Files
ml/LLM.md

12 KiB
Raw Permalink Blame History

LLM.md — hanzoai/ml

Fast, multi-backend tensor & ML framework for Rust (CPU · CUDA · Metal · ROCm · Vulkan) with quantization (GGUF/GGML/AFQ/GPTQ/AWQ). The compute core beneath Hanzo inference.

Canonical role

  • This repo is the canonical implementation of the Hanzo Rust ML/tensor core. One impl, one place — discovery/wrapper repos link here, never copy the code.
  • Rust is the 2nd-most-complete ecosystem (Python → Rust → C++ → Go → …).
  • Crate hanzo-ml (crates.io) · docs at docs.rs/hanzo-ml.

Install / run

  • Core: cargo add hanzo-ml (Tensor/Device); add hanzo-nn to build models.
  • GPU: --features cuda (+ cudnn), or metal / rocm / vulkan.
  • Examples: cargo run --example quantized --release (see hanzo-ml-examples/).

Key entry points

  • hanzo-ml/ — core ops, devices, Tensor.
  • hanzo-nn/ — layers & model building.
  • hanzo-transformers/ — model implementations.
  • hanzo-kernels/, hanzo-flash-attn/ — CUDA kernels & FlashAttention v2.
  • hanzo-onnx/, hanzo-datasets/, hanzo-ml-wasm-examples/.
  • hanzo-research/ — the /v1/research client, the Rust peer of the Python hanzo-research. A record batch goes to its kind's route under its kind's key ({"runs": [...]}), the body hanzo-inc/cloud apps/research decodes; tests/record.rs is a stand-in that answers as those Go handlers do. Canonical JSON is RFC 8785 (numbers as ECMAScript writes the double, Ryu's digits), held to the Python package's bytes and WAL by tests/vectors/python.json and wal.jsonl (tests/vectors/gen.py regenerates them). serde_json's float_roundtrip is on so a WAL line reads back as the double that was hashed; it unifies into every consumer's build.
  • hanzo-train/ — training on hanzo-ml. cluster: one model across heterogeneous machines (Metal, CUDA, ROCm, CPU) by local SGD with an outer Nesterov step (DiLoCo). A coordinator owns the plan, θ_global in F32, the outer step and checkpoints; workers join any time, take batches one at a time, send θ_local − θ_global in bf16 with error feedback; deltas sum in join order, so a run equals its rounds replayed in one process (cluster::replay) bit for bit, and a checkpoint (outer momentum and carry, merged batches, round, every worker's AdamW state) resumes bit-identically. A worker that drops has its batches requeued; a round waits for no one past its deadline plus grace (a silent member is dropped, its link shut), a join under a member's name replaces it, links have keepalive and bounded reads/writes (net::link), and a worker whose link fails joins again on its own with backoff. A model plugs in behind cluster::Model (vars F32 masters, rate per parameter — None frozen, step over a batch of row indices the caller planned); the coordinator's owner behind cluster::Keep (round line, validation, save). adam: AdamW with per-parameter rates and open moments (hanzo-nn's optim::adamw update, fused on Metal). gpu: free-memory floor. dspark: the DSpark draft on the runtime (hanzo-train fit / join). Tests: cargo test -p hanzo-train (tiny classifier and synthetic DSpark cache, all CPU). Kai (hanzoai/decision train) is one user.

Backends and models

  • Vulkan dtypes. Storage is F32 or U32 (BF16 is held as F32, I64 as U32). U32 reductions run their own kernel (reduce_u32.comp; argmax/argmin take the first index on ties, as the CPU does) and U32 binary ops take the CPU fallback; any other storage type in a reduction is an error, never an F32 answer. cmp, affine, powf and unary kernels are F32 only. hanzo-ml/tests/vulkan_dtype_tests.rs holds each against the CPU.
  • Masked softmax on Vulkan. A row's running max starts at -inf (softmax_rows_blk.spv), so a row masked throughout is uniform, not NaN (softmax_masked_row_is_uniform). The same -3.4e38 start is still in the DSL (hanzo-kernel dag.rs Red::Max, attn.rs, flash.rs, quant.rs) and in sdpa_blk.spv, flash_attn_dsl_d128.spv and moe_route.spv: fix it in the DSL and regenerate, with a fully masked row in each test.
  • ModernBert past its trained positions. Config::yarn(s) gives a longer sequence YaRN rotary tables (NTK-by-parts, beta 32 and 1, attention factor 0.1 ln s + 1) on the global layers and extended frequencies on the local ones; within max_position_embeddings every pass reads the trained tables bit for bit.
  • ModernBert's trained rotary tables take their angles from the host. Each angle p · f is one F32 product on the host, its sine and cosine taken on the device in F32, then cast: the bits an exact F32 GEMM t · inv_freq gave on the CPU, ROCm, Metal and CUDA (f32_tables_are_the_exact_gemm_bits). The device GEMM it replaces ran on Vulkan's fp16 matrix cores, rounding positions past 2,048 (rows of 4,096 and 8,192 tokens read 25% and 37% from the CPU's states), and in BF16 rounded positions past 256 (to multiples of 2 from 257, more above) and the angles themselves well below 256, so every BF16 pass on every device read other positions than an F32 one: a CUDA BF16 pass read 5.7%, 25% and 36% from the CPU's F32 states at 256, 512 and 1,000 tokens with the old tables, 1.35%, 1.2% and 3.4% with these. trained_tables_hold_every_position_on_every_device holds every device's tables to the exact angles (--features rocm|vulkan|metal|cuda): F32 within 1e-3 (2e-3 on CUDA, whose kernels build with --use_fast_math), BF16 within 1e-2.
  • Past 2^31 elements on CUDA. A tensor of more than 2^31 elements miscomputes on CUDA with no error: the whole padded F32 pass at 16,384 tokens (12 heads, 3.2e9 scores) read 56% from the CPU's states. Tiles keep every score block far below it; a follow-up should refuse such a tensor rather than compute it.
  • Vulkan's pool trims. Device::trim gives Vulkan's idle buffers back too (VulkanDevice::trim, keeping 1 GiB, as ROCm's pool keeps): it flushes and waits, so between forwards, never in one. Its pool otherwise keeps up to 3/5 of the heap idle (30 GiB after a 32,768-token pass on evo).
  • ModernBert's long padded pass, a tile at a time. A padded pass whose row's scores pass LOGITS (2^26, 256 MiB in F32: a row past 2,364 tokens over 12 heads) attends a tile of queries at a time (Tiles): a global tile reads every key, at most 2^26 scores over its rows; a local tile holds 512 queries and the keys within their window. The keys' mask is [B, 1, 1, L] and each local tile's window [t, t + 2w] (tiles alike share one, built on the pass's device), so no [L, L] tensor exists; a shorter pass is the whole pass, bit for bit. Kᵀ and V are laid out once a layer (Vulkan's matmul copies a strided operand per call), and the raw scores drop once masked: two blocks live at once. (Dropping the masked scores too once the softmax is queued grew Metal's pool past 13 GiB at 16,384 tokens.) Each stage is a TRACE span (layer with at, qkv, rope, tile over scores, mask, softmax, values, then join, proj, mlp), which a subscriber can meter. tests/modernbert_memory.rs counts a 20,480-token pass's heap (no block of L x L bytes, less than one [L, L] F32 tensor held) and holds it to the packed pass within 5e-5 of the scale.
  • ModernBert's residual stream. ModernBert::stream(xs, mask, at) stops the pass forward runs at the input of layer at (Stream: packed or padded, as that pass lays it out), and ModernBert::finish runs it on: finish(stream(.., at)) is forward bit for bit at every at, packed and padded. Stack::new(vb, config, at) loads layers at.. and the final norm apart (a trainable copy of a top, for hanzo-inc/kai's routed yes/no reader), reading another pass's stream; Stream::rows and Stream::packed take a packed stream apart by sequence and put it back, the same bits.

Next: synthesis as a training loop

Capability training today is dataset → batches → optimizer → checkpoint, with curricula generated and scored by scripts outside it (hanzo-inc/kai's jailbreak round, 2026-10-03). Its pieces belong here, in this order:

  1. hanzo-train keeps per-example signals (loss, margin, prediction, gradient norm) instead of folding them into the batch loss, and reports balanced accuracy and MCC per slice.
  2. A dataset rows can join and leave mid-run, each row carrying its provenance and its factors (what a generated row varies), with weights a sampler reads.
  3. Contrast pairs as a training item: two rows and a margin on their score gap.
  4. Micro-fits with successive halving (hanzo-train sweep): many curricula at a tenth of a run, the best continued.
  5. Gradient embeddings over a capability's parameters only, to select the generated rows whose updates align with real errors. Generation itself (propose, mutate, verify against more than its own model) is a separate crate on top of these; a generator that only checks itself learns nothing new.

Releasing

  • Registry is crates.io, owner zeekay. There is no Hanzo cargo registry: no [registries] in .cargo/config.toml, no publish = [...] allow-list, and the sibling Rust repo (hanzoai/engine) publishes the same way.
  • Each crate names everything itself: no workspace inheritance. A crate's manifest holds its own version, package fields, lints and every dependency spec (a sibling as path plus the exact version it was published at); the root holds only members, exclude, resolver and profiles. What a crate ships is what its own manifest says.
  • Each crate carries its own version and moves by a patch bump from the version it last released. Crates change at different rates, so their numbers differ — that is information, not drift. Never renumber a crate to match another.
  • scripts/publish-order is the release set: every crate that does not say publish = false, topologically sorted. publish = false is the one way to keep a crate off the registry (examples, demos, the book, the PyPI extension module, and tensor-tools, whose crates.io name belongs to the upstream candle author).
  • Publishing is CI's job: .github/workflows/publish.yml, on a v* tag or a manual dispatch (gh workflow run publish.yml -R hanzoai/ml --ref main), walks that order with cargo publish --no-verify (GPU build scripts can't run on crates.io builders). It publishes a crate only when its manifest version is ahead of crates.io, so re-running is safe. The token is CARGO_REGISTRY_TOKEN at /@prod in org hanzo's KMS, read as the hanzo-ci identity (universe kms-grants.yaml).
  • A crate whose source changes takes a patch bump in the same commit. The ratchet only sees versions: a crate edited at an already-published version never ships, and its consumers pinning crates.io read older code than main (six crates had drifted, among them hanzo-rocm-kernels, before 027c26d9e). Consumers pin published versions, never a git rev (hanzo-inc/kai since 9b382b4).
  • Run a crate's own tests before its version moves. ROCm needs LD_LIBRARY_PATH=/opt/rocm/core-7.13/lib — libhiprtc.so.7 lives there, not in /opt/rocm/lib, so cargo test -p hanzo-kernel --features rocm otherwise dies at load time with the test binary already built.

Brand rules (enforce in all docs)

  • Hanzo is the Open AI Cloud / full AI SDK — never an "LLM gateway", never positioned vs LiteLLM, never an "OpenAI-compatible proxy". Purge that framing.
  • Paths are /v1/, never /api/.
  • Zen models are our own family — don't present upstream model names as ours.

Spec: ~/work/hanzo/SDK-ARCHITECTURE.md — the canonical one-way SDK model.

Local active inference

hanzo-active-inference/ owns finite-state Bayesian filtering, variational free energy, exact risk/ambiguity and information-gain EFE, finite open-loop policy posteriors and Dirichlet A/B learning. CPU f64 values, no tensor/backend dependency. cargo test -p hanzo-active-inference holds the mathematical identities, refusal boundaries and learning conservation. Engine hosts consume this crate, never copy its math. Its README states the limits: no contingent trees, parameter epistemic value or global loopy graph inference. It has no execution interface.