#nvme

Viperblock: High-performance block storage with S3 compatibility for edge to on-premise deployments. Features log-structured writes, block-to-object mapping, and configurable chunk sizes. Supports multiple backends with checksumming and intelligent caching for VM and container storage needs.

Train and serve MoE models that do not fit in VRAM: fused 4-bit experts, QLoRA, CPU/NVMe offload, and fast inference on consumer NVIDIA GPUs.

Filesystem IO for Rust storage engines: journal substrate, io_uring, NVMe passthrough, atomic writes, cross-platform durability.

Triton kernels for 4-bit MoE inference: grouped NF4/MXFP4 GEMM, INT4 GEMV, FP8 paged attention, and CPU/NVMe expert streaming.