code wiki / q4k

topic: q4k

22 modules sharing the q4k name family (derived from the tree's prefix discipline).

The 'q4k' topic family focuses on implementing and optimizing integer-based GEMM operations within the Nishi sovereign ecosystem, emphasizing efficiency and fidelity in quantized neural network inference. Modules like nx_q4k_dot_simd leverage SIMD acceleration for high-performance dot products, while nx_q4k_fused_fidelity_gate investigates whether fused integer GEMM accurately represents real-trained weights. nx_q4k_linear provides a core integer linear transformation, bridging quantized weights and activations to produce precise outputs.

auto-narrated by the local model from this topic's module headers; links verified against the wiki index.

narrated overview -- maintained by the narration lane, module links verified against this wiki.

moduledescriptionlinesfuncs
nx_q4k_dot_simd.nxSIMD-accelerated Q4_K fused dot (vpmaddwd), bit-exact vs nx_q4k_dot_row_col.1653
nx_q4k_dot_simd2_lib.nxthe fused Q4_K SIMD dot with the COMPILER-EMITTED unpack (2026-09-02, LM4c).1051
nx_q4k_dot_simd_lib.nxthe fused Q4_K SIMD dot as a LIBRARY (extracted 2026-09-02 from nx_q4k_dot_simd.nx,1465
nx_q4k_fused_fidelity_gate.nxDOES THE FUSED INTEGER GEMM TELL THE TRUTH ON REAL TRAINED WEIGHTS?1765
nx_q4k_fused_vs_x4_gate.nxTHE HEAD-TO-HEAD the SOTA decision actually rests on:2576
nx_q4k_gemm_mt.nxmulti-threaded integer GEMM: wire nx_thread into the Q4_K integer dot.1453
nx_q4k_ggml_kat.nxPERMANENT golden regression guard for Q4_K dequant.34211
nx_q4k_linear.nxthe integer GEMM linear: real Q4_K weights x Q10 activation -> Q10 output.1213
nx_q4k_linear_hp.nxhigher-precision integer linear (Q14 activation -> Q24 output).1394
nx_q4k_matmul.nxfused-dequant Q4_K row × column dot product.1403
nx_q4k_matmul_bench.nxfused vs materialize-then-dot bench.160
nx_q4k_matmul_bench_test.nxthe fused-vs-materialized bench.1883
nx_q4k_matmul_mt_gate.nxadversarial gate for the multi-threaded2407
nx_q4k_matmul_rate.nxisolate the SSE Q4_K matmul rate (no 491MB model load) to pinpoint why the LM743
nx_q4k_matmul_test.nxsmoke + bit-exact KAT for fused Q4_K dot.1302
nx_q4k_matmul_x4_gate.nxadversarial gate for the PACKED-SIMD3148
nx_q4k_q24_gate.nxprove the exact Q24 super-scale decode on the REAL Q4_K super-scales.281
nx_q4k_real_gemm.nxthe CPU SPEED PATH proven on a REAL Qwen weight: integer fused-dequant-dot vs f32.1312
nx_q4k_simd2_gate.nxthe REFEREE for __q4k_unpack32s (the nx_cc intrinsic) and nx_q4k_dot_simd2 (2026-09-02, LM4c).2037
nx_q4k_speed_bench.nxmeasure the CPU speedup of the integer fused-dequant-dot vs emulated-f32.1212
nx_q4k_to_f32.nxbits-up Q4_K dequantization producing IEEE 7542976
nx_q4k_to_f32_test.nxsmoke for nx_q4k_to_f32.nx.971