[package]
edition = "2024"
rust-version = "1.85"
name = "memra-server"
version = "0.82.0"
build = "build.rs"
autolib = false
autobins = false
autoexamples = false
autotests = false
autobenches = false
description = "OpenAI-compatible HTTP serving for the memra CUDA inference engine - single-GPU multi-model step-interleave scheduling on RTX 50-series"
readme = false
license = "MIT"
repository = "https://github.com/avifenesh/memra"
resolver = "2"
[package.metadata.binstall]
pkg-url = "{ repo }/releases/download/v{ version }/memra-server-v{ version }-{ target }.tar.gz"
bin-dir = "memra-server-v{ version }-{ target }/{ bin }"
pkg-fmt = "tgz"
[[bin]]
name = "memra-server"
path = "src/main.rs"
[dependencies.async-stream]
version = "0.3"
[dependencies.axum]
version = "0.7"
[dependencies.cudarc]
version = "0.19"
features = [
"std",
"cuda-13010",
"dynamic-loading",
"driver",
"cublas",
"cublaslt",
"nvrtc",
]
default-features = false
[dependencies.futures-core]
version = "0.3"
[dependencies.libc]
version = "0.2"
[dependencies.llguidance]
version = "1.7.6"
[dependencies.memra-engine]
version = "=0.82.0"
[dependencies.memra-gguf]
version = "=0.82.0"
[dependencies.memra-lanes]
version = "=0.82.0"
[dependencies.memra-tokenizer]
version = "=0.82.0"
[dependencies.serde]
version = "1"
features = ["derive"]
[dependencies.serde_json]
version = "1"
features = ["preserve_order"]
[dependencies.sha2]
version = "0.10"
[dependencies.tokio]
version = "1"
features = [
"rt-multi-thread",
"macros",
"net",
"sync",
"time",
"signal",
]
[dependencies.toml]
version = "0.8"
[dev-dependencies.tower]
version = "0.5"
features = ["util"]