[package]
name = "infer-vindex"
version.workspace = true
edition.workspace = true
authors.workspace = true
license.workspace = true
description = "Vindex — the queryable model format. Storage, KNN index, build, load, mutate."
keywords = ["transformer", "vindex", "knn", "interpretability"]
categories = ["science"]

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

# Matrix types (all compute via infer-compute)
ndarray = "0.16"

# Parallelism for the direct-Q4K matmul kernel (feature walk).
rayon = "1.10"

# Tokenizer (for embeddings loading + token resolution)
tokenizers = "0.21"

# Per-expert SVD summarization for MoE models
rand = "0.8"

# Checksums
sha2 = "0.10"

# base64 encoding for the HF preupload file-sample field.
base64 = "0.22"

# Model weights (for safetensors loading during extract)
safetensors = "0.5"
memmap2 = "0.9"

# OS-level mmap hints (madvise)
[target.'cfg(unix)'.dependencies]
libc = "0.2"

# HuggingFace Hub integration
hf-hub = "0.5"
reqwest = { version = "0.12", features = ["blocking", "json"] }

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

[dev-dependencies]
criterion = "0.5"

[[bench]]
name = "vindex_ops"
harness = false

[[bench]]
name = "vindex_scaling"
harness = false

[[bench]]
name = "memit_solve"
harness = false

[[bench]]
name = "extract_throughput"
harness = false

[[bench]]
name = "q4k_vs_f32"
harness = false