minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[model-gateway] Optimize HashRing construction to reduce heap allocations (#17575)

Browse Source
This commit is contained in:
Praneth Paruchuri
2026-01-27 22:46:40 -08:00
committed by GitHub
parent 897c35b457
commit 6f009961bb
1 changed files with 7 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
sgl-model-gateway/src/core/worker_registry.rs
+7 -2
View File
@@ -57,11 +57,16 @@ impl HashRing {
for worker in workers { for worker in workers {
// Create Arc<str> once per worker, share across all virtual nodes // Create Arc<str> once per worker, share across all virtual nodes
let url: Arc<str> = Arc::from(worker.url()); let url: Arc<str> = Arc::from(worker.url());
let url_bytes = url.as_bytes();
// Create multiple virtual nodes per worker // Create multiple virtual nodes per worker
for vnode in 0..VIRTUAL_NODES_PER_WORKER { for vnode in 0..VIRTUAL_NODES_PER_WORKER {
let vnode_key = format!("{}#{}", url, vnode); let mut hasher = blake3::Hasher::new();
let pos = Self::hash_position(&vnode_key); hasher.update(url_bytes);
hasher.update(b"#");
hasher.update(&(vnode as u64).to_le_bytes());
let hash = hasher.finalize();
let pos = u64::from_le_bytes(hash.as_bytes()[..8].try_into().unwrap());
entries.push((pos, Arc::clone(&url))); entries.push((pos, Arc::clone(&url)));
} }
} }