Concepts

This section explains the framework design in depth. Read these pages in order for a complete mental model of Grilly.

• Architecture and Runtime
  • What Grilly is
  • Core runtime objects
  • Execution flow
  • CPU fallback model
  • Design choices
  • Resource lifecycle
  • How the high-level APIs fit
• Design Choices
  • 1. Vulkan-first backend
  • 2. PyTorch-like UX with explicit internals
  • 3. CPU fallback paths
  • 4. NumPy as primary tensor interchange
  • 5. Specialized subsystems in one framework
  • 6. Determinism tools for experimental pipelines
• Tensor Model and Shapes
  • Primary tensor type
  • Shape conventions
  • Why shape discipline matters
  • Example input guards
  • Parameter and gradient storage
• Module System and Functional API
  • Two ways to build models
  • grilly.nn overview
  • grilly.functional overview
  • When to choose each API
  • Minimal composition example
  • Functional example
• Training and Optimization
  • Training model in Grilly
  • Canonical loop
  • Optimizers
  • Gradient lifecycle
  • Numerical and runtime tips
  • Related advanced learning ops
• Attention, Transformers, and Decoding
  • Attention stack
  • Why Flash Attention 2 matters
  • Transformer-facing modules
  • Basic attention example
  • Backend flash attention example
  • Decoding usage
• Convolution, Pooling, and Normalization
  • What is covered
  • Data layout
  • Backward support
  • Simple CNN block
  • Design notes
• Spiking Neural Networks
  • SNN support in Grilly
  • LIF neuron model
  • Design choices
  • High-level SNN pipeline
  • Example: high-level usage
  • Example: module usage
  • Practical guidance
• Memory, Retrieval, and FAISS-style Search
  • Retrieval primitives
  • FAISS-style APIs
  • Memory APIs
  • Example search flow
  • Use cases
• Multimodal, Capsule, and VLM Systems
  • Multimodal module families
  • Capsule-related components
  • When to use which approach
  • Design choices
  • Example: simple multimodal fusion
  • Design considerations
• LoRA and Adapter Fine-Tuning
  • Why LoRA
  • Core classes
  • Basic LoRALinear flow
  • Managing inference overhead
  • Checkpointing adapters
• Experimental Language and Cognition
  • Scope
  • Primary components
  • Instant language model
  • Cognitive controller
  • Minimal ingestion example
  • Operational guidance
• Stable Hashing and Ingestion Checkpoints
  • Why stable hashing exists
  • Stable hash utilities
  • Design choices
  • Ingestion checkpoint system
  • Checkpoint flow example
  • When to use it
• Specialized Operations
  • Signal and spectral ops
  • Neuroscience-inspired cells
  • Bridge and conversion ops
  • Domain routing and affective ops
  • Embedding and lookup operations
  • Design choices
  • Practical recommendation
• Compatibility and Integrations
  • PyTorch-style compatibility layer
  • Device management
  • Tensor conversion helpers
  • HuggingFace bridge
  • When to use integrations
• Performance, Debugging, and Testing
  • Performance model
  • Design choices
  • Profiling strategy
  • Debugging checklist
  • Testing workflow
  • Reproducibility tips