rocev2

Network-faithful simulation of LLM serving and training deployments

Early-warning bottleneck profiler for GPU training nodes: GPU + RDMA fabric telemetry with job-level classification

Rust library and CLI for the FERoCE (Front-End RoCE v2) FPGA network stack

KAI Data Center Builder

Performance metrics for AI / ML cluster

Packet-level simulator (ns-3 + RoCEv2/DCQCN/PFC/ECN) comparing fat-tree vs rail-optimized topologies for AI training fabrics. Validated against NCCL on real GPUs.

A distributed hardware-software co-design fabric for MoE models (DeepSeek-V3, Mixtral) that eradicates NCCL All-to-All communication stalls via distributed RoCEv2 RDMA virtual address MUX and JAX/XLA SPMD sharding.

Software examples for RDMA senders and receivers using RDMA WRITE and RDMA READ

CLI tool for RoCEv2 network analysis

Ultra-Ethernet compliance test architecture eliminating expensive HBM memory buffers. Design utilizes FPGAs, PRBS payloads, state hash tables, and virtual RDMA.

A simple experiment applying compressible flow principles to soften distributed gradient communication stalls.

Lossless-Ethernet benchmark harness for RoCEv2 fabrics: sweeps queue thresholds, ECN marking, and oversubscription to locate throughput collapse and PFC pressure.

PowerShell-based toolkit for Storage Spaces Direct strict bundle deployment, validation, and operational checks in Windows Server 2025 environments.

RoCEv2 control plane simulator — DCQCN congestion control on fat-tree GPU clusters with NCCL Ring AllReduce workloads

GLM-5.3-Flash (320B MoE) at TP=4 across four ASUS Ascent GX10 / DGX Spark nodes on a 100G RoCE fabric: 76 tok/s structured, 3.9M-token fp8 KV pool, 1M context, DFlash2 drafting. Field notes on what broke and why.