Keeping the benchmark somewhere

backends/v3/Cargo.toml

@@ -36,7 +36,13 @@ serde_json = "1.0.107"
 slotmap = "1.0.7"
 thiserror = "1.0.48"
 tokenizers = { workspace = true }
-tokio = { version = "1.32.0", features = ["rt", "rt-multi-thread", "parking_lot", "signal", "sync"] }
+tokio = { version = "1.32.0", features = [
+  "rt",
+  "rt-multi-thread",
+  "parking_lot",
+  "signal",
+  "sync",
+] }
 tokio-stream = "0.1.14"
 tower-http = { version = "0.5.1", features = ["cors"] }
 tracing = "0.1.37"
@@ -44,7 +50,9 @@ tracing-opentelemetry = "0.21.0"
 tracing-subscriber = { version = "0.3.17", features = ["json", "env-filter"] }
 utoipa = { version = "4.2.0", features = ["axum_extras"] }
 utoipa-swagger-ui = { version = "6.0.0", features = ["axum"] }
-init-tracing-opentelemetry = { version = "0.14.1", features = ["opentelemetry-otlp"] }
+init-tracing-opentelemetry = { version = "0.14.1", features = [
+  "opentelemetry-otlp",
+] }
 minijinja = { version = "2.0.2" }
 minijinja-contrib = { version = "2.0.2", features = ["pycompat"] }
 futures-util = "0.3.30"
@@ -60,8 +68,16 @@ tower = "^0.4"
 tonic-build = "0.10.1"
 prost-build = "0.12.1"
 
+[dev-dependencies]
+criterion = "0.3"
+itertools = "0.13"
+
 [features]
 default = ["ngrok"]
 ngrok = ["text-generation-router/ngrok"]
 google = ["text-generation-router/google"]
 kserve = ["text-generation-router/kserve"]
+
+[[bench]]
+name = "prefix_cache"
+harness = false

backends/v3/benches/prefix_cache.rs

@@ -0,0 +1,45 @@
+use std::sync::Arc;
+
+use criterion::{black_box, criterion_group, criterion_main, Criterion};
+use itertools::Itertools;
+use rand::seq::SliceRandom;
+use rand::Rng;
+
+use text_generation_router_v3::block_allocator::{Allocator, RadixAllocator};
+
+fn prefix_cache_benchmark(c: &mut Criterion) {
+    let prefixes: Vec<Vec<u32>> = (0..8192)
+        .chunks(256)
+        .into_iter()
+        .map(|c| c.collect())
+        .collect();
+
+    let mut cache = RadixAllocator::new(1, 262144, None);
+
+    c.bench_function("fib 20", |b| {
+        b.iter_batched(
+            || {
+                //prefixes
+                //    .choose_multiple(&mut rand::thread_rng(), 5)
+                //    .fold(Vec::new(), |mut v, s| {
+                //        v.extend(s);
+                //        v
+                //    })
+
+                (0..7936)
+                    .map(|_| rand::thread_rng().gen_range(0..1024))
+                    .collect::<Vec<u32>>()
+            },
+            |prefill| {
+                let alloc = cache.allocate(prefill.len() as u32 + 13, Some(Arc::new(prefill)));
+                if let Some(alloc) = alloc {
+                    cache.free(alloc.0, alloc.3);
+                }
+            },
+            criterion::BatchSize::SmallInput,
+        );
+    });
+}
+
+criterion_group!(benches, prefix_cache_benchmark);
+criterion_main!(benches);

backends/v3/src/block_allocator.rs

@@ -204,7 +204,7 @@ impl Allocator for SimpleAllocator {
     }
 }
 
-struct RadixAllocator {
+pub struct RadixAllocator {
     allocation_id: u64,
 
     allocations: HashMap<u64, RadixAllocation>,

backends/v3/src/lib.rs

@@ -1,5 +1,5 @@
 mod backend;
-mod block_allocator;
+pub mod block_allocator;
 mod client;
 mod queue;
 mod radix;

router/Cargo.toml

@@ -28,7 +28,13 @@ serde = "1.0.188"
 serde_json = "1.0.107"
 thiserror = "1.0.48"
 tokenizers = { workspace = true }
-tokio = { version = "1.32.0", features = ["rt", "rt-multi-thread", "parking_lot", "signal", "sync"] }
+tokio = { version = "1.32.0", features = [
+  "rt",
+  "rt-multi-thread",
+  "parking_lot",
+  "signal",
+  "sync",
+] }
 tokio-stream = "0.1.14"
 tower-http = { version = "0.5.1", features = ["cors"] }
 tracing = "0.1.40"
@@ -37,7 +43,9 @@ tracing-subscriber = { version = "0.3.18", features = ["json", "env-filter"] }
 utoipa = { version = "4.2.0", features = ["axum_extras"] }
 utoipa-swagger-ui = { version = "6.0.0", features = ["axum"] }
 ngrok = { version = "0.13.1", features = ["axum"], optional = true }
-init-tracing-opentelemetry = { version = "0.14.1", features = ["opentelemetry-otlp"] }
+init-tracing-opentelemetry = { version = "0.14.1", features = [
+  "opentelemetry-otlp",
+] }
 minijinja = { version = "2.0.2" }
 minijinja-contrib = { version = "2.0.2", features = ["pycompat"] }
 futures-util = "0.3.30"
@@ -46,7 +54,11 @@ once_cell = "1.19.0"
 image = "0.25.1"
 base64 = { workspace = true }
 sysinfo = "0.30.13"
-uuid = { version = "1.9.1", default-features = false, features = ["v4", "fast-rng", "macro-diagnostics"] }
+uuid = { version = "1.9.1", default-features = false, features = [
+  "v4",
+  "fast-rng",
+  "macro-diagnostics",
+] }
 csv = "1.3.0"
 ureq = "=2.9"