How snap-research/grid works
GRID is a generative-recommendation framework whose main workflow turns item text into language-model embeddings, learns hierarchical semantic IDs by quantization, then trains and runs a TIGER transformer to generate recommendations from user histories. The README documents this staged handoff and supported semantic-ID approaches; source-index references and excerpts show the runtime entry point, clustering abstractions, model blocks, and data contracts. Several feature internals are unsampled, so the graph preserves their documented roles without asserting unverified implementation-level calls.
Main entry points: tiger_generation_model.py TIGER generation model · base_clustering_module.py Clustering base · inference.py Inference entry · semantic_embedding_inference_module.py Embedding inference
Components 21
Runtime orchestration
- Training entry
src/train.py - Inference entry
src/inference.py - Pipeline launcher
src/utils/launcher_utils.py
Data and embeddings
- Prepared datasets
- Sequence data module
src/data/loading/datamodules/sequence_datamodule.py - Loading components
src/data/loading/components - Embedding inference
src/modules/semantic_embedding_inference_module.py - HuggingFace language model
src/components/network_blocks/hf_language_model.py - Item embeddings
Semantic ID learning
- Residual quantization
src/modules/clustering/residual_quantization.py - Vector quantization
src/modules/clustering/vector_quantization.py - Mini-batch k-means
src/models/modules/clustering/mini_batch_kmeans.py - Clustering base
src/models/modules/clustering/base_clustering_module.py - Clustering primitives
src/components/clustering_initializers.py - Semantic IDs
Generative recommendation
- TIGER generation model
src/models/modules/semantic_id/tiger_generation_model.py - Transformer base
src/models/modules/huggingface/transformer_base_module.py - Embedding aggregation
src/models/components/network_blocks/embedding_aggregator.py - Generated recommendations
- Model interfaces
src/models/components/interfaces.py
Other components
- Researcher
Connections 22
- Researcher → Inference entry: runs inference
README.md - Researcher → Training entry: runs training
README.md - Training entry → Pipeline launcher: launches pipeline
src/train.py - Inference entry → Pipeline launcher: launches pipeline
src/inference.py - Researcher → Prepared datasets: prepares data
README.md - Prepared datasets → Sequence data module: loads sequences
README.md - Sequence data module → Loading components: uses components
src/data/loading/datamodules/sequence_datamodule.py - Researcher → Embedding inference: generates embeddings
README.md - Embedding inference → HuggingFace language model: uses model
src/components/network_blocks/hf_language_model.py - Embedding inference → Item embeddings: produces embeddings
README.md - Item embeddings → Residual quantization: learns IDs from
README.md - Residual quantization → Clustering base: extends clustering
src/modules/clustering/residual_quantization.py - Vector quantization → Clustering base: uses base module
src/modules/clustering/vector_quantization.py - Mini-batch k-means → Clustering base: extends base module
src/models/modules/clustering/mini_batch_kmeans.py - Clustering base → Clustering primitives: initializes centroids
src/models/modules/clustering/base_clustering_module.py - Residual quantization → Semantic IDs: produces IDs
README.md - Semantic IDs → TIGER generation model: trains on IDs
README.md - TIGER generation model → Transformer base: uses base
src/models/modules/semantic_id/tiger_generation_model.py - Transformer base → Embedding aggregation: uses aggregator
src/models/modules/huggingface/transformer_base_module.py - TIGER generation model → Model interfaces: uses interfaces
src/models/modules/semantic_id/tiger_generation_model.py - TIGER generation model → Generated recommendations: generates recommendations
README.md - Inference entry → Generated recommendations: returns predictions
README.md