The AgentX/InferenceX Kimi-K3 GB300 submission
is published as SLURM (srt-slurm) recipes. If you run Kubernetes and want the
closest deployable equivalent, this package is that translation: the same
topologies on the same hardware class, built from the upstream
recipes/kimi-k3
Dynamo recipe and a public NGC image, field-tested on a GB300 NVL72 Kubernetes
cluster (Amazon EC2 P6e-GB300 UltraServer capacity, 4 GPUs per node, all nodes
in one NVLink domain). Everything here was deployed and verified with live
Last updated: 28 July 2026
Race Mode does not collect, transmit, store or share any personal information.
Race Mode runs entirely inside your browser. When you are on youtube.com it reads the text of video titles, description snippets and search suggestions already on
Last updated: 28 July 2026
Race Mode does not collect, transmit, store or share any personal information.
Race Mode runs entirely inside your browser. When you are on youtube.com it reads the text of video titles, description snippets and search suggestions already on
InferenceX is SemiAnalysis's public dashboard of NVIDIA-vs-AMD LLM inference benchmarks (throughput, accuracy evals, GPU reliability). The front end is a JavaScript single-page app, so it looks like you'd need a headless browser — but the whole dashboard is backed by three plain, unauthenticated JSON endpoints. No API key, no browser, no scraping of rendered HTML. Just HTTP GET.
This gist documents those endpoints and ships a dependency-free Python client
Goal: Simulate 100 multi-turn conversations averaging 12 turns each, with 60K avg ISL / 1K avg OSL, hitting 95% KV cache reuse — a large shared system prompt across all conversations plus accumulating turn-over-turn history.
This is a useful pattern for stress-testing KV-aware routing in disaggregated inference, prefix-cache hit rates, and goodput under realistic agentic-conversation load. It's also the rough shape of real coding-agent traffic (Claude Code, Cursor, Cline, etc.) — long shared system prompt + turn-over-turn growth.
aiperf profile \Date: 2026-02-15/16
Hardware: 8× NVIDIA B200 183GB, NV18 NVSwitch, Xeon 8570 (224 cores), 2TB RAM
GPU Clocks: SM 1965 MHz, MEM 3996 MHz (locked at max)
Driver: 570.195.03
Framework: SGLang v0.5.8.post1 (Docker lmsysorg/sglang:v0.5.8.post1)
Benchmark: sglang.bench_serving, random 1K input / 1K output tokens
OS: Ubuntu 24.04.3 LTS
When an AI agent runs long tasks (benchmarks, deployments, CI pipelines), context overflow is the #1 failure mode. Compaction only runs between turns — if a single turn accumulates too much tool output, the agent dies with no recovery.
A sub-agent doing GPU benchmarks might chain tool calls like:
docker pull→ 200 lines of layer progressdocker logs→ 500 lines of model loading / warmupbench_serving→ 1000+ lines of progress bars
Date: 2026-02-13
Model: zai-org/GLM-5-FP8 (744B MoE, 40B active parameters)
Framework: SGLang (lmsysorg/sglang:glm5-blackwell)
Hardware: 8× NVIDIA B200 183GB, NV18 NVSwitch, Xeon 8570 (224 cores), 2TB RAM
Reference: SGLang GLM-5 Cookbook (8×H200 141GB)
AI Agent: Claude-4.6-Opus Running on OpenClaw