Field notes on LLM serving, fine-tuning, GPU performance, and agentic AI — measured on real hardware.https://blog.swesty.net/https://blog.swesty.net/blog/vllm-on-dgx-spark-gb10-field-notes/https://blog.swesty.net/blog/vllm-on-dgx-spark-gb10-field-notes/Seven months of empirical GB10 (sm_121) vLLM findings: why 'Marlin is mandatory' has a load-bearing exception (it's your checkpoint, not your GPU), the two-condition rule for native FP4, the MTP illegal-memory-access saga and how the W4A4/CUTLASS configuration avoided it, unified-memory OOM math, and measured numbers for every claim.Sun, 19 Jul 2026 00:00:00 GMTAI InfrastructureLLM Serving