Articles for people who
ship the stack
Rewrites on React, Node.js, TypeScript and AI. A translated body is shown when it exists. If it does not, you see the English text and a note that says so.
Tagged: GPU
Why Dual RTX 3090s Hit 55 tok/s Instead of 220 with SGLang and DFlash2
A hands-on look at running Qwen3.8-27B on two 3090s with SGLang, and how PCIe x4 links, missing P2P, tight KV cache and power spikes crushed the advertised speed.
7 min readvLLM Serve Flags Explained, With Sensible Starting Values
A practical tour of the vllm serve flags you'll actually tune: dtype, memory use, context length, parallelism, API keys and quantization, with defaults to try.
4 min readLocal LLM Workstation: From 20 to 60 Tokens per Second with Two GPUs
A used Xeon box with two RTX 5060 Ti cards, Qwen 27B, tensor parallelism and NCCL: hardware rules, real speed gains, and a break-even calculation against API pricing.
9 min readFlashAttention and PagedAttention: Two Fixes for Two Different Bottlenecks
Learn how FlashAttention speeds up attention math through tiling and online softmax, how PagedAttention manages KV cache memory, and why vLLM-style stacks use both.
7 min readCerebras Wafer vs GPU for LLM Inference: A GPT-OSS Benchmark
Compare Cerebras' wafer-scale engine with GPU inference using GPT-OSS-120B across five workloads, then see how Ultrafast speeds up GPT-5.6 Sol through the API.
4 min readRunning Local Coding Models on a 12GB RTX 3060 With a Model Scheduler
How to pick models that fit a 12GB GPU and why switching between GGUF and SafeTensors backends behind one OpenAI-compatible endpoint turns into a scheduling problem.
5 min readQwen3.8-27B on One RTX 3090: 262K Context at 65-80 Tokens per Second
A recipe for running a dense 27B model with full context on a single 24 GB GPU: a tuned llama.cpp fork, a 14.5 GB quant, Turbo3 cache and MTP drafting.
8 min readQwen3.8-27B on One RTX 5090: vLLM, LACT Fan Control and Concurrency Benchmarks
Serve Qwen3.8-27B from a headless RTX 5090 with vLLM: Docker GPU setup, Prometheus and Grafana, fixing thermals with LACT, and where throughput saturates under load.
11 min readKV cache is the unseen database eating GPU memory in LLM decode
The KV cache quietly eats GPU memory during LLM inference: prefill vs decode, KV math, GQA and MQA, PagedAttention, quantization and eviction policies.
9 min read
About these articles
Request a 24h estimate
Need the same stack in a production operator layer? Send the brief — estimate within 24 hours.