v0.7.1

Hologram v0.7.1

Installation

Via Git Dependency

Add to your Cargo.toml:

[dependencies]
# Depend on the specific crates you need (all share one version):
hologram-compiler = { git = "https://github.com/Hologram-Technologies/hologram", tag = "v0.7.1" }
hologram-exec     = { git = "https://github.com/Hologram-Technologies/hologram", tag = "v0.7.1" }
hologram-backend  = { git = "https://github.com/Hologram-Technologies/hologram", tag = "v0.7.1" }

CLI Binary

Download the appropriate binary for your platform from the assets below and add it to your PATH.

Changes

Commits since v0.7.0

• chore: bump version to v0.7.1 (f06cdef)
• Merge pull request #35 from Hologram-Technologies/perf/m1-gemv (0efbdf2)
• fix(sdk): moduleResolution "node" -> "bundler" for TypeScript 7 (1730424)
• perf(exec): keep the input-cache hit path on an immutable borrow (2e82863)
• perf: fix two perf-gate regressions (memo-hit output copy, fused-act epilogue) (763bcb5)
• fix(exec): drop breaking ExecError variant; refresh public-API snapshot (3abc203)
• perf(cpu): add x86-64 AVX2 lane to matmul_i8_per_channel (was scalar) (a61bc99)
• perf(cpu): fix fused matmul_dequant losing to the unfused path (432380f)
• perf(exec): eliminate per-walk/per-step allocations + O(L²) paging; fail-loud port dtype (04ec0f7)
• perf(exec): live tier-policy routing + LRU-bounded graph memo (b2dea87)
• docs(backend): correct the Clip/FusedSwiGlu kernel-guard comments (1c981c7)
• perf(cpu): reduce_float full-reduction SIMD + im2col/col2im/pool f32 fast paths (8f79cb7)
• perf(exec): fuse matmul epilogue for bf16/f16 (2-op forms), proven bit-identical (7ebf34f)
• bench(cpu): kernel_perf — throughput guard for the vectorized f32 paths (93dfc46)
• perf(cpu,exec): wasm elementwise lanes, RoPE + GroupNorm f32 paths, output slice-copy (665fc68)
• perf(cpu): vectorize f32 sigmoid/silu/tanh/gelu via deterministic exp (049a722)
• perf(cpu): vectorize norm/softmax/reduce/broadcast/attention f32 paths (ffd3493)
• perf(backend): vectorize the f32 RmsNorm/AddRmsNorm decode reductions (a1441bc)
• perf(backend): vectorize the x86 f32 packed-matmul row-remainder (M=1) (61d9b56)
• perf(backend): vectorize the x86 f32 matmul row-remainder (M=1 decode GEMV) (9dfd502)

What's Changed

• perf: M=1 decode GEMV fix + full CPU/exec vectorization & completeness sweep by @afflom in #35

Full Changelog: v0.7.0...v0.7.1