lattice-inference 0.9.0

Pure Rust transformer inference engine — safetensors loading, SIMD matmul, BGE/Qwen3 embeddings
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

//! Forward-kernel module index for batch prefill, BitNet, CPU, f16 CPU, q8 CPU, GPU, Metal, and NEON paths.
pub mod batch_prefill;
pub mod bitnet_kernel;
pub mod cpu;
pub mod cpu_f16;
pub mod cpu_q8;
// Host-side f32 chunkwise GDN parity oracle (issue #175). Test/bench-only —
// adds zero production surface to the default build.
#[cfg(any(test, feature = "bench-internals"))]
pub mod gdn_chunk_ref;
pub mod gpu;
pub mod gpu_gemm;
pub mod metal;
pub mod metal_gemm;
pub mod metal_qwen35;
pub mod moe_expert_cache;
pub mod neon;
pub mod neon_forward;
// Every call site lives inside `metal_qwen35::inner`, gated the same way
// (`target_os = "macos"` + `metal-gpu`) — mirror that gate here instead of
// compiling the module everywhere and suppressing the resulting dead-code
// lint. Smaller surface, nothing to keep in sync with `mod inner`'s own gate.
#[cfg(all(target_os = "macos", feature = "metal-gpu"))]
pub(crate) mod signpost;

// Re-export from cpu for backward compat (was `layers`)
pub use self::cpu::*;