How sparticle62ops/pssa works
PSSA is a Rust language-model runtime centered on recurrent state-space processing, episodic memory, and plastic updates. Its main workflow is corpus preparation and tokenization, model training, checkpointing, then generation or held-out evaluation. A parameter-matched transformer baseline supports comparison. The graph preserves major README capabilities; relationships not visible in the supplied excerpts or README are omitted rather than inferred.
Main entry points: cli.rs CLI workflows · pssa.rs PSSA recurrent model · comparison.rs Model comparison · dataset.rs Dataset and tokenizer
Components 22
Data preparation
- Dataset and tokenizer
src/dataset.rs
PSSA model
- PSSA recurrent model
src/pssa.rs - Episodic memory bank
src/memory.rs - Plastic adapter
src/adapter.rs
Training and runtime
- CLI workflows
src/cli.rs - PSSA training
src/training.rs - Sequence batching
src/sequence_batch.rs - PSSA checkpoints
src/checkpoint.rs - PSSA generation
src/inference.rs - Held-out evaluation
src/evaluation.rs - Loss reporting
src/loss_csv.rs - Training interface
src/ui.rs
Transformer baseline
- Transformer model
src/transformer.rs - Transformer training
src/transformer_training.rs - Transformer checkpoints
src/transformer_checkpoint.rs - Transformer generation
src/transformer_inference.rs - Model comparison
src/comparison.rs
Numerical compute
- GPU batch kernels
src/gpu_batch.rs - Compute backend
src/backend.rs - Linear algebra
src/linalg.rs
Other components
- User
- Text corpus
Connections 40
- User → CLI workflows: runs workflows
src/cli.rs - Text corpus → Dataset and tokenizer: loads and tokenizes
src/dataset.rs - CLI workflows → Dataset and tokenizer: loads corpus
src/cli.rs - CLI workflows → PSSA training: starts training
src/cli.rs - PSSA training → Dataset and tokenizer: plans token chunks
src/training.rs - PSSA training → PSSA recurrent model: trains model
- PSSA recurrent model → Episodic memory bank: reads and writes
src/pssa.rs - PSSA recurrent model → Plastic adapter: uses adapter
src/pssa.rs - PSSA recurrent model → Compute backend: dispatches compute
src/pssa.rs - PSSA recurrent model → Linear algebra: uses operations
src/pssa.rs - Sequence batching → PSSA checkpoints: uses checkpoint types
src/sequence_batch.rs - GPU batch kernels → PSSA recurrent model: batches model work
src/gpu_batch.rs - GPU batch kernels → Episodic memory bank: uses memory operations
src/gpu_batch.rs - GPU batch kernels → Compute backend: dispatches kernels
src/gpu_batch.rs - GPU batch kernels → Linear algebra: uses operations
src/gpu_batch.rs - CLI workflows → PSSA checkpoints: loads and saves
src/cli.rs - CLI workflows → PSSA generation: generates text
src/cli.rs - CLI workflows → Held-out evaluation: scores models
src/cli.rs - CLI workflows → Loss reporting: configures loss output
src/cli.rs - CLI workflows → Training interface: renders interface
src/cli.rs - PSSA generation → Dataset and tokenizer: uses tokenizer
src/inference.rs - PSSA generation → PSSA recurrent model: runs model
src/inference.rs - Held-out evaluation → Dataset and tokenizer: encodes corpus
src/evaluation.rs - Held-out evaluation → PSSA recurrent model: scores model
src/evaluation.rs - CLI workflows → Transformer training: starts baseline training
src/cli.rs - Transformer training → Dataset and tokenizer: loads token data
src/transformer_training.rs - Transformer training → Transformer model: trains model
src/transformer_training.rs - Transformer training → Transformer checkpoints: resumes model
src/transformer_training.rs - Transformer training → PSSA checkpoints: imports tokenizer
src/transformer_training.rs - Transformer generation → Transformer checkpoints: loads model
src/transformer_inference.rs - Transformer generation → Transformer model: runs model
src/transformer_inference.rs - Transformer generation → PSSA generation: reuses sampler
src/transformer_inference.rs - Transformer generation → Dataset and tokenizer: loads corpus
src/transformer_inference.rs - Transformer generation → Held-out evaluation: evaluates model
src/transformer_inference.rs - Model comparison → Held-out evaluation: compares scores
src/comparison.rs - Model comparison → PSSA training: runs PSSA training
src/comparison.rs - Model comparison → Transformer training: runs baseline training
src/comparison.rs - Model comparison → Transformer model: compares model
src/comparison.rs - Model comparison → Transformer checkpoints: uses baseline checkpoint
src/comparison.rs - Model comparison → PSSA checkpoints: uses PSSA checkpoint
src/comparison.rs