code wiki / (root) / nx_lowrank_weight_spectrum.nx

nx_lowrank_weight_spectrum.nx

buildroot/runtime/nx_lowrank_weight_spectrum.nx

8419 B185 linesdepth 8pulls 19 transitivereach 0 importersview sourcekind tooltopic lowrank

docs dependencies structs consts functions

about

nx_lowrank_weight_spectrum.nx -- the FORWARD-FREE VRAM lever: are the real model's WEIGHT matrices low-rank? (FlashSVD / over-parameterization: arXiv 2508.01506). Unlike the KV-cache lever, this needs NO forward -- just load a weight tensor + measure its singular spectrum -- so it iterates without the slow software-f32 forward. Measures attn_q (square 896x896) AND ffn_gate (4864x896, the big MLP weight where most params live) for Qwen2.5-0.5B: cumulative covariance energy in the top-r of the smaller (cols) dimension. High energy at small r => low-rank => SVD-compressible. Sovereign (nx_cc->nxasm, no gcc). NOTE: SVD funcs copied from nx_lowrank_kv_real (proven f32+reorth) -- TODO extract both to nx_lowrank_svd.nx (DRY #15). license_tier: ORIGINAL

dependencies 8 imports · 0 importers

imports: nx_syscalls.nxnx_gguf.nxnx_gguf_load.nxnx_gguf_meta.nxnx_gguf_load_f32.nxnx_f32.nxnx_f32_div.nxnx_f32_hw.nx

imported by: nobody (leaf or entry point)

call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown

structs

none

consts

none

functions

17	func w_puts(s: *u8) -> i64 { var n: i64 = 0; while s[n] != (0 as u8) { n = n + 1 } sys_write(1, s, n); return 0 } called by 3: w_spectrum w_measure main calls 1: sys_write
18	func w_putn(v: i64) -> i64 called by 2: w_spectrum w_measure calls 2: sys_write sys_mmap
30	func r_is_zero(x: i64) -> i64 { if (x & 0x7fffffff) == 0 { return 1 } return 0 } called by 3: r_normalize r_power_one w_spectrum
31	func r_dot(a: i64, b: i64, n: i64) -> i64 called by 2: r_normalize r_power_one calls 2: nx_f32_add nx_f32_mul
37	func r_normalize(v: *i64, n: i64) -> i64 called by 1: r_power_one calls 6: r_dot r_is_zero nx_f32_sqrt nx_f32_div f32_of nx_f32_mul
47	func r_gmatvec(G: i64, v: i64, d: i64, out: *i64) -> i64 called by 1: r_power_one calls 2: nx_f32_add nx_f32_mul
58	func r_deflate1(v: i64, basis: i64, found: i64, d: i64) -> i64 called by 1: r_reorth calls 3: nx_f32_add nx_f32_mul nx_f32_sub
70	func r_reorth(v: i64, basis: i64, found: i64, d: i64) -> i64 called by 1: r_power_one calls 1: r_deflate1
73	func r_power_one(G: i64, d: i64, basis: i64, found: i64, v: i64, w: i64) -> i64 called by 1: w_spectrum calls 6: f32_of r_reorth r_normalize r_gmatvec r_is_zero r_dot
98	func r_gram(W: i64, rows: i64, cols: i64, G: i64) -> i64 called by 1: w_spectrum calls 2: nx_f32_add nx_f32_mul
113	func r_trace(G: *i64, d: i64) -> i64 called by 1: w_spectrum calls 1: nx_f32_add
121	func w_spectrum(W: i64, rows: i64, cols: i64, label: u8) -> i64 called by 1: w_measure calls 12: sys_mmap r_gram r_trace w_puts w_putn r_is_zero+6
147	func w_measure(buf: u8, hdr: NxGgufHeader, name: u8, name_len: nx_int, label: u8) -> i64 called by 1: main calls 7: nx_gguf_find_tensor w_puts nx_gguf_tensor_at sys_mmap nx_gguf_load_tensor_to_f32 w_putn+1
166	func main() -> i64 calls 6: sys_mmap sys_read_file w_puts nx_gguf_parse w_measure sys_exit