Skip to content

Hologram

A content-addressed, UOR-native tensor runtime — address identical computation once, reuse it everywhere.

Hologram compiles a tensor graph to a .holo archive and executes it through a content-addressed runtime: every value carries a UOR-ADDR κ-label, so identical computation is addressed once and reused (memoized, deduplicated, replayed) instead of recomputed. The same .holo archive runs on x86_64, WebAssembly (no_std), and ARM bare-metal.

The runtime is one content-addressed buffer pool. A value lives in a single aligned buffer; a slot binds to it by κ-label. Re-executing identical inputs rebinds rather than recomputes — a graph-level memo hit is O(1) in graph size — and constants are pinned for the session’s lifetime. Performance comes from content-addressing, not micro-optimization: redundant compute is eliminated by identity.

Where a function has a finite quantum domain it is materialized once as a lookup table — the compute-once form of the function, built bit-identically from the reference implementation:

result = LUT[encoded_input] // O(1), bit-identical to compute
  • f16 / bf16 transcendentals (Sigmoid, Tanh, GELU, SiLU, Exp, Erf): the 16-bit domain has 65536 points, so the activation is materialized once as a [u16; 65536] table (128 KB, L2-resident) — ~28× faster on bf16 GELU.
  • Byte (≤8-bit) domain: a 256-entry table.
  • Quantized inference: a Dequantize → activation chain densifies into a ≤256-entry table keyed on the quantized byte — ~27× faster.
  • f32 is computed (a 4 GB table is infeasible); reuse is structural, via the κ-label memo at the graph level.
FeatureWhat it does
Content-addressed executionEvery value carries a UOR-ADDR κ-label; identical computation is addressed once and replayed in O(1)
LUT materializationPure functions over a finite quantum domain become a static table, bit-identical to compute
FusionCompile-time desugar + algebraic elision, plus load-time content-addressed fusion (matmul epilogue, dequant→matmul, dequant→activation, broadcast→binary)
Cache-oblivious matmulBlocked SIMD kernel (AVX-512 → AVX2 → NEON → scalar) with compile-time panel-packed constant weights
Warm startThe constant-only cone is folded into the archive so the runtime cache is never cold
Cross-platformno_std + alloc library stack, a C ABI, and WASM bindings for embedded and browser targets

Eleven crates make up the workspace; depend on the individual crates you need — there is no umbrella crate.

CrateRole
hologram-hostPlatform/host bounds (register widths, capacities)
hologram-typesShared types: dtype tags, memory tiers
hologram-opsUOR-native op taxonomy, semantics, backward rules
hologram-graphTensor graph IR, desugaring, algebraic elision, scheduling
hologram-compilerGraph → .holo (lowering, fusion, workspace planning)
hologram-archive.holo binary format, UOR-ADDR κ-labels, BLAKE3 footer
hologram-backendKernel backends (CPU SIMD + LUT; optional wgpu / Metal)
hologram-execContent-addressed executor, buffer pool, warm-start
hologram-ffiC ABI bindings
hologram-clihologram compile / execute / inspect / bench
hologram-benchCriterion benchmark suites