start: renumber node ranks within the launch set — a subset launch (dgx3+dgx4 = global ranks 2/3 from 4nodes-switch) sent --node-rank 2 with --nnodes 2 and vLLM rejects it ('must be between 0 and 1'); dist flags/head selection/headless-ness now use the serve-local index; container names/labels follow the launch set (_0/_1)

src/docker/start.rs

@@ -42,6 +42,21 @@ pub(crate) async fn start_cluster(
         }
     };
     let nnodes = nodes.len() as u32;
+    // Renumber to zero-based ranks within the launch set: dist flags
+    // (`--node-rank`), head selection and headless-ness must use the
+    // serve-local index, not the node's global cluster rank. A subset launch
+    // (required_nodes=2 out of a 4-node cluster, dgx3+dgx4 = global ranks
+    // 2/3) otherwise sends `--node-rank 2` with `--nnodes 2` — vLLM rejects
+    // it ("must be between 0 and 1").
+    let launch_nodes: Vec<Node> = nodes
+        .iter()
+        .enumerate()
+        .map(|(i, n)| {
+            let mut n = (*n).clone();
+            n.rank = i as u32;
+            n
+        })
+        .collect();
     // A multi-node recipe without an explicit distributed backend would launch
     // N INDEPENDENT single-node engines (no dist flags, each loading the full
     // checkpoint) — the silent root cause of the glm53-flash-fp4-local-2x
@@ -84,8 +99,8 @@ pub(crate) async fn start_cluster(
     let resolved = command::resolve_extra_mounts(config, recipe)?;
     command::clear_extra_mounts(config, &nodes, &resolved, opts.clear, opts.dry_run).await?;
     let mut handles = Vec::new();
-    for node in &nodes {
-        let node = (*node).clone();
+    for node in &launch_nodes {
+        let node = node.clone();
         let config = config.clone();
         let rn = recipe_name.to_string();
         let dry = opts.dry_run;