소스 검색

fix: drop_caches after image sync (clear docker load page cache)

Alexander Korolev 3 달 전
부모
커밋
126fe81115
3개의 변경된 파일과 15개의 추가작업 그리고 13개의 파일을 삭제
  1. 1 1
      recipes/glm52-nf3-hybrid-context/recipe.toml
  2. 13 11
      src/docker/mod.rs
  3. 1 1
      src/docker/sync.rs

+ 1 - 1
recipes/glm52-nf3-hybrid-context/recipe.toml

@@ -62,7 +62,7 @@ args = [
     "--max-num-seqs", "4",
     "--max-num-batched-tokens", "4096",
     "--max-cudagraph-capture-size", "32",
-    "--gpu-memory-utilization", "0.92",
+    "--gpu-memory-utilization", "0.88",
     # Engine backends
     "--attention-backend", "B12X_MLA_SPARSE",
     "--moe-backend", "b12x",

+ 13 - 11
src/docker/mod.rs

@@ -82,17 +82,6 @@ pub async fn serve(
     )
     .await?;
 
-    // Drop page cache on all nodes before starting.
-    if config.drop_caches && !opts.dry_run {
-        info!("Dropping page cache on all nodes...");
-        for node in &cluster.nodes {
-            let cmd = "echo 3 | sudo tee /proc/sys/vm/drop_caches".to_string();
-            if let Err(e) = run_ssh(config, node, cmd, false).await {
-                warn!("drop_caches on {} failed: {}", node.name, e);
-            }
-        }
-    }
-
     use self::sync::sync_image_to_nodes;
 
     // Build image on specified node if not found on any cluster node.
@@ -168,6 +157,19 @@ pub async fn serve(
         )
         .await?;
     }
+
+    // Drop page cache on all nodes right before starting containers.
+    // Must run after image sync (docker load fills page cache).
+    if config.drop_caches && !opts.dry_run {
+        info!("Dropping page cache on all nodes...");
+        for node in &cluster.nodes {
+            let cmd = "echo 3 | sudo tee /proc/sys/vm/drop_caches".to_string();
+            if let Err(e) = run_ssh(config, node, cmd, false).await {
+                warn!("drop_caches on {} failed: {}", node.name, e);
+            }
+        }
+    }
+
     let cache_dir = config.resolved_hf_cache_dir();
     start_cluster(
         config,

+ 1 - 1
src/docker/sync.rs

@@ -65,7 +65,7 @@ pub async fn sync_image_to_nodes(
             created[b]
                 .as_deref()
                 .unwrap_or("")
-                .cmp(&created[a].as_deref().unwrap_or(""))
+                .cmp(created[a].as_deref().unwrap_or(""))
         });
         *indexed
             .first()