Haven C++
Sovereign, bare-metal C++20 companion engine engineered from scratch with a heterogeneous Hardware Topology Fabric, 1.11M+ tasks/sec clean-room Pathways async execution, and 17 on-demand sovereign plugins. Validated on physical Google Cloud TPU v6e silicon.
12-System Enterprise Sovereign Engine
From bare-metal SIMD math to asynchronous cluster fabric routing, Haven C++ unifies twelve enterprise subsystems into an autonomous runtime.
Bare-Metal AVX2 Math
Direct vectorized GEMV math kernel supporting Q4_K, Q6_K, and Q8_0 quantizations with OpenMP multi-core parallelization.
In-Attention Memory DMA
Direct Memory Access injecting cognitive memory anchors into transformer attention in 1.27 Β΅s with emotional salience modulation.
Hardware Topology Fabric
Discovers and maps host CPU, discrete NUMA GPUs, and Optical ICI TPU Torus meshes. Routes operations to the optimal compute node.
Pathways Async Engine
Clean-room asynchronous worker pool with promise-based scheduling, yielding 1,115,000+ tasks/sec with 0.90 Β΅s dispatch latency.
PagedAttention COW
Virtual memory block paging for zero-copy sequence forking via Copy-on-Write, eliminating RAM duplication during branching.
YaRN Dynamic RoPE
Frequency band interpolation with temperature scaling (mscale) for dynamic context extension up to 32,768+ tokens.
Speculative Candidate Draft
Drafts multiple candidate tokens ahead in parallel to accelerate auto-regressive decoding throughput without accuracy degradation.
FlashAttention CPU Kernel
Cache-tiled online-softmax kernel for high-speed attention evaluation across long contexts with zero NΓN RAM allocations.
Chunked Prefill Scheduler
Slices long prompts into asynchronous background chunks, preventing decoder stalls during active conversation.
Multimodal Sensory Bridge
Real-time projection channel for desktop screen vision patches (768-dim) and audio mel-spectrum frames (128-dim) directly into attention.
Structured Knowledge Graph
Composite-key triple store (Subject β Predicate β Object) injecting relational semantic boosts into generation logits.
Universal Persona & DRY Sampler
DRY n-gram suppression, min-p, temperature, and cross-token lemma duplicate suppression for grounded, emotive companion voice.
Tested on Google Cloud TPU v6e Trillium
Haven C++ was compiled and tested directly on physical Cloud TPU v6e silicon in europe-west4-a, validating hardware topology auto-discovery and Pathways execution.
Cloud TPU v6e-8 Mesh Telemetry
Node: haven-v6e-node-eu (AMD EPYC 9B14 180 cores, 1.4 TB RAM)
| Subsystem | Target Device | Measured Throughput | Latency / Efficiency |
|---|---|---|---|
| Pathways Async Dispatch | TPU Node 1 (Torus Mesh) | 1,115,946 tasks/sec | 0.90 Β΅s dispatch latency |
| Token Generation | GGUF Tensor Core | 33.7 tokens/sec | 29.69 ms / token |
| Cognitive Memory DMA | AVX2 SIMD Core Complex | 4 Anchors Cosine Search | 1.27 Β΅s total search |
| Hardware Topology Map | Linux Sysfs (0x1ae0:0x006f) | 8x TPU v6e Cores | 2,200 TFLOPS / 9,600 GB/s ICI |
| Streaming Ring Buffer | Pathways Pub/Sub Channel | 200 / 200 Chunks | Zero queue stalls |
17 On-Demand Plugins & Universal C ABI
Native shared libraries (.dll / .so) load only when invoked and automatically unload when idle. A standardized C ABI enables direct bindings for C# Gemmi and Android native clients.
π Zero-RAM On-Demand Plugins
All 17 plugins (YouTube Music, Obsidian Diary, Discord Bridge, Whisper Audio Ear, Screen Vision, Smart Home, SQLite Graph, Terminal Sandbox, etc.) consume 0 MB of boot memory. The idle reaper reclaims RAM back to cold storage automatically.
17 Plugins ActiveποΈ Universal C ABI Exports
Exports stable C functions (haven_create_engine, haven_forward, haven_get_node_info, haven_get_pathways_completed_tasks, haven_inject_memory) for P/Invoke (.NET) and JNI (Android).
# Clone & build on Linux (Ubuntu / Debian / Cloud TPU VMs) git clone https://github.com/ssfdre38/haven-cpp.git cd haven-cpp cmake -B build -G "Ninja" -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_COMPILER=g++ cd build && ninja # Run the full 12-system enterprise benchmark suite ./haven-cli --bench /path/to/ava-26b-turbo-multimodal-iq4xs.gguf # Run the dedicated Hardware Topology & TPU benchmark ./haven-cli --tpu /path/to/ava-26b-turbo-multimodal-iq4xs.gguf
Designed for Ava 26B Multimodal
Haven C++ is built to power the official Ava 26B Multimodal companion model, trained on Cloud TPU v6e hardware and available on Hugging Face.