code wiki / _hdl_build / nx_vram_frontier.nx
nx_vram_frontier.nx
buildroot/runtime/_hdl_build/nx_vram_frontier.nx
about
nx_vram_frontier.nx -- MAP THE FRONTIER: the achievable set on a fixed consumer GPU is not set
by the hardware, it is PUSHED OUTWARD by smarter math. For a budget, this walks a ladder of
stacked techniques (fp16 -> Q4 weights -> +Q8 KV -> +GQA -> +Q4 K-cache) and shows, at each
rung, (a) the LARGEST model that fits at a fixed context and (b) the LONGEST context that fits
for a fixed model. Each rung should dominate the one below: same silicon, more capability.
The deep-research adds further rungs (2-bit/BitNet weights, MLA, KV eviction, weight streaming).
license_tier: ORIGINAL
dependencies 1 imports · 1 importers
imports: nx_vram_budget.nx
imported by: nx_vram_frontier_test.nx
structs
| none |
consts
| none |
functions
| 12 | func vfr_usable_mb(budget_gb: i64, act_mb: i64, overhead: i64) -> i64 called by 1: main |
| 17 | func vfr_max_model_m(usable_mb: i64, kv_mb: i64, w_bits: i64) -> i64 called by 1: main |
| 24 | func vfr_max_ctx(usable_mb: i64, weights_mb: i64, kv_at_ref_mb: i64, ref_ctx: i64) -> i64 called by 1: main |