code wiki / (root) / nx_f32_lazy_weight.nx

nx_f32_lazy_weight.nx

buildroot/runtime/nx_f32_lazy_weight.nx

41152 B886 linesdepth 9pulls 28 transitivereach 62 importersview sourcekind librarytopic f32

docs dependencies structs consts functions

about

nx_f32_lazy_weight.nx -- format-tagged weight tensor + matmul dispatcher. The foundation brick for lazy-Q4_K integration into the forward path. Current model bindings (NxF32LlamaLayer) hold raw *i64 pointers to eagerly-dequanted f32 storage. Lazy variants hold either f32 storage OR a (GGUF buffer + offset) pair for on-the- fly Q4_K dequant during matmul. A future v4 forward + binder will use NxF32LazyWeight throughout. This brick provides the type + dispatcher; integration with the per-layer struct + forward is a separate brick to keep scope tight. genealogy_id: tagged_union_dispatch + ggml_format_taxonomy lineage_id: substrate_f32_lazy_weight_v1

dependencies 9 imports · 55 importers

diagram shows first 10 each side; +0 more imports, +45 more importers in the complete lists below.

imports: nx_syscalls.nxnx_tier.nxnx_f32_matmul.nxnx_f32_matmul_t.nxnx_f32_q4k_matmul.nxnx_q5_0_to_f32.nxnx_q8_0_to_f32.nxnx_le.nxnx_pteam.nx

imported by: nx_batched_gate.nxnx_batched_tput.nxnx_batchscale_kat.nxnx_chat_warmcache.nxnx_coder_gen_f32.nxnx_companion_chat.nxnx_companion_voice.nxnx_f32_ffn_path_gate.nxnx_f32_lazy_weight_test.nxnx_f32_llama_block_v4.nxnx_f32_llama_block_v4_test.nxnx_f32_llama_layer_lazy_load.nxnx_f32_llama_layer_lazy_load_test.nxnx_f32_llama_stack_v4.nxnx_f32_llama_stack_v4_test.nxnx_f32_llama_v4b.nxnx_f32_llama_v4p.nxnx_f32_llm_bench_test.nxnx_f32_llm_chat_test.nxnx_f32_llm_cohere_test.nxnx_f32_llm_embed_diag_test.nxnx_f32_llm_live_gen_test.nxnx_f32_llm_live_load_only_test.nxnx_f32_llm_live_load_test.nxnx_f32_llm_probe.nxnx_f32_llm_run_v3.nxnx_f32_llm_run_v3_test.nxnx_f32_llm_serve.nxnx_f32_llm_v4.nxnx_f32_llm_v4_test.nxnx_f32_qwen_capital_gate.nxnx_llm_argmax_probe.nxnx_llm_batch_serve.nxnx_llm_diff_probe.nxnx_llm_forward_profile.nxnx_llm_france_probe.nxnx_llm_layerlens_probe.nxnx_llm_sched_gate.nxnx_lowrank_kv_real.nxnx_lw_cache_gate.nxnx_lwq24_probe.nxnx_natmm_bench.nxnx_paged_fwd_gate.nxnx_q5q8_ab.nxnx_q8_coldwarm.nxnx_q8_matmul_micro.nxnx_q8_mix_ab.nxnx_q8_scaling.nxnx_q8_st_bw.nxnx_qwen_native_gate.nxnx_qwen_wsl_timing_gate.nxnx_reason_paged_probe.nxnx_reason_probe.nxnx_reasoning_gate.nxnx_specdec_gate.nx

structs

43	struct NxF32LazyWeight

consts

26	const NX_LW_DTYPE_F32: nx_int = 0
27	const NX_LW_DTYPE_Q4_K: nx_int = 1
28	const NX_LW_DTYPE_Q5_0: nx_int = 2 // 2026-07-08: keep Q5_0 quantized (3MB vs 34.9MB F32), fused dequant-dot
29	const NX_LW_DTYPE_Q8_0: nx_int = 3 // 2026-07-08: keep Q8_0 quantized; SIMD __f32_i8dot32 dequant-dot = 10.2x (lm_head)
31	const NX_LW_OK: nx_int = 0
32	const NX_LW_ERR_BAD_TYPE: nx_int = 1
33	const NX_LW_ERR_NULL: nx_int = 2
34	const NX_LW_ERR_INNER: nx_int = 3 // inner matmul verdict != OK (e.g. Q4_K ERR_ALIGN -- was silently OK before 2026-07-07)
35	const NX_LW_N_VERDICTS: nx_int = 4
55	const NX_LW_BYTES: nx_int = 64 // 8 fields * 8
130	const NX_LW_NATIVE_POOL_W: i64 = 8
178	const NX_LW_CACHE_BUDGET_DEFAULT: i64 = 0 // 0 = stream quantized weights (re-measured 2026-08-01)
636	const NX_Q8_ST_MAC_MAX: i64 = 0

functions

37	func nx_lw_verdict_is_valid(v: nx_int) -> nx_int called by 1: main
57	func nx_f32_lazy_weight_new_f32(storage: i64, rows: nx_int, cols: nx_int) -> NxF32LazyWeight called by 5: main main main _flll_build_lazy _make_zero_lazy_layer calls 1: sys_mmap
70	func nx_f32_lazy_weight_new_q4k(buf: *u8, offset: i64, called by 5: main _flll_build_lazy nx_f32_llm_load_weights_v4_from_gguf lwc_run main calls 1: sys_mmap
88	func nx_f32_lazy_weight_new_q5_0(buf: *u8, offset: i64, called by 2: nx_f32_llm_load_weights_v4_from_gguf q5_weight calls 1: sys_mmap
102	func nx_f32_lazy_weight_new_q8_0(buf: *u8, offset: i64, called by 9: _flll_build_lazy nx_f32_llm_load_weights_v4_from_gguf main q8_weight cw_weight main+3 calls 1: sys_mmap
132	func nx_lw_shared_pool() -> *NxThreadPool called by 18: main main _lw_try_fill _lw_cached_matmul _lw_q5_0_matmul _lw_q8_0_matmul_pool_force+12 calls 1: nx_pool_is_native
183	func nx_lw_set_cache_budget(bytes: i64) -> i64 called by 2: main lwc_run
188	func nx_lw_cache_used() -> i64 called by 4: main main lwc_run main
192	func _lw_budget() -> i64 called by 1: _lw_try_fill
200	func _lw_fill_task(ctx_i: i64) -> i64 calls 1: nx_q4k_to_f32_packed
218	func _lw_fill_f32_task(ctx_i: i64) -> i64
244	func _lw_try_fill(W: *NxF32LazyWeight, k: nx_int, n: nx_int) -> i64 called by 1: nx_f32_lazy_matmul calls 7: _lw_budget sys_mmap nx_lw_shared_pool nx_pool_n_completed nx_pool_submit nx_pool_wait+1
307	func _lw_dot_x8(a: u8, b: u8, count: i64, acc: *u8) -> i64 called by 1: _lw_dot_task
330	func _lw_dot_task(ctx_i: i64) -> i64 calls 3: sys_mmap _lw_dot_x8 sys_munmap
359	func _lw_cached_matmul(W: NxF32LazyWeight, A: i64, C: *i64, called by 1: nx_f32_lazy_matmul calls 8: nx_lw_shared_pool sys_mmap _fq4m_pack_q10 nx_pool_n_completed nx_pool_submit nx_pool_wait+2
435	func _lw_q5_consts() -> *u8 called by 1: _lw_q5_0_task calls 1: sys_mmap
455	func _lw_q5_0_dot(qbuf: u8, qoff: i64, Apk: u8, k: i64, i8scr: u8, consts: u8) -> i64 called by 1: _lw_q5_0_task calls 2: nx_f16_to_f32 nx_le_read_u16
472	func _lw_q5_0_task(ctx_i: i64) -> i64 calls 4: sys_mmap _lw_q5_consts _lw_q5_0_dot sys_munmap
494	func _lw_q5_0_matmul(W: NxF32LazyWeight, A: i64, C: *i64, called by 1: nx_f32_lazy_matmul calls 7: nx_lw_shared_pool sys_mmap _fq4m_pack_a nx_pool_n_completed nx_pool_submit nx_pool_wait+1
553	func _lw_q8_0_dot(qbuf: u8, qoff: i64, Apk: u8, k: i64) -> i64 called by 3: _lw_q8_0_task _lw_q8_0_matmul_st _lw_q8_0_pteam_band calls 2: nx_f16_to_f32 nx_le_read_u16
569	func _lw_q8_0_task(ctx_i: i64) -> i64 calls 1: _lw_q8_0_dot
599	func _q8st_apk(need: i64) -> *u8 called by 3: _lw_q8_0_matmul_st _lw_q8_0_matmul_pool_force _lw_q8_0_matmul_pteam calls 1: sys_mmap
607	func _lw_q8_0_matmul_st(W: NxF32LazyWeight, A: i64, C: *i64, called by 6: _lw_q8_0_matmul main main main main main calls 3: _q8st_apk _fq4m_pack_a _lw_q8_0_dot
647	func _q8mt_ctxs(need: i64) -> *u8 called by 2: _lw_q8_0_matmul_pool_force _lw_q8_0_matmul_pteam calls 1: sys_mmap
652	func _lw_q8_0_matmul_pool_force(W: NxF32LazyWeight, A: i64, C: *i64, called by 4: _lw_q8_0_matmul main main main calls 7: nx_lw_shared_pool _q8st_apk _fq4m_pack_a _q8mt_ctxs nx_pool_n_completed nx_pool_submit+1
692	func _q8_pteam() -> *NxPTeam called by 1: _lw_q8_0_matmul_pteam calls 1: nx_pteam_new
696	func _lw_q8_0_pteam_band(band: i64, ctx: i64) -> i64 calls 1: _lw_q8_0_dot
718	func _lw_q8_0_matmul_pteam(W: NxF32LazyWeight, A: i64, C: *i64, called by 3: _lw_q8_0_matmul main main calls 5: _q8_pteam _q8st_apk _fq4m_pack_a _q8mt_ctxs nx_pteam_run
737	func _lw_q8_0_matmul(W: NxF32LazyWeight, A: i64, C: *i64, called by 2: nx_f32_lazy_matmul main calls 4: nx_pool_is_native _lw_q8_0_matmul_pteam _lw_q8_0_matmul_st _lw_q8_0_matmul_pool_force
756	func nx_f32_lazy_matmul(A: i64, W: NxF32LazyWeight, C: *i64, called by 13: main main nx_f32_llama_block_forward_v4 nx_f32_llama_block_forward_v4b nx_f32_llm_forward_v4b nx_f32_llama_block_forward_v4p+7 calls 7: nx_f32_matmul_t_pool nx_lw_shared_pool _lw_try_fill _lw_cached_matmul nx_f32_q4k_matmul_pool_fused _lw_q5_0_matmul+1