topic: qwen
9 modules sharing the qwen name family (derived from the tree's prefix discipline).
The 'qwen' topic family in the Nishi sovereign ecosystem focuses on integrating and optimizing the Qwen model within the NishiLang framework, utilizing hybrid precision and modular architecture. Modules like nx_qwen_extract and nx_qwen_hybrid_attn enable extraction and attention processing of Qwen weights, while nx_qwen_hybrid_ffn implements the SwiGLU sub-layer with mixed-precision. These modules work together to support efficient, native execution of Qwen on the Nishi platform.
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_qwen_blk0_types.nx | dump the ggml_type + dims of every blk.0 weight of the real Q4_K_M Qwen. | 85 | 3 |
| nx_qwen_extract.nx | R5: the pretrained sovereign Qwen as the RESEARCHER'S EXTRACTION stage (importable by | 135 | 5 |
| nx_qwen_extract_gate.nx | unit test for nx_qwen_extract: build a synthetic qabench g[] (1 paragraph, 2 | 47 | 2 |
| nx_qwen_hybrid_attn.nx | the hybrid attention SUB-LAYER on real Qwen weights, gated vs the f32 path. | 210 | 3 |
| nx_qwen_hybrid_ffn.nx | the hybrid FFN (SwiGLU) sub-layer on real Qwen weights, gated vs the f32 path. | 202 | 3 |
| nx_qwen_hybrid_qkv.nx | the hybrid mixed-precision Q/K/V projection stage on real Qwen weights. | 129 | 3 |
| nx_qwen_native_gate.nx | NATIVE WINDOWS " Paris" forward. | 105 | 2 |
| nx_qwen_reader_bench_gate.nx | IN-PROCESS sovereign benchmark of the pretrained no-float Qwen2.5-0.5B as an | 419 | 11 |
| nx_qwen_wsl_timing_gate.nx | WSL2 forward on the DEFAULT threaded pool (the | 90 | 2 |