Deploy open-weight models on hardware you own.

same config and tensors, so the recipe serves it unchanged. How to swap the weights on a running GB10 fleet (memory-capped transfers, Engram copies with the fleet's own row ranges, one 17-minute window), the same 66 tok/s, and where SGLang's prompt encoder now differs from DeepSeek's reference.

Claude Code's follow-up turns start in about a second instead of a minute — SGLang hoisted Claude Code's per-turn system notes into the top of the prompt, so the prefix cache matched ~4K of 160K tokens and every turn re-read the conversation; now 166K tokens come from the cache. Plus prompt-lookup drafting on top of DSpark (+13–22 % on coding and copy-heavy answers), opt-in after an image-placeholder crash it caused (fixed).

a rank crash under long prompts traced to GB10 page cache the GPU driver will not evict (fixed; 13–20 GB free per node), decode attention on each GPU's 16 real heads instead of padded to 64 (220 tok/s at 8 streams), and YOCO-style prefill that skips the late layers where decode never reads them (128K prompt 52 → 45 s)