Loading the document…
bankml0.2.7
Verified low-bit inference for the CPU you already have — the documentation, as of 0.2.7.GitHub · Hugging Face · why bankML · the thesis · minaiml (Space)
The milestone: native serving complete. Everything Savante and mindX ask of llama-server, answered by bankML's own engine, identical to llama-server b11192; every stage of the gate passed.
auto takes the 1-bit files tooscientific.diagnostic to 18 decimalsExact and precise first, then efficient and optimized. Port what works, and only what works; optimization, succinct, and verified response; the standing machine is the CPU you already have; consume the inference and prefer the better answer; know the limit and code around it; improve by increments with the oracle in the loop; use the machine at hand; prove the data, keep the data.
the Thesis and its sixteen contributions · the argument, with the field · hear it read
.memory keeps its versions (a private repository of its own, gitmind’s ledger, no backup sources), every stage and every oracle passed · gate record 0.4.3: browseML — bankML in the browser, verified and with a receipt, 9 / 9 answers identical to llama-server b11192 on one thread and on Web Workers, every WebAssembly kernel bit-exact (all 65,536 f16 values, about a million fused multiply-add lanes), every oracle passed · gate record 0.4.2: logprobs on Ollama’s API with llama-server b11192’s numbers (19 / 19), receipts that hash them, max_tokens: 0 as llama-server (7 / 7), a request whose client has gone stopped (6 / 6), ternary on the GPU bit-exact, every oracle passed · gate record 0.4.0, the milestone: native serving complete — 1-bit decode at least at llama-server’s speed (pinned 8B A/B, 3 / 3 rounds, answers identical), a q8_0 KV cache 6 / 6, logprobs 14 / 14, slots 19 / 19, sessions 14 / 14, every oracle passed · gate record 0.3.6: the penalties — repeat, frequency and presence, token-identical to llama-server on 56 / 56 answers on each of three models, live 85 / 85 · gate record 0.3.5: JSON schemas — llama-server’s grammar on each template (173 / 173 schemas × 3 templates against llama.cpp’s own code) and its answers token-identical on all five native models (28 / 28, 11 / 11, 56 / 56 × 3), the content rule matched on 30,063 / 30,063 texts per template · bankml create: mindX’s persona layer, made from mindXtrain’s merged output, token-identical end to end (27 / 27, two ways) · gate record 0.3.4: mindX’s own model natively — mindx-gen39 and SmolLM2-135M-Instruct (the Llama graph in F16) and Bonsai-1.7B (tied embeddings) token-identical to llama-server b11192: whole model 800 / 800 and 840 / 840 rows bit-exact, F16 products 552,268 / 552,268 elements bit-exact against ggml, seeded sampling 40 / 40 on each, conversations 9 / 9, JSON mode 23 / 23 · JSON mode — answers under llama-server’s own json_object grammar token-identical, greedy and seeded (23 / 23 answers on the 1-bit model, 13 / 13 on the ternary), and llama.cpp’s grammar sampler matched on 1,645 / 1,645 whole-vocabulary masks · a C API, libbankml, identical to serve --native · Savante answered by bankML’s own forward pass — whole conversations identical to llama-server on 9 / 9 turns (text, token counts, prompt-cache reuse) · own forward pass 1,064 / 1,064 rows bit-exact for the 1-bit and the ternary model · seeded sampling identical on 40 / 40 continuations · 8,188,239,872 ternary weights and 762 / 762 dot products bit-exact against llama.cpp b11192’s compiled library · one ternary token 0.222 s vs 2.139 sLoading the document…