Skip to content

DEV Community

Jahn

Inference performance engineering on NVIDIA Blackwell. Conatus AI.

Joined on Aug 25, 2026

Jahn

Sep 9

Name the Blackwell serving cell you are actually in

#nvidia #gpu #llm #inference

5 min read

Jahn

Sep 2

DGX Spark (GB10) memory sizing for LLM serving: the numbers

#nvidia #llm #inference #gpu

7 min read

Jahn

Aug 26

DGX Spark (GB10) bare-metal vLLM: the install that works, two landmines, measured timings

#nvidia #llm #gpu #vllm

2 min read

Jahn

Aug 26

SGLang outputs endless repetition on NVFP4 models: the FP8 lm_head bug

#llm #gpu #machinelearning #debugging

3 min read

Jahn

Aug 25

The sm_120 shared-memory cliff: why FP8 KV cache crashes vLLM on workstation Blackwell

#cuda #llm #gpu #performance

3 min read

Jahn

Aug 25

Did FP8 make the model dumber? A per-prompt regression check for quantized serving

#llm #machinelearning #performance #python

3 min read

Jahn

Aug 25

Qwen3-8B on workstation Blackwell: vLLM vs SGLang vs llama.cpp, plus an FP8 pass

#machinelearning #cuda #performance #llm

3 min read

loading...