chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Super tiny remove unused select_worker_pair (#14609)

Browse Source
This commit is contained in:
fzyzcjy
2025-12-08 12:32:38 +08:00
committed by GitHub
parent 8fbf7dd56f
commit c08b780fe0
3 changed files with 0 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
sgl-model-gateway/src/policies/bucket.rs
View File

@@ -312,25 +312,6 @@ impl LoadBalancingPolicy for BucketPolicy {
workers.iter().position(|w| w.url() == prefill_url)
}
fn select_worker_pair(
&self,
prefill_workers: &[Arc<dyn Worker>],
decode_workers: &[Arc<dyn Worker>],
request_text: Option<&str>,
) -> Option<(usize, usize)> {
let prefill_idx = self.select_worker(prefill_workers, request_text)?;
let healthy_decode = get_healthy_worker_indices(decode_workers);
if healthy_decode.is_empty() {
return None;
}
let mut rng = rand::rng();
let decode_idx = rng.random_range(0..healthy_decode.len());
Some((prefill_idx, decode_idx))
}
fn name(&self) -> &'static str {
"bucket"
}

31
sgl-model-gateway/src/policies/cache_aware.rs
View File

@@ -347,37 +347,6 @@ impl LoadBalancingPolicy for CacheAwarePolicy {
}
}
fn select_worker_pair(
&self,
prefill_workers: &[Arc<dyn Worker>],
decode_workers: &[Arc<dyn Worker>],
request_text: Option<&str>,
) -> Option<(usize, usize)> {
// DEPRECATED: This method is no longer used when separate policies are configured.
// The PD router now uses separate policies for prefill and decode selection.
// This implementation remains for backward compatibility when a single policy is used.
// In PD mode with single policy:
// - Prefill: Use cache-aware routing for better cache utilization
// - Decode: Use least-load routing for better load distribution
// Select prefill worker using cache-aware logic
let prefill_idx = self.select_worker(prefill_workers, request_text)?;
// Select decode worker using least-load logic
let healthy_decode = get_healthy_worker_indices(decode_workers);
if healthy_decode.is_empty() {
return None;
}
let decode_idx = healthy_decode
.iter()
.min_by_key(|&&idx| decode_workers[idx].load())
.copied()?;
Some((prefill_idx, decode_idx))
}
fn on_request_complete(&self, worker_url: &str, success: bool) {
// Could track success rates per worker for more intelligent routing
if !success {

16
sgl-model-gateway/src/policies/mod.rs
View File

@@ -39,22 +39,6 @@ pub trait LoadBalancingPolicy: Send + Sync + Debug {
request_text: Option<&str>,
) -> Option<usize>;
/// Select a pair of workers (prefill and decode) for PD routing
///
/// Returns indices of (prefill_worker, decode_worker) from their respective arrays.
/// Default implementation uses select_worker for each array independently.
fn select_worker_pair(
&self,
prefill_workers: &[Arc<dyn Worker>],
decode_workers: &[Arc<dyn Worker>],
request_text: Option<&str>,
) -> Option<(usize, usize)> {
// Default implementation: independently select from each pool
let prefill_idx = self.select_worker(prefill_workers, request_text)?;
let decode_idx = self.select_worker(decode_workers, request_text)?;
Some((prefill_idx, decode_idx))
}
/// Update policy state after request completion
///
/// This is called when a request completes (successfully or not) to allow