[package]
name = "infer-inference"
version.workspace = true
edition.workspace = true
authors.workspace = true
license.workspace = true
description = "Transformer inference engine — forward pass, attention, FFN, layer norm from safetensors weights"
keywords = ["transformer", "inference", "llm", "interpretability"]
categories = ["science"]

[dependencies]
infer-models = { path = "../infer-models" }
infer-compute = { path = "../infer-compute" }
infer-core = { path = "../infer-core" }
infer-vindex = { path = "../infer-vindex" }
serde = { workspace = true }
serde_json = { workspace = true }
thiserror = { workspace = true }

# Model weights
safetensors = "0.5"
memmap2 = "0.9"

# System
libc = "0.2"

# Matrix ops (BLAS-accelerated)
ndarray = { version = "0.16", features = ["blas"] }

# Parallelism for the walk loop (per-feature work is embarrassingly parallel).
rayon = "1.10"

# Tokenizer
tokenizers = "0.21"

# Remote FFN backend (RemoteWalkBackend -> POST /v1/walk-ffn)
reqwest = { version = "0.12", features = ["blocking", "json"] }

# WASM expert registry
wasmtime = { version = "29", default-features = false, features = ["cranelift", "runtime"] }
wasmtime-wasi = "29"
anyhow = "1"


[target.'cfg(target_os = "linux")'.dependencies]
blas-src = { version = "0.10", features = ["openblas"], default-features = false }
openblas-src = { version = "0.10", features = ["system"] }

[target.'cfg(target_os = "macos")'.dependencies]
blas-src = { version = "0.10", features = ["accelerate"] }

[features]
default = []
metal = ["infer-compute/metal"]

[dev-dependencies]
assert_approx_eq = "1"
tempfile = "3"