Edge naming and Handles Rework

[ManjotSingh08x]

Description

The current edge system does not track naming consistently across multiple edges originating from same handle. Furthermore, complex modules require handles with independent shape_compute functions. This issue covers a revision of edge/handle architecture which makes shape propagation more flexible and generalized.

Tasks

• Implement consistent edge naming:
  • All edges that originate from the same handle should share a linked name or naming group
  • Edge metadata should indicate: origin handle, target handle, origin node, target node.
• Create a Handle class:
  • Accepts and optional handle_shape_compute function.
  • Falls back to class default shape_compute when none is provided
  • Alternative implementation could include a HandleSchema of the class which describes the construction of Handles by the CreateLayerComponent factory.
• Extend shape propagation:
  • Multi-output modules should have multiple handles with independent handle_shape_compute functions.
  • shape computation and verification systems should be adapted to reflect this change.

Expected Outcomes

• Edges visually and internally show consistent names for handles.
• for a group of edges originating from the same handle, changing name of one edge, changes the names of other edges instantly.
• if a multi-output module cannot be found in PyTorch documentation, create a dummy module which simulates this behavior by combining functionalities of 2 other layers (e. g. by passing an image tensor, the node outputs its flattened dimensions as well as its convolutional output.) Demonstrate that all systems can handle this new module
• Multi-output layers compute shapes correctly for each output handle independently.

[advikasinha]

Have implemented the following in commit on my branch

please check and revert

Task 1: Consistent Edge Naming (done)

• Implementation: Edge.tsx
• Edges read labels from node.data.__handleLabels[handleId]
• Parse handleId from edge ID (e.g., conv_out, flat_out)
• All edges from the same handle share the same label source
• Edge metadata: Includes source, sourceHandle, target, targetHandle

Task 2: Handle Schema System (done)

• New types: handleTypes.ts
• HandleDefinition with defaultLabel, position, optional shapeCompute
• HandleSchema for inputs/outputs
• Integration: BaseClass.tsx
• LayerDefinition.handleSchema field
• Falls back to default shapeCompute when per-handle not provided
• Factory support: CreateNodeComponent.tsx
• Constructs handles from HandleSchema

Task 3: Shape Propagation (done)

• Multi-output support: BaseClass.tsx
• shapeCompute returns number[] | Record<string, number[]>
• Verification: shape_verifier.ts
• Tracks sourceHandle when propagating
• Uses byHandle for per-handle shapes
• Demo node: DualOutputConv.tsx
• Conv2d + Flatten outputs
• Returns { conv_out: [B,C,H,W], flat_out: [B,CHW] }

[ManjotSingh08x]

You haven't shown example of modules which output 2 values. this still covers the outputs in two lines.
show working of a module of the form: output1, output2 = self.layer(input).
also verify that when taking out a 2 edges from a single handle, changing the name of one edge also changes the name of the other head simultaneously.

[advikasinha]

show working of a module of the form: output1, output2 = self.layer(input).

tuple unpacking in line `_34pi, tMI0943 = self.layer_1(x)`

also verify that when taking out a 2 edges from a single handle, changing the name of one edge also changes the name of the other head simultaneously

works

edit: tag @ManjotSingh08x

[ManjotSingh08x]

okay i will look into the source code. you can create a PR on this