feat: RadixMLP

Cargo.toml

@@ -2,6 +2,7 @@
 members = [
     "backends",
     "backends/candle",
+    "backends/candle-bench",
     "backends/ort",
     "backends/core",
     "backends/python",
@@ -54,6 +55,7 @@ candle-index-select-cu =  { version = "0.0.1", features = ["cuda-11"], default-f
 candle-rotary = { version = "0.0.1" }
 candle-flash-attn-v1 = { version = "0.0.1" }
 half = { version = "2.3.1", features = ["num-traits"] }
+radix_mlp = { version = "0.0.2" }
 
 [patch.crates-io]
 cudarc = { git = "https://github.com/Narsil/cudarc" , rev = "8b4f18b4bcd5e4b1a9daf40abc3a2e27f83f06e9"}

backends/candle-bench/Cargo.toml

@@ -0,0 +1,29 @@
+[package]
+name = "candle-bench"
+version = "0.1.0"
+edition = "2021"
+publish = false
+
+[dependencies]
+text-embeddings-backend-candle = { path = "../candle" }
+text-embeddings-backend = { path = ".." }
+text-embeddings-backend-core = { path = "../core" }
+text-embeddings-core = { path = "../../core" }
+anyhow = { workspace = true }
+cudarc = { workspace = true, optional = true }
+hf-hub = { workspace = true , features = ["ureq"] }
+radix_mlp = { workspace = true }
+serde_json = "*"
+tracing = "*"
+
+[dev-dependencies]
+criterion = "0.5"
+
+[[bench]]
+name = "radix_mlp_benchmark"
+harness = false
+
+[features]
+metal = ["text-embeddings-backend-candle/metal"]
+cuda = ["text-embeddings-backend-candle/cuda", "text-embeddings-backend-candle/flash-attn", "dep:cudarc","cudarc?/dynamic-linking"]
+candle = ["text-embeddings-backend/candle"]

backends/candle-bench/Readme.md

@@ -0,0 +1,3 @@
+```
+cargo bench --manifest-path backends/candle-bench/Cargo.toml --bench radix_mlp_benchmark --features "candle,cuda"
+```