[RMP] Add support for ranking models in PyTorch

[marcromeyn]

Problem:

Goal:

Add support for ranking models in PyTorch.

New Functionality

• Models
  • Support Ranking Models
• Systems
  • Make sure torch ranking models can be served

Starting Point:

Implement base-classes of block-API in PyTorch

People: @marcromeyn

Currently the block-API is T4Rec is using a similar design to Keras to allow for modules that lazily initialize their variables. We would like to deprecate this in favor of a native way to achieve the same thing that could launched recently.

• Introducing Block models#1087
• Adding ParallelBlock models#1088
• Adding registry models#1090
• Adding Link models#1091
• Adding simple aggregations: concat & stack models#1092
• Adding sample_batch & sample_features models#1095
• Adding SchemaTrackingMixin models#1109
• Add Encoder & Predictor models#1112
• Adding RouterBlock models#1096

Input-blocks

People: @marcromeyn

• Implement Continuous & EmbeddingTables
• Implement TabularInputBlock
• Add Encoder & Predictor models#1112
• Add schema tracing models#1149
• Add support for sequential-features in input-blocks
• Do performance testing of holding multiple features in a single embedding-table

Output-blocks

People: @edknv & @marcromeyn

• Add BinaryOutput models#1099
• Add RegressionOutput models#1115
• Add CategoricalOutput models#1158
• Port TabularOutputBlock (for multi-task learning)
• Adding MMOE & PLE models#1173

Blocks

• Adding MLPBlock models#1093
• Adding CrossBlock (used in DCN-v2) models#1172
• Add DLRM block models#1162

Models

• Add DLRM Model models#1171
• Add DCNModel
• Export Model with Input / Output Schemas (Serve in Systems Ensemble)
  • Ensure model outputs are compatible with Triton (tensor or NamedTuple instead of dict)

Examples

• Getting Started from #1159
• Serving Ranking Model with Systems (similar to Serving Ranking Models With Merlin Systems (TF) )
  • multi-task model (multiple model named outputs)