topic: llm
25 modules sharing the llm name family (derived from the tree's prefix discipline).
The 'llm' topic family in the Nishi sovereign ecosystem focuses on ensuring the correctness, performance, and reliability of large language models within a modular, decentralized architecture. Modules like nx_llm_arch_census track architectural coverage and benchmarking, while nx_llm_argmax_probe validates pure-forward correctness without sampling, and nx_llm_capcheck assesses deep capabilities of specific models. These modules collectively support rigorous validation, optimization, and deployment of LLMs across the ecosystem.
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_llm_arch_census.nx | SOVEREIGN LLM-ARCHITECTURE COVERAGE CENSUS (July 2026 leaderboard). Grounded in | 75 | 6 |
| nx_llm_argmax_probe.nx | PURE-FORWARD correctness test. Bypasses run_v3's sampler + repetition-penalty: | 93 | 3 |
| nx_llm_batch_serve.nx | the SOVEREIGN CONTINUOUS-BATCHING LLM SEAT. | 288 | 10 |
| nx_llm_capcheck.nx | SOVEREIGN DEEP CAPABILITY CHECKER for the no-float Qwen2.5-0.5B (our own "Unsloth-style" | 150 | 5 |
| nx_llm_diff_probe.nx | NUMERICAL cross-check vs the transformers oracle (same GGUF). Dumps our forward's | 112 | 7 |
| nx_llm_forward_profile.nx | clean per-decode-forward timing on the | 105 | 4 |
| nx_llm_france_probe.nx | DECISIVE correctness probe (post QKV-bias fix). Feeds the REAL factual prompt | 75 | 3 |
| nx_llm_gap_registry.nx | where the team needs an LLM to INTERCEDE, marked so a future CUSTOM LLM | 66 | 6 |
| nx_llm_gap_registry_test.nx | the team records WHERE it needs an LLM (this session's real gaps), | 54 | 3 |
| nx_llm_landscape_research.nx | THIN structured research organ for the NEXT-STEPS decision: which small | 27 | 1 |
| nx_llm_layerlens_probe.nx | SOVEREIGN logit-lens bisection (no external oracle). Prefills "The capital of | 115 | 4 |
| nx_llm_loopb_gate.nx | THE LOOP-B CLOSURE: load OUR sovereignly-trained model (nx_f32_qwen2_train_gate's | 103 | 4 |
| nx_llm_ppl_bench_gate.nx | THE QUALITY BENCHMARK: perplexity + next-token accuracy of ANY gguf model on a | 142 | 9 |
| nx_llm_probe.nx | one-shot probe of the LIVE sovereign LLM seat (127.0.0.1:11434) to MEASURE whether the | 76 | 5 |
| nx_llm_provider.nx | a BACKUP to Claude for the LLM-rung work (operator: "figure out a plan to have | 86 | 9 |
| nx_llm_provider_local_test.nx | the identified LOCAL backup is preferred over paid APIs. With a | 36 | 3 |
| nx_llm_provider_test.nx | the team has a BACKUP to Claude for the LLM rung and prefers sovereignty. | 40 | 3 |
| nx_llm_run.nx | end-to-end LLM inference scaffold. | 277 | 3 |
| nx_llm_run_v2.nx | end-to-end Llama-class inference composer (real | 217 | 2 |
| nx_llm_run_v2_test.nx | end-to-end smoke for nx_llm_run_v2.nx. | 243 | 3 |
| nx_llm_sched.nx | sovereign CONTINUOUS-BATCHING scheduler (Orca / vLLM | 193 | 7 |
| nx_llm_sched_gate.nx | MEASURED gate for the CONTINUOUS-BATCHING | 228 | 5 |
| nx_llm_train_estimate.nx | 66 | 6 | |
| nx_llm_train_setup_gate.nx | 77 | 3 | |
| nx_llm_wired_expand_gate.nx | 95 | 5 |