nx_live_fire_gguf.nx source

1// nx_live_fire_gguf.nx -- live-fire end-to-end GGUF round-trip on real disk. 2// 3// This is NOT a synthetic in-memory smoke. It exercises: 4// 1. Build a Llama-shape GGUF in a buffer (3 top-level tensors) 5// 2. WRITE the buffer to /tmp/nx_live_fire.gguf via sys_write 6// 3. CLOSE the file 7// 4. READ it back via sys_read_file (re-opens, sys_read into mmap) 8// 5. Parse the header via nx_gguf_parse 9// 6. LAZY-load every tensor handle via nx_gguf_load_tensor_lazy 10// (Phase B of the conductor arc; zero-RAM tensor handles) 11// 7. MATERIALIZE on demand via nx_placed_tensor_get 12// 8. VERIFY the dequantized values bit-exactly match the source 13// we wrote 4 syscalls ago (proves disk round-trip + dequant 14// pipeline is end-to-end correct) 15// 9. MEASURE elapsed monotonic time via sys_clock_gettime_mono + 16// assert each phase is under generous budget 17// 18// Why this matters (live-fire vs synthetic-smoke discipline): 19// * Synthetic in-memory smokes (all 9 bricks shipped so far this 20// session) prove ALGORITHMIC correctness on hand-built buffers. 21// * THIS smoke proves the same primitives survive a kernel-mediated 22// file-system round-trip (write(2) -> close(2) -> openat(2) -> 23// read(2)) which is the actual deployment path for real-model 24// loading. 25// * If any of: file-descriptor leakage, buffered-write inversion, 26// mmap-vs-read alignment, byte-order on RV64 host -- bit-exact 27// comparison after disk round-trip catches it. 28// * Per the honest-perf-verdict cardinal: timing assertions in the 29// smoke are SIGNAL-ONLY, not the verdict. The verdict is 30// correctness; timing is recorded for trend visibility. 31// 32// Bits-up composition: 33// nx_syscalls.nx -- sys_openat_wr / sys_write / sys_close 34// / sys_read_file / sys_clock_gettime_mono 35// nx_le.nx -- header byte writers 36// nx_gguf.nx + nx_gguf_load.nx 37// -- parse + per-tensor dequant 38// nx_placement.nx -- lazy placement 39// nx_gguf_load_lazy.nx -- Phase B lazy loader 40// 41// genealogy_id: posix_file_round_trip + ggml_format_canon 42// lineage_id: substrate_live_fire_v1_disk_round_trip 43 44// nx_safety_envelope: 45// intended_use: "Live-fire end-to-end disk-backed GGUF 46// round-trip + lazy load + materialize + 47// bit-exact value verification + timing 48// budget assertion" 49// sil_target: SIL2 50// evidence: [real_file_system_io_exercised, 51// bit_exact_disk_round_trip_verified, 52// timing_budget_signal_only_per_honest_perf] 53// verdict: NOT_YET_EVALUATED 54 55import "nx_syscalls.nx" 56import "nx_tier.nx" 57import "nx_le.nx" 58import "nx_tensor.nx" 59import "nx_gguf.nx" 60import "nx_gguf_load.nx" 61import "nx_placement.nx" 62import "nx_gguf_load_lazy.nx" 63 64// Generous timing budgets (signal-only per the honest-perf cardinal). 65// On qemu-riscv64 these are pessimistic; on native they'll be much 66// faster. If we ever exceed them something has gone structurally 67// wrong (sync-write to network FS, swap thrashing, etc.). 68const NX_LF_WRITE_BUDGET_MS: i64 = 2000 69const NX_LF_READ_BUDGET_MS: i64 = 2000 70const NX_LF_PARSE_BUDGET_MS: i64 = 500 71const NX_LF_LAZY_LOAD_BUDGET_MS: i64 = 500 72const NX_LF_MATERIALIZE_BUDGET_MS: i64 = 500