nx_zimage_real_qkv.nx
buildroot/runtime/nx_zimage_real_qkv.nx
about
nx_zimage_real_qkv.nx -- REAL Qwen layer-0 Q/K/V projections on REAL weights, sovereignly.
sd-server -> Nishi migration: extends the real forward from Q to the FULL attention input -- computes
Q, K and V projections of token-0's real embedding against blk.0.attn_{q,k,v}.weight (all Q6_K, ~335MB
into the file) with our own f32 matmul, and records the out-dims (confirming GQA: q_out=4096 > kv_out).
Real Qwen weights -> our sovereign matmuls -> the real attention inputs. Bounded 384MB prefix; only the
slices we need are dequantized.
license_tier: ORIGINAL
dependencies 13 imports · 0 importers
diagram shows first 10 each side; +3 more imports, +0 more importers in the complete lists below.
imports: nx_syscalls.nxnx_tier.nxnx_le.nxnx_strconv.nxnx_tensor.nxnx_gguf.nxnx_gguf_load.nxnx_gguf_meta.nxnx_placement.nxnx_gguf_load_lazy.nxnx_q6_k_to_f32.nxnx_f32.nxnx_f32_cvt.nx
imported by: nobody (leaf or entry point)
structs
| none |
consts
| none |
functions
| 23 | func qkv_emit(fd: i64, key: *u8, key_len: i64, value: i64) -> i64 |
| 39 | func qkv_proj4(buf: *u8, w_off: i64, HID: i64, embed: *i64, out: *i64) -> i64 |
| 54 | func main() -> i64 |