How kapildhinwa2010-dev/fluidtech works
FluidTech is a memory layer for causal language models. The recommended workflow attaches BaseKVLayer to hidden states and token IDs: its exact tier replays seen contexts, its bounded KV window evicts old entries, and long-term memory retains each eviction with its hidden state. The README also documents optional graft, Mamba, KV-only, Ultra, and combined layers, plus quantization, acceleration, and multi-GPU support. Main-layer internals were not sampled, so those edges reflect README statements; optional implementation links are limited to explicit excerpts or index references.
Main entry points: base_kv_layer.py Base + KV · combined_layer.py Combined layer · blocks.py Fluid block · diffusion.py Diffusion
Components 24
Main memory layer
- Base + KV
fluidlayer/base_kv_layer.py - Exact replay
fluidlayer/memory.py - Bounded KV
fluidlayer/base_kv_layer.py - Long-term memory
fluidlayer/base_kv_layer.py
Optional layers
- Base graft
fluidlayer/graft_bulletproof.py - Mamba layer
experiments/superfluid_mamba_layer.py - KV wrapper
kv_cache/constant_kv_superfluid.py - Ultra stack
experiments/ultra_fluid_layer.py - Combined layer
combined/combined_layer.py
Fluid architecture
- Fluid model
fluidlayer/model.py - Fluid block
fluidlayer/blocks.py - Topology routing
fluidlayer/topology.py - Sheaf geometry
fluidlayer/sheaf.py - Diffusion
fluidlayer/diffusion.py - Autodiff engine
fluidlayer/autodiff.py - Graft variants
fluidlayer/graft.py - Refactored graft
fluidlayer/graft_refactored.py - Loss functions
fluidlayer/losses.py - Optimizer
fluidlayer/optim.py - Corpus workflow
experiments/e6_real_corpus.py
Runtime features
- Quantization
fluidlayer/quantization.py - Acceleration
fluidlayer/acceleration.py - Sharded KV
fluidlayer/multi_gpu.py
Other components
- Causal model
Connections 28
- Causal model → Base + KV: attaches to
- Base + KV → Exact replay: replays contexts
- Base + KV → Bounded KV: updates window
- Base + KV → Long-term memory: stores evictions
- Base + KV → Base graft: optional base-only
- Base + KV → Mamba layer: optional alternative
- Base + KV → KV wrapper: optional alternative
- Base + KV → Ultra stack: optional alternative
- Base + KV → Combined layer: optional alternative
- Fluid model → Fluid block: uses blocks
fluidlayer/model.py - Fluid block → Autodiff engine: uses tensors
fluidlayer/blocks.py - Fluid block → Topology routing: builds geometry
fluidlayer/blocks.py - Fluid block → Diffusion: filters bands
fluidlayer/blocks.py - Topology routing → Sheaf geometry: uses geometry
fluidlayer/topology.py - Graft variants → Diffusion: uses diffusion
fluidlayer/graft.py - Refactored graft → Autodiff engine: uses tensors
fluidlayer/graft_refactored.py - Refactored graft → Diffusion: uses diffusion
fluidlayer/graft_refactored.py - Loss functions → Diffusion: uses diffusion
fluidlayer/losses.py - Corpus workflow → Graft variants: runs graft
experiments/e6_real_corpus.py - Corpus workflow → Optimizer: uses optimizer
experiments/e6_real_corpus.py - KV wrapper → Mamba layer: constructs memory
kv_cache/constant_kv_superfluid.py - Combined layer → Mamba layer: constructs branch
combined/combined_layer.py - Combined layer → Ultra stack: constructs stack
combined/combined_layer.py - Combined layer → Base graft: constructs branch
combined/combined_layer.py - Combined layer → KV wrapper: constructs cache
combined/combined_layer.py - Base + KV → Quantization: supports quantization
- Base + KV → Acceleration: supports acceleration
- Base + KV → Sharded KV: supports sharding