Network-faithful simulation of LLM serving and training deployments
-
Updated
Oct 8, 2026 - Python
Network-faithful simulation of LLM serving and training deployments
RDMA fork version of Strix Halo inference engine. Qwen Flash Next Q4_K_XL: >2100pp …
JONATHAN GPUBASE Fabric — open-core Ethernet/RoCEv2 GPU cluster network fabric manager by ACRYL
Early-warning bottleneck profiler for GPU training nodes: GPU + RDMA fabric telemetry with job-level classification
Open-source RoCE v2 engine for FPGAs: line-rate RDMA between an FPGA and a host or GPU over 100G Ethernet, vendor-neutral SystemVerilog, MIT
KAI Data Center Builder
NCCL over a switchless 4-node DGX Spark / GB10 ring using both PCIe halves of every QSFP cable: ~193 Gb/s per cable instead of ~112, up to +34% vLLM prefill. One patch on top of switchless-nccl.
Reference design for the Opsero 2x QSFP28 FMC on the KCU116: 100G RDMA with rocelet, Holoscan Sensor Bridge mode, and a live phased-array beamforming demo on a GPU
Performance metrics for AI / ML cluster
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.
Two-node Soft-RoCE lab: RoCEv2 packet evidence, RDMA benchmarks, and NCCL AllReduce over RDMA vs TCP sockets.
Line-rate out-of-band network packet classification & RoCEv2 isolation engine offloaded to SmartNIC/DPU cores.
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 simple experiment applying compressible flow principles to soften distributed gradient communication stalls.
Ultra-Ethernet compliance test architecture eliminating expensive HBM memory buffers. Design utilizes FPGAs, PRBS payloads, state hash tables, and virtual RDMA.
Visual, exam-focused study guide for Cisco 300-640 DCAI v1.0
Production-grade High-Frequency Streaming Telemetry Ingestion (TIG Stack: Telegraf, InfluxDB v2, Grafana) & Sub-Second Buffer Incast Inspection for AI Data Center Fabrics (Cisco Nexus Streaming Telemetry aligned).
To associate your repository with the rocev2 topic, visit your repo's landing page and select "manage topics."