Decode · tok/s
53
Prefill
—
tok/100W
22.1
Underdog Saluki 27B on one DGX Spark: 24 → 53 tok/s. Stock llama.cpp with z-lab's DFlash2 drafter + ngram-mod. Full 262k context uses only 26.5 GB.
First-person one-Spark print. Rank is the quoted 53 with the DFlash2 drafter. 24 is the prior path. Dense 27B in ~7.9 GB.
