router/Cargo.toml

[package]
name = "text-embeddings-router"
description = "Text Embedding Webserver"
build = "build.rs"
version.workspace = true
edition.workspace = true
authors.workspace = true
homepage.workspace = true

[lib]
path = "src/lib.rs"

[[bin]]
name = "text-embeddings-router"
path = "src/main.rs"

[dependencies]
anyhow = { workspace = true }
text-embeddings-backend = { path = "../backends", features = ["clap"] }
text-embeddings-core = { path = "../core" }
clap = { workspace = true }
futures = "^0.3"
init-tracing-opentelemetry = { version = "0.18.1", features = ["opentelemetry-otlp"] }
hf-hub = { workspace = true }
http = "1.0.0"
num_cpus = { workspace = true }
metrics = { workspace = true }
metrics-exporter-prometheus = { version = "0.15.1", features = [] }
opentelemetry = "0.23.0"
opentelemetry_sdk = { version = "0.23.0", features = ["rt-tokio"] }
opentelemetry-otlp = "0.16.0"
reqwest = { version = "0.12.5", features = [] }
simsimd = "4.4.0"
serde = { workspace = true }
serde_json = { workspace = true }
thiserror = { workspace = true }
tokenizers = { workspace = true }
tokio = { workspace = true }
tracing = { workspace = true }
tracing-opentelemetry = "0.24.0"
tracing-subscriber = { version = "0.3.16", features = ["json", "env-filter"] }
veil = "0.1.6"

# HTTP dependencies
axum = { version = "0.7.4", features = ["json"], optional = true }
axum-tracing-opentelemetry = { version = "0.18.1", optional = true }
base64 = { version = "0.22.1", optional = true }
tower-http = { version = "0.5.1", features = ["cors"], optional = true }
utoipa = { version = "4.2", features = ["axum_extras"], optional = true }
utoipa-swagger-ui = { version = "7.1", features = ["axum", "vendored"], optional = true }

# gRPC dependencies
async-stream = { version = "0.3.5", optional = true }
prost = { version = "0.12.1", optional = true }
tonic = { version = "0.11.0", optional = true }
tonic-health = { version = "0.11.0", optional = true }
tonic-reflection = { version = "0.11.0", optional = true }
tokio-stream = { version = "0.1.14", optional = true }

# Optional
cudarc = { workspace = true, optional = true }
intel-mkl-src = { workspace = true, optional = true }

# Malloc trim hack for linux
[target.'cfg(target_os = "linux")'.dependencies]
libc = "0.2.149"
# else use mimalloc
[target.'cfg(not(target_os = "linux"))'.dependencies]
mimalloc = { version = "*", features = ["no_thp"] }

[dev-dependencies]
insta = { git = "https://github.com/OlivierDehaene/insta", rev = "f4f98c0410b91fb5a28b10df98e4422955be9c2c", features = ["yaml"] }
is_close = "0.1.3"
reqwest = { version = "0.12.5", features = ["json"] }
serial_test = { workspace = true }

[build-dependencies]
vergen = { version = "8.0.0", features = ["build", "git", "gitcl"] }
tonic-build = { version = "0.11.0", optional = true }

[features]
default = ["candle", "http", "dynamic-linking"]
http = ["dep:axum", "dep:axum-tracing-opentelemetry", "dep:base64", "dep:tower-http", "dep:utoipa", "dep:utoipa-swagger-ui"]
grpc = ["metrics-exporter-prometheus/http-listener", "dep:prost", "dep:tonic", "dep:tonic-health", "dep:tonic-reflection", "dep:tonic-build", "dep:async-stream", "dep:tokio-stream"]
metal = ["text-embeddings-backend/metal"]
mkl = ["text-embeddings-backend/mkl"]
accelerate = ["text-embeddings-backend/accelerate"]
python = ["text-embeddings-backend/python"]
ort = ["text-embeddings-backend/ort"]
candle = ["text-embeddings-backend/candle"]
candle-cuda = ["candle", "text-embeddings-backend/flash-attn", "dep:cudarc"]
candle-cuda-turing = ["candle", "text-embeddings-backend/flash-attn-v1", "dep:cudarc"]
candle-cuda-volta = ["candle", "text-embeddings-backend/cuda", "dep:cudarc"]
static-linking = ["cudarc?/static-linking", "intel-mkl-src?/mkl-static-lp64-iomp"]
dynamic-linking = ["cudarc?/dynamic-linking", "intel-mkl-src?/mkl-dynamic-lp64-iomp"]
google = []