diff --git a/README.ko.md b/README.ko.md
index f3a3f92..c425156 100644
--- a/README.ko.md
+++ b/README.ko.md
@@ -266,7 +266,7 @@ dry-run은 적용 권한을 요구하지 않습니다. 이 모드는 조작 실
 
 lmd-top은 [ai-rm](https://github.com/ai-computing/ai-rm)의 **`l5-ops` 계층, kind `tool`** 구성요소이자 관측 콘솔이며,
 메타데이터는 [`ai-rm.yaml`](ai-rm.yaml)에 있습니다. 예전 rm_monitor의 역할은 나뉘었습니다. 모니터링 스택(Prometheus,
-exporter, Grafana 대시보드)은 [llm-d-setup](https://github.com/ai-computing/llm-d-setup) `monitoring/`(L4)으로, 부하 도구는
+exporter, Grafana 대시보드)은 [hetero-llm-d](https://github.com/ai-computing/hetero-llm-d) `monitoring/`(L4)으로, 부하 도구는
 rm-bench(L5)로 옮겼습니다. lmd-top은 그 스택이 모은 지표를 읽습니다. rm-core를 import하지 않고 rm-core 형식을 읽습니다.
 
 **ai-rm 패널 (protocol v0.1, 읽기 전용).** 클러스터에 ai-rm 소스가 하나라도 있으면 Infra ▸ Devices 아래에 ai-rm 패널이 나옵니다.
diff --git a/README.md b/README.md
index 832f639..3613713 100644
--- a/README.md
+++ b/README.md
@@ -276,7 +276,7 @@ project, and `:Routing policy` opens the configuration and rollout preview form.
 
 lmd-top is the **`l5-ops` layer, kind `tool`** of [ai-rm](https://github.com/ai-computing/ai-rm) and its observation
 console; metadata is in [`ai-rm.yaml`](ai-rm.yaml). The old rm_monitor role was split: the monitoring stack
-(Prometheus, exporters, Grafana dashboards) moved to [llm-d-setup](https://github.com/ai-computing/llm-d-setup)
+(Prometheus, exporters, Grafana dashboards) moved to [hetero-llm-d](https://github.com/ai-computing/hetero-llm-d)
 `monitoring/` (L4) and load generation to rm-bench (L5). lmd-top reads what that stack collects; it does not import
 rm-core, it reads rm-core formats.
 
diff --git a/src/airm.rs b/src/airm.rs
index 8996ad9..a72c168 100644
--- a/src/airm.rs
+++ b/src/airm.rs
@@ -192,13 +192,13 @@ pub struct Decision {
 }
 
 /// Parse ConfigMap `airm-placement`. Two layouts are read:
-/// - llm-d-setup's documented one: key `decisions.json` holding
+/// - hetero-llm-d's documented one: key `decisions.json` holding
 ///   `{protocol, generation, placement, decisions: [{pool, workload: {kind, ...}, acceleratorIds,
 ///   kvTransport, reason}]}`; each decision is shown as `<pool>/<kind>`.
 /// - one key per workload whose value is a `PlacementDecision`
 ///   `{accelerator_ids, kv_transport, reason}` (camelCase accepted too).
 ///
-/// An empty ConfigMap or an empty `decisions` list (the v0.1 llm-d-setup default) yields no
+/// An empty ConfigMap or an empty `decisions` list (the v0.1 hetero-llm-d default) yields no
 /// decisions. A value that does not parse is kept with the parse failure as its reason, so a bad
 /// write stays visible.
 pub fn parse_placement(cm: &serde_json::Value) -> Vec<Decision> {
diff --git a/src/app/train.rs b/src/app/train.rs
index 4103f1f..95481b2 100644
--- a/src/app/train.rs
+++ b/src/app/train.rs
@@ -1,6 +1,6 @@
 //! LoRA training Job — `--plan train`.
 //!
-//! The trainer is npu-lora (upstream: `llm-d-setup/npu-lora`), vendored into
+//! The trainer is npu-lora (upstream: ai-computing/npu-train, package `npu-lora`), vendored into
 //! `assets/npu-lora/` by its `scripts/vendor-into-lmd-top.sh` and embedded here, so this
 //! binary alone can produce a runnable Job: the official furiosa-llm image already carries
 //! furiosa-torch, torch and transformers, and the package rides along as a ConfigMap.
diff --git a/src/config.rs b/src/config.rs
index 67b4489..19d288d 100644
--- a/src/config.rs
+++ b/src/config.rs
@@ -48,7 +48,7 @@ impl Default for Config {
         let ns = s("LMD_NS", "namespace", "llm-serving");
         Config {
             prom: s("LMD_PROM", "prometheus", "localhost:9090"),
-            // llm-d-setup writes airm-placement into the llm-d namespace, so follow LMD_NS by default
+            // hetero-llm-d writes airm-placement into the llm-d namespace, so follow LMD_NS by default
             airm_ns: s("LMD_AIRM_NS", "airm_namespace", &ns),
             ns,
             grafana: s("LMD_GRAFANA", "grafana", "http://localhost:3000"),