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.
| module | description | lines | funcs |
|---|---|---|---|
| nx_q4k_dot_simd.nx | SIMD-accelerated Q4_K fused dot (vpmaddwd), bit-exact vs nx_q4k_dot_row_col. | 165 | 3 |
| nx_q4k_dot_simd2_lib.nx | the fused Q4_K SIMD dot with the COMPILER-EMITTED unpack (2026-09-02, LM4c). | 105 | 1 |
| nx_q4k_dot_simd_lib.nx | the fused Q4_K SIMD dot as a LIBRARY (extracted 2026-09-02 from nx_q4k_dot_simd.nx, | 146 | 5 |
| nx_q4k_fused_fidelity_gate.nx | DOES THE FUSED INTEGER GEMM TELL THE TRUTH ON REAL TRAINED WEIGHTS? | 176 | 5 |
| nx_q4k_fused_vs_x4_gate.nx | THE HEAD-TO-HEAD the SOTA decision actually rests on: | 257 | 6 |
| nx_q4k_gemm_mt.nx | multi-threaded integer GEMM: wire nx_thread into the Q4_K integer dot. | 145 | 3 |
| nx_q4k_ggml_kat.nx | PERMANENT golden regression guard for Q4_K dequant. | 342 | 11 |
| nx_q4k_linear.nx | the integer GEMM linear: real Q4_K weights x Q10 activation -> Q10 output. | 121 | 3 |
| nx_q4k_linear_hp.nx | higher-precision integer linear (Q14 activation -> Q24 output). | 139 | 4 |
| nx_q4k_matmul.nx | fused-dequant Q4_K row × column dot product. | 140 | 3 |
| nx_q4k_matmul_bench.nx | fused vs materialize-then-dot bench. | 16 | 0 |
| nx_q4k_matmul_bench_test.nx | the fused-vs-materialized bench. | 188 | 3 |
| nx_q4k_matmul_mt_gate.nx | adversarial gate for the multi-threaded | 240 | 7 |
| nx_q4k_matmul_rate.nx | isolate the SSE Q4_K matmul rate (no 491MB model load) to pinpoint why the LM | 74 | 3 |
| nx_q4k_matmul_test.nx | smoke + bit-exact KAT for fused Q4_K dot. | 130 | 2 |
| nx_q4k_matmul_x4_gate.nx | adversarial gate for the PACKED-SIMD | 314 | 8 |
| nx_q4k_q24_gate.nx | prove the exact Q24 super-scale decode on the REAL Q4_K super-scales. | 28 | 1 |
| nx_q4k_real_gemm.nx | the CPU SPEED PATH proven on a REAL Qwen weight: integer fused-dequant-dot vs f32. | 131 | 2 |
| nx_q4k_simd2_gate.nx | the REFEREE for __q4k_unpack32s (the nx_cc intrinsic) and nx_q4k_dot_simd2 (2026-09-02, LM4c). | 203 | 7 |
| nx_q4k_speed_bench.nx | measure the CPU speedup of the integer fused-dequant-dot vs emulated-f32. | 121 | 2 |
| nx_q4k_to_f32.nx | bits-up Q4_K dequantization producing IEEE 754 | 297 | 6 |
| nx_q4k_to_f32_test.nx | smoke for nx_q4k_to_f32.nx. | 97 | 1 |