A persistent key/value cache
Self-learning LLM runtime — TurboQuant KV-cache (6-8x compression), SONA adaptive learning, FlashAttention, speculative decoding, GGUF inference
- Supports [Cloudflare KV](https://developers.cloudflare.com/kv/) cache for `@envelop/response-cache` plugin - Suitable for graphql servers running on [Cloudflare Workers](https://workers.cloudflare.com/)
CDN-edge client hints, tier detection, and KV cache
Improve Pi prompt/KV cache hit rates with stable prompts, OpenAI-compatible cache keys, proxy compat warnings, and footer cache stats.
Routes requests to KV assets
Self-learning LLM runtime — TurboQuant KV-cache (6-8x compression), SONA adaptive learning, FlashAttention, speculative decoding, GGUF inference
Cloudflare KV cache layer for Anvil — key-value caching with TTL
AI agent memory & session orchestrator for MCP — persistent KV-Cache, Soul Board, immutable Ledger
Deno KV cache adapter for Flight Framework - native Deno key-value store
Cloudflare Workers KV cache adapter for Flight Framework
Serverless-simple middleware for managing a KV cache layer
Invert the key/value of an object. Example: `{foo: 'bar'}` → `{bar: 'foo'}`
Local-first MCP toolbelt: delegates summarise/classify/extract/transform tasks from any MCP client to a local oMLX inference server. Apple Silicon, MLX-accelerated, KV-cache persistent.
The next generation web framework for Cloudflare Workers
Memory / Redis abstraction for Directus
Return Cached Key/Value pairs when requested from Chia DataLayer
Utility functions for the workspace
Generate a stable REPOMAP.md context file for any codebase — maximizes LLM KV-cache hits across requests.
LRU Cache
PMLL Memory MCP Server v2 — persistent memory logic loop with short-term KV cache (peek pattern), Q-promise deduplication, Context+ long-term semantic memory graph, and solution engine. Four-way benchmarked: Combined Context+ + PMLL/peek delivers 36ms (TS
A cache object that deletes the least-recently-used items.
kv library - this library implements all the base functionality for NATS KV javascript clients
A lightweight cache for file metadata, ideal for processes that work on a specific set of files and only need to reprocess files that have changed since the last run
LLM inference in Rust
TurboQuant KV-Cache Quantization — 3-bit compression with zero accuracy loss (Zandieh et al., ICLR 2026)
IMECE v0.1.0 - Local-First Autonomous Agent Framework. Zero API dependency. Zero cloud lock-in. Full sovereignty.
Trait interface for compressed KV-cache implementations in mistral.rs
High-performance key-value cache for LLM inference
Experimental Rust implementation of the FibQuant radial-angular vector quantization core
LLM serving runtime with Ruvector integration - Paged attention, KV cache, and SONA learning
C ABI over WombatKV. Produces libwombatkv.dylib + wombatkv.h so C/C++ engines (ds4, llama.cpp, custom) integrate without a Rust path dependency. Headline ABI surface for the system.
Core types, errors, and shared primitives for WombatKV (object-storage-native KV cache system for LLM inference). Leaf crate, no internal deps.
WombatKV daemon binary + SHM/TCP/HTTP listeners for the daemon deployment mode. Sits between multiple engine clients (ds4, future llama.cpp, vLLM/SGLang via wire) and a shared S3 bucket.
Wire-format codecs for WombatKV: 16-byte universal envelope (magic + version + CRC32C + len per RFC 0018), blake3 chain hashing, rkyv archive helpers. Internal to the wombatkv-* workspace; consumers should depend on wombatkv-cabi or wombatkv-node.
Since using kv_cache, Mami need't to warry my db crashing anymore. :)
RRRMatey is an ODM (Object Document Mapper) Framework for Riak, using the Basho Cache Proxy to provide reliable persistence using Riak KV with the speed and accessibility of Redis. Riak's Solr integration provides for fast listings as well as relations.
Use Mysql AUTO_INCREMENT to support key value cache, which should be combined by an integer and string. It means to reduce the database storage size, and improve query performance. All cache will store in process memory, and will never be expired, until the process dies, so the less kvs you use, the better performance you will get. BTW, 100,000 general strings use 10MB memory. Some relatived articles: http://en.wikipedia.org/wiki/Correlation_database Usage ------------------------------------------ ## setup ```ruby create_table :kv_browser_names, :options => 'ENGINE=MyISAM DEFAULT CHARSET=utf8' do |t| t.string :name t.timestamps end class KvBrowserName < ActiveRecord::Base include IdNameCache end ``` or ```ruby create_table :common_tag, :options => 'ENGINE=MyISAM DEFAULT CHARSET=utf8' do |t| t.integer :tagid t.string :tagname end class CommonTag < ActiveRecord::Base self.table_name = :common_tag self.primary_key = :tagid include IdNameCache; set_key_value :tagid, :tagname # include IdNameCache; set_key_value_without_create :tagid, :tagname # if you dont want create it automately end ``` ### use cases ```text ruby-1.9.3-rc1 :001 > QuizTag[1] QuizTag Load (0.3ms) SELECT `common_tag`.* FROM `common_tag` WHERE `common_tag`.`tagid` = 1 LIMIT 1 => "Android" ruby-1.9.3-rc1 :002 > QuizTag[1] => "Android" ruby-1.9.3-rc1 :003 > QuizTag['Android'] QuizTag Load (0.5ms) SELECT `common_tag`.* FROM `common_tag` WHERE `common_tag`.`tagname` = 'Android' LIMIT 1 => 1 ruby-1.9.3-rc1 :004 > QuizTag['Android'] => 1 ``` == Copyright MIT, David Chen at eoe.cn
ignis-dl is the deep-learning layer of the Ignis ecosystem: NN modules (Linear, Embedding, LayerNorm, RMSNorm, Dropout), optimizers (SGD/Adam/AdamW), losses, and a transformer stack (multi-head + grouped-query attention, RoPE, SwiGLU, KV cache) with HuggingFace weight loaders (GPT-2, Llama). Loads real GPT-2 and Llama-3.2 checkpoints and matches HuggingFace logits, and trains transformers from scratch — in Ruby, on native Windows. Installing this pulls the whole stack (ignis + ignis-autograd), so it also serves as the meta-gem.
No description provided.
No description provided.
No description provided.
No description provided.
No description provided.
No description provided.
No description provided.
No description provided.
No description provided.