target_node_id": "mul@time_embedding#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-58", "source_node_id": "input", "target_node_id": "time_embedding", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-59", "source_node_id": "input", "target_node_id": "class_embedding", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-60", "source_node_id": "time_embedding.down", "target_node_id": "add@TinyDiT#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-61", "source_node_id": "time_embedding", "target_node_id": "add@TinyDiT#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-62", "source_node_id": "class_embedding", "target_node_id": "add@TinyDiT#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-63", "source_node_id": "flatten@TinyDiT#0", "target_node_id": "block.silu", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-64", "source_node_id": "flatten@TinyDiT#0", "target_node_id": "block.attention_norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-65", "source_node_id": "flatten@TinyDiT#0", "target_node_id": "add@block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-66", "source_node_id": "flatten@TinyDiT#0", "target_node_id": "block", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-67", "source_node_id": "add@TinyDiT#0", "target_node_id": "block.silu", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-68", "source_node_id": "add@TinyDiT#0", "target_node_id": "block.attention_norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-69", "source_node_id": "add@TinyDiT#0", "target_node_id": "add@block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-70", "source_node_id": "add@TinyDiT#0", "target_node_id": "block", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-71", "source_node_id": "add@block#3", "target_node_id": "silu@decoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-72", "source_node_id": "add@block#3", "target_node_id": "decoder.norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-73", "source_node_id": "block", "target_node_id": "decoder", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-74", "source_node_id": "add@TinyDiT#0", "target_node_id": "silu@decoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-75", "source_node_id": "add@TinyDiT#0", "target_node_id": "decoder.norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-76", "source_node_id": "add@TinyDiT#0", "target_node_id": "decoder", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-77", "source_node_id": "python@decoder#1", "target_node_id": "output", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-78", "source_node_id": "decoder", "target_node_id": "output", "source_port": "y", "target_port": "x", "metadata": {} } ], "inputs": [], "modules": [], "operations": [], "metadata": {} }, "view": { "graph_id": "snapshot", "node_positions": { "input": { "x": 0, "y": 160 }, "patchify": { "x": 340, "y": 160 }, "flatten@TinyDiT#0": { "x": 680, "y": 160 }, "time_embedding": { "x": 1020, "y": 160 }, "class_embedding": { "x": 1360, "y": 160 }, "add@TinyDiT#0": { "x": 1700, "y": 160 }, "block": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "time_embedding.up": { "x": 24, "y": 214 }, "mul@time_embedding#0": { "x": 24, "y": 364 }, "python@time_embedding#0": { "x": 24, "y": 514 }, "mul@time_embedding#1": { "x": 24, "y": 664 }, "cat@time_embedding#0": { "x": 24, "y": 814 }, "time_embedding.silu": { "x": 24, "y": 964 }, "time_embedding.down": { "x": 24, "y": 1114 }, "block.silu": { "x": 24, "y": 214 }, "block.modulation": { "x": 24, "y": 364 }, "python@block#0": { "x": 24, "y": 514 }, "block.attention_norm": { "x": 24, "y": 664 }, "mul@block#0": { "x": 24, "y": 814 }, "add@block#0": { "x": 24, "y": 964 }, "block.attention": { "x": 24, "y": 1114 }, "mul@block#1": { "x": 24, "y": 1264 }, "add@block#1": { "x": 24, "y": 1414 }, "block.ff_norm": { "x": 24, "y": 1564 }, "mul@block#2": { "x": 24, "y": 1714 }, "add@block#2": { "x": 24, "y": 1864 }, "block.up": { "x": 24, "y": 2014 }, "block.gelu": { "x": 24, "y": 2164 }, "block.down": { "x": 24, "y": 2314 }, "mul@block#3": { "x": 24, "y": 2464 }, "add@block#3": { "x": 24, "y": 2614 }, "block.attention.query": { "x": 24, "y": 214 }, "python@block.attention#0": { "x": 24, "y": 364 }, "block.attention.key": { "x": 24, "y": 514 }, "python@block.attention#1": { "x": 24, "y": 664 }, "div@block.attention#0": { "x": 24, "y": 814 }, "block.attention.value": { "x": 24, "y": 964 }, "python@block.attention#2": { "x": 24, "y": 1114 }, "block.attention.softmax": { "x": 24, "y": 1264 }, "block.attention.project": { "x": 24, "y": 1414 }, "silu@decoder#0": { "x": 24, "y": 214 }, "decoder.norm": { "x": 24, "y": 364 }, "decoder.modulation": { "x": 24, "y": 514 }, "python@decoder#0": { "x": 24, "y": 664 }, "mul@decoder#0": { "x": 24, "y": 814 }, "add@decoder#0": { "x": 24, "y": 964 }, "decoder.project": { "x": 24, "y": 1114 }, "python@decoder#1": { "x": 24, "y": 1264 } }, "collapsed_modules": [ "time_embedding", "block", "decoder", "block.attention" ], "selection": [], "layout": {} } }, "snapshotId": "snapshot-111a8b31c7bf143f", "collaboration": { "walkthrough": { "title": "Use a Transformer as the diffusion backbone", "steps": [ { "id": "patches", "title": "Turn the noisy latent into four tokens", "note": "A 2 × 2 stride-two convolution projects latent patches into eight features. Fixed 2D sinusoidal positions preserve the row and column coordinates. Unlike ViT classification, this backbone has no class token.", "view": { "camera": { "x": -7.5, "y": 46.5, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "patchify": { "x": 340, "y": 160 }, "flatten@TinyDiT#0": { "x": 680, "y": 160 }, "time_embedding": { "x": 1020, "y": 160 }, "class_embedding": { "x": 1360, "y": 160 }, "add@TinyDiT#0": { "x": 1700, "y": 160 }, "block": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "time_embedding.up": { "x": 24, "y": 214 }, "mul@time_embedding#0": { "x": 24, "y": 364 }, "python@time_embedding#0": { "x": 24, "y": 514 }, "mul@time_embedding#1": { "x": 24, "y": 664 }, "cat@time_embedding#0": { "x": 24, "y": 814 }, "time_embedding.silu": { "x": 24, "y": 964 }, "time_embedding.down": { "x": 24, "y": 1114 }, "block.silu": { "x": 24, "y": 214 }, "block.modulation": { "x": 24, "y": 364 }, "python@block#0": { "x": 24, "y": 514 }, "block.attention_norm": { "x": 24, "y": 664 }, "mul@block#0": { "x": 24, "y": 814 }, "add@block#0": { "x": 24, "y": 964 }, "block.attention": { "x": 24, "y": 1114 }, "mul@block#1": { "x": 24, "y": 1264 }, "add@block#1": { "x": 24, "y": 1414 }, "block.ff_norm": { "x": 24, "y": 1564 }, "mul@block#2": { "x": 24, "y": 1714 }, "add@block#2": { "x": 24, "y": 1864 }, "block.up": { "x": 24, "y": 2014 }, "block.gelu": { "x": 24, "y": 2164 }, "block.down": { "x": 24, "y": 2314 }, "mul@block#3": { "x": 24, "y": 2464 }, "add@block#3": { "x": 24, "y": 2614 }, "block.attention.query": { "x": 24, "y": 214 }, "python@block.attention#0": { "x": 24, "y": 364 }, "block.attention.key": { "x": 24, "y": 514 }, "python@block.attention#1": { "x": 24, "y": 664 }, "div@block.attention#0": { "x": 24, "y": 814 }, "block.attention.value": { "x": 24, "y": 964 }, "python@block.attention#2": { "x": 24, "y": 1114 }, "block.attention.softmax": { "x": 24, "y": 1264 }, "block.attention.project": { "x": 24, "y": 1414 }, "silu@decoder#0": { "x": 24, "y": 214 }, "decoder.norm": { "x": 24, "y": 364 }, "decoder.modulation": { "x": 24, "y": 514 }, "python@decoder#0": { "x": 24, "y": 664 }, "mul@decoder#0": { "x": 24, "y": 814 }, "add@decoder#0": { "x": 24, "y": 964 }, "decoder.project": { "x": 24, "y": 1114 }, "python@decoder#1": { "x": 24, "y": 1264 } }, "expandedModuleIds": [], "selectedNodeIds": [ "patchify" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "patchify" }, "snapshotId": "snapshot-111a8b31c7bf143f", "exportedAt": "2026-09-17T06:51:52.437778+00:00", "modelName": "TinyDiT", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"patchify\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"patchify\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"flatten@TinyDiT#0\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"patchify\",\"kind\":\"layer\",\"op\":\"Conv2d\",\"params\":{\"in_channels\":2,\"kernel_size\":2,\"out_channels\":8,\"stride\":2}}}" } ] }, { "id": "condition", "title": "Combine diffusion time with the class label", "note": "The timestep passes through sinusoidal features and an MLP. A learned class embedding is added to it. This shared conditioning vector controls every spatial token, without becoming an image patch.", "view": { "camera": { "x": -1027.5, "y": 60, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "patchify": { "x": 340, "y": 160 }, "flatten@TinyDiT#0": { "x": 680, "y": 160 }, "time_embedding": { "x": 1020, "y": 160 }, "class_embedding": { "x": 1360, "y": 160 }, "add@TinyDiT#0": { "x": 1700, "y": 160 }, "block": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "time_embedding.up": { "x": 24, "y": 214 }, "mul@time_embedding#0": { "x": 24, "y": 364 }, "python@time_embedding#0": { "x": 24, "y": 514 }, "mul@time_embedding#1": { "x": 24, "y": 664 }, "cat@time_embedding#0": { "x": 24, "y": 814 }, "time_embedding.silu": { "x": 24, "y": 964 }, "time_embedding.down": { "x": 24, "y": 1114 }, "block.silu": { "x": 24, "y": 214 }, "block.modulation": { "x": 24, "y": 364 }, "python@block#0": { "x": 24, "y": 514 }, "block.attention_norm": { "x": 24, "y": 664 }, "mul@block#0": { "x": 24, "y": 814 }, "add@block#0": { "x": 24, "y": 964 }, "block.attention": { "x": 24, "y": 1114 }, "mul@block#1": { "x": 24, "y": 1264 }, "add@block#1": { "x": 24, "y": 1414 }, "block.ff_norm": { "x": 24, "y": 1564 }, "mul@block#2": { "x": 24, "y": 1714 }, "add@block#2": { "x": 24, "y": 1864 }, "block.up": { "x": 24, "y": 2014 }, "block.gelu": { "x": 24, "y": 2164 }, "block.down": { "x": 24, "y": 2314 }, "mul@block#3": { "x": 24, "y": 2464 }, "add@block#3": { "x": 24, "y": 2614 }, "block.attention.query": { "x": 24, "y": 214 }, "python@block.attention#0": { "x": 24, "y": 364 }, "block.attention.key": { "x": 24, "y": 514 }, "python@block.attention#1": { "x": 24, "y": 664 }, "div@block.attention#0": { "x": 24, "y": 814 }, "block.attention.value": { "x": 24, "y": 964 }, "python@block.attention#2": { "x": 24, "y": 1114 }, "block.attention.softmax": { "x": 24, "y": 1264 }, "block.attention.project": { "x": 24, "y": 1414 }, "silu@decoder#0": { "x": 24, "y": 214 }, "decoder.norm": { "x": 24, "y": 364 }, "decoder.modulation": { "x": 24, "y": 514 }, "python@decoder#0": { "x": 24, "y": 664 }, "mul@decoder#0": { "x": 24, "y": 814 }, "add@decoder#0": { "x": 24, "y": 964 }, "decoder.project": { "x": 24, "y": 1114 }, "python@decoder#1": { "x": 24, "y": 1264 } }, "expandedModuleIds": [], "selectedNodeIds": [ "class_embedding" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "class_embedding" }, "snapshotId": "snapshot-111a8b31c7bf143f", "exportedAt": "2026-09-17T06:51:52.437778+00:00", "modelName": "TinyDiT", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"class_embedding\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"add@TinyDiT#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"class_embedding\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"class_embedding\",\"kind\":\"layer\",\"op\":\"Embedding\",\"params\":{\"embedding_dim\":8,\"num_embeddings\":4}}}" } ] }, { "id": "gates", "title": "Start each residual block as the identity", "note": "Conditioning produces six vectors: attention and feed-forward shifts, scales and residual gates. Zero initialization closes both residual branches, so the block starts as identity. A recorded nonzero intervention makes the effect of class and time visible.", "view": { "camera": { "x": 308.5, "y": -144, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "patchify": { "x": 340, "y": 160 }, "flatten@TinyDiT#0": { "x": 680, "y": 160 }, "time_embedding": { "x": 1020, "y": 160 }, "class_embedding": { "x": 1360, "y": 160 }, "add@TinyDiT#0": { "x": 1700, "y": 160 }, "block": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "time_embedding.up": { "x": 24, "y": 214 }, "mul@time_embedding#0": { "x": 24, "y": 364 }, "python@time_embedding#0": { "x": 24, "y": 514 }, "mul@time_embedding#1": { "x": 24, "y": 664 }, "cat@time_embedding#0": { "x": 24, "y": 814 }, "time_embedding.silu": { "x": 24, "y": 964 }, "time_embedding.down": { "x": 24, "y": 1114 }, "block.silu": { "x": 24, "y": 214 }, "block.modulation": { "x": 24, "y": 364 }, "python@block#0": { "x": 24, "y": 514 }, "block.attention_norm": { "x": 24, "y": 664 }, "mul@block#0": { "x": 24, "y": 814 }, "add@block#0": { "x": 24, "y": 964 }, "block.attention": { "x": 24, "y": 1114 }, "mul@block#1": { "x": 24, "y": 1264 }, "add@block#1": { "x": 24, "y": 1414 }, "block.ff_norm": { "x": 24, "y": 1564 }, "mul@block#2": { "x": 24, "y": 1714 }, "add@block#2": { "x": 24, "y": 1864 }, "block.up": { "x": 24, "y": 2014 }, "block.gelu": { "x": 24, "y": 2164 }, "block.down": { "x": 24, "y": 2314 }, "mul@block#3": { "x": 24, "y": 2464 }, "add@block#3": { "x": 24, "y": 2614 }, "block.attention.query": { "x": 24, "y": 214 }, "python@block.attention#0": { "x": 24, "y": 364 }, "block.attention.key": { "x": 24, "y": 514 }, "python@block.attention#1": { "x": 24, "y": 664 }, "div@block.attention#0": { "x": 24, "y": 814 }, "block.attention.value": { "x": 24, "y": 964 }, "python@block.attention#2": { "x": 24, "y": 1114 }, "block.attention.softmax": { "x": 24, "y": 1264 }, "block.attention.project": { "x": 24, "y": 1414 }, "silu@decoder#0": { "x": 24, "y": 214 }, "decoder.norm": { "x": 24, "y": 364 }, "decoder.modulation": { "x": 24, "y": 514 }, "python@decoder#0": { "x": 24, "y": 664 }, "mul@decoder#0": { "x": 24, "y": 814 }, "add@decoder#0": { "x": 24, "y": 964 }, "decoder.project": { "x": 24, "y": 1114 }, "python@decoder#1": { "x": 24, "y": 1264 } }, "expandedModuleIds": [ "block" ], "selectedNodeIds": [ "block.modulation" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "block.modulation" }, "snapshotId": "snapshot-111a8b31c7bf143f", "exportedAt": "2026-09-17T06:51:52.437778+00:00", "modelName": "TinyDiT", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"block.modulation\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"python@block#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"block.silu\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"block.modulation\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"block.modulation\",\"kind\":\"layer\",\"op\":\"Linear\",\"params\":{\"in_features\":8,\"out_features\":48},\"parent\":\"block\"}}" } ] }, { "id": "spatial", "title": "Return tokens to noise and variance-parameter maps", "note": "Adaptive normalization and a linear projection emit each patch’s spatial values. Unpatchification restores the 4 × 4 grid with two epsilon and two raw variance-parameter channels. This final projection is also zero-initialized.", "view": { "camera": { "x": 308.5, "y": -894, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "patchify": { "x": 340, "y": 160 }, "flatten@TinyDiT#0": { "x": 680, "y": 160 }, "time_embedding": { "x": 1020, "y": 160 }, "class_embedding": { "x": 1360, "y": 160 }, "add@TinyDiT#0": { "x": 1700, "y": 160 }, "block": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "time_embedding.up": { "x": 24, "y": 214 }, "mul@time_embedding#0": { "x": 24, "y": 364 }, "python@time_embedding#0": { "x": 24, "y": 514 }, "mul@time_embedding#1": { "x": 24, "y": 664 }, "cat@time_embedding#0": { "x": 24, "y": 814 }, "time_embedding.silu": { "x": 24, "y": 964 }, "time_embedding.down": { "x": 24, "y": 1114 }, "block.silu": { "x": 24, "y": 214 }, "block.modulation": { "x": 24, "y": 364 }, "python@block#0": { "x": 24, "y": 514 }, "block.attention_norm": { "x": 24, "y": 664 }, "mul@block#0": { "x": 24, "y": 814 }, "add@block#0": { "x": 24, "y": 964 }, "block.attention": { "x": 24, "y": 1114 }, "mul@block#1": { "x": 24, "y": 1264 }, "add@block#1": { "x": 24, "y": 1414 }, "block.ff_norm": { "x": 24, "y": 1564 }, "mul@block#2": { "x": 24, "y": 1714 }, "add@block#2": { "x": 24, "y": 1864 }, "block.up": { "x": 24, "y": 2014 }, "block.gelu": { "x": 24, "y": 2164 }, "block.down": { "x": 24, "y": 2314 }, "mul@block#3": { "x": 24, "y": 2464 }, "add@block#3": { "x": 24, "y": 2614 }, "block.attention.query": { "x": 24, "y": 214 }, "python@block.attention#0": { "x": 24, "y": 364 }, "block.attention.key": { "x": 24, "y": 514 }, "python@block.attention#1": { "x": 24, "y": 664 }, "div@block.attention#0": { "x": 24, "y": 814 }, "block.attention.value": { "x": 24, "y": 964 }, "python@block.attention#2": { "x": 24, "y": 1114 }, "block.attention.softmax": { "x": 24, "y": 1264 }, "block.attention.project": { "x": 24, "y": 1414 }, "silu@decoder#0": { "x": 24, "y": 214 }, "decoder.norm": { "x": 24, "y": 364 }, "decoder.modulation": { "x": 24, "y": 514 }, "python@decoder#0": { "x": 24, "y": 664 }, "mul@decoder#0": { "x": 24, "y": 814 }, "add@decoder#0": { "x": 24, "y": 964 }, "decoder.project": { "x": 24, "y": 1114 }, "python@decoder#1": { "x": 24, "y": 1264 } }, "expandedModuleIds": [ "decoder" ], "selectedNodeIds": [ "decoder.project" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "decoder.project" }, "snapshotId": "snapshot-111a8b31c7bf143f", "exportedAt": "2026-09-17T06:51:52.437778+00:00", "modelName": "TinyDiT", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"add@decoder#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"decoder.project\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"decoder.project\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"python@decoder#1\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"decoder.project\",\"kind\":\"layer\",\"op\":\"Linear\",\"params\":{\"in_features\":8,\"out_features\":16},\"parent\":\"decoder\"}}" } ] } ] }, "threads": [], "savedViews": [ { "id": "gates", "title": "Start each residual block as the identity", "note": "Conditioning produces six vectors: attention and feed-forward shifts, scales and residual gates. Zero initialization closes both residual branches, so the block starts as identity. A recorded nonzero intervention makes the effect of class and time visible.", "view": { "camera": { "x": 308.5, "y": -144, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "patchify": { "x": 340, "y": 160 }, "flatten@TinyDiT#0": { "x": 680, "y": 160 }, "time_embedding": { "x": 1020, "y": 160 }, "class_embedding": { "x": 1360, "y": 160 }, "add@TinyDiT#0": { "x": 1700, "y": 160 }, "block": { "x": 2040, "y": 160 }, "decoder": { "x": 2380, "y": 160 }, "output": { "x": 2720, "y": 160 }, "time_embedding.up": { "x": 24, "y": 214 }, "mul@time_embedding#0": { "x": 24, "y": 364 }, "python@time_embedding#0": { "x": 24, "y": 514 }, "mul@time_embedding#1": { "x": 24, "y": 664 }, "cat@time_embedding#0": { "x": 24, "y": 814 }, "time_embedding.silu": { "x": 24, "y": 964 }, "time_embedding.down": { "x": 24, "y": 1114 }, "block.silu": { "x": 24, "y": 214 }, "block.modulation": { "x": 24, "y": 364 }, "python@block#0": { "x": 24, "y": 514 }, "block.attention_norm": { "x": 24, "y": 664 }, "mul@block#0": { "x": 24, "y": 814 }, "add@block#0": { "x": 24, "y": 964 }, "block.attention": { "x": 24, "y": 1114 }, "mul@block#1": { "x": 24, "y": 1264 }, "add@block#1": { "x": 24, "y": 1414 }, "block.ff_norm": { "x": 24, "y": 1564 }, "mul@block#2": { "x": 24, "y": 1714 }, "add@block#2": { "x": 24, "y": 1864 }, "block.up": { "x": 24, "y": 2014 }, "block.gelu": { "x": 24, "y": 2164 }, "block.down": { "x": 24, "y": 2314 }, "mul@block#3": { "x": 24, "y": 2464 }, "add@block#3": { "x": 24, "y": 2614 }, "block.attention.query": { "x": 24, "y": 214 }, "python@block.attention#0": { "x": 24, "y": 364 }, "block.attention.key": { "x": 24, "y": 514 }, "python@block.attention#1": { "x": 24, "y": 664 }, "div@block.attention#0": { "x": 24, "y": 814 }, "block.attention.value": { "x": 24, "y": 964 }, "python@block.attention#2": { "x": 24, "y": 1114 }, "block.attention.softmax": { "x": 24, "y": 1264 }, "block.attention.project": { "x": 24, "y": 1414 }, "silu@decoder#0": { "x": 24, "y": 214 }, "decoder.norm": { "x": 24, "y": 364 }, "decoder.modulation": { "x": 24, "y": 514 }, "python@decoder#0": { "x": 24, "y": 664 }, "mul@decoder#0": { "x": 24, "y": 814 }, "add@decoder#0": { "x": 24, "y": 964 }, "decoder.project": { "x": 24, "y": 1114 }, "python@decoder#1": { "x": 24, "y": 1264 } }, "expandedModuleIds": [ "block" ], "selectedNodeIds": [ "block.modulation" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "block.modulation" }, "snapshotId": "snapshot-111a8b31c7bf143f", "exportedAt": "2026-09-17T06:51:52.437778+00:00", "modelName": "TinyDiT", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"block.modulation\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"python@block#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"block.silu\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"block.modulation\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"block.modulation\",\"kind\":\"layer\",\"op\":\"Linear\",\"params\":{\"in_features\":8,\"out_features\":48},\"parent\":\"block\"}}" } ] } ], "openingViewId": "gates" } }