nx_f32_llm_run_v3.nx
buildroot/runtime/nx_f32_llm_run_v3.nx
about
nx_f32_llm_run_v3.nx -- lazy-aware autoregressive runner.
Same API as nx_f32_llm_run_v2 (NxF32SamplerCfg + recent-tokens
ring + repetition penalty), but invokes nx_f32_llm_forward_v4
internally. Use with v4-loaded models (lazy Q4_K weights).
genealogy_id: standard_decode_loop + lazy_dispatch
lineage_id: substrate_f32_llm_run_v3_lazy
dependencies 12 imports · 10 importers
diagram shows first 10 each side; +2 more imports, +0 more importers in the complete lists below.
imports: nx_syscalls.nxnx_tier.nxnx_bpe.nxnx_f32.nxnx_f32_kv_cache.nxnx_f32_lazy_weight.nxnx_f32_llama_block_v4.nxnx_f32_llama_stack_v4.nxnx_f32_llm.nxnx_f32_llm_v4.nxnx_f32_sampler.nxnx_f32_llm_run_v2.nx
imported by: nx_coder_gen_f32.nxnx_f32_llm_bench_test.nxnx_f32_llm_cohere_test.nxnx_f32_llm_live_gen_test.nxnx_f32_llm_run_v3_test.nxnx_f32_llm_serve.nxnx_f32_qwen_capital_gate.nxnx_llm_france_probe.nxnx_qwen_native_gate.nxnx_qwen_wsl_timing_gate.nx
structs
| none |
consts
| 23 | const NX_FRN3_OK_BASE: nx_int = 0 |
| 24 | const NX_FRN3_ERR_NULL: nx_int = -1 |
| 25 | const NX_FRN3_ERR_BAD_DIM: nx_int = -2 |
| 26 | const NX_FRN3_ERR_FORWARD: nx_int = -3 |
| 30 | const NX_FRN3_PF_CHUNK: nx_int = 32 |
functions
| 32 | func _frn3_pick(cfg: *NxF32SamplerCfg, logits: *i64, vocab_size: nx_int, called by 1: nx_f32_llm_run_v3 calls 3: nx_f32_sampler_argmaxnx_f32_sampler_sample_top_pnx_f32_sampler_sample_top_k |
| 48 | func _frn3_apply_penalty(cfg: *NxF32SamplerCfg, logits: *i64, vocab_size: nx_int, |
| 57 | func _frn3_recent_push(recent: *i64, recent_cap: nx_int, called by 1: nx_f32_llm_run_v3 |
| 69 | func nx_f32_llm_run_v3(model: *NxF32LlamaModel, |