diff --git a/configs/basic-single-node/k8s.yaml b/configs/basic-single-node/k8s.yaml
index f7e2810f6..ed3d5d501 100644
--- a/configs/basic-single-node/k8s.yaml
+++ b/configs/basic-single-node/k8s.yaml
@@ -1,12 +1,16 @@
+# memory is in bytes, not a "32Gi" quantity string: the published bundle is
+# parsed straight into the backend's PhysicalResources (`memory: int`), and
+# only the SDK path converts suffixes — a string passes local `compute-config
+# create -f` and 422s on launch. 8Gi = 8589934592, 32Gi = 34359738368.
 head_node:
   required_resources:
     CPU: 4
-    memory: 8Gi
+    memory: 8589934592
   resources:
     CPU: 0
 worker_nodes:
 - name: cpu_worker
   required_resources:
     CPU: 8
-    memory: 32Gi
+    memory: 34359738368
   max_nodes: 1
diff --git a/configs/getting-started/k8s.yaml b/configs/getting-started/k8s.yaml
index d7e7aaa8b..b422dcacd 100644
--- a/configs/getting-started/k8s.yaml
+++ b/configs/getting-started/k8s.yaml
@@ -1,12 +1,16 @@
+# memory is in bytes, not a "32Gi" quantity string: the published bundle is
+# parsed straight into the backend's PhysicalResources (`memory: int`), and
+# only the SDK path converts suffixes — a string passes local `compute-config
+# create -f` and 422s on launch. 8Gi = 8589934592, 32Gi = 34359738368.
 head_node:
   required_resources:
     CPU: 4
-    memory: 8Gi
+    memory: 8589934592
   resources:
     CPU: 0
 worker_nodes:
 - name: cpu_worker
   required_resources:
     CPU: 8
-    memory: 32Gi
+    memory: 34359738368
   max_nodes: 2
diff --git a/configs/pytorch-fsdp/k8s.yaml b/configs/pytorch-fsdp/k8s.yaml
index 1b5649d18..965ce8d5c 100644
--- a/configs/pytorch-fsdp/k8s.yaml
+++ b/configs/pytorch-fsdp/k8s.yaml
@@ -1,14 +1,18 @@
+# memory is in bytes, not a "16Gi" quantity string: the published bundle is
+# parsed straight into the backend's PhysicalResources (`memory: int`), and
+# only the SDK path converts suffixes — a string passes local `compute-config
+# create -f` and 422s on launch. 8Gi = 8589934592, 16Gi = 17179869184.
 head_node:
   required_resources:
     CPU: 4
-    memory: 8Gi
+    memory: 8589934592
   resources:
     CPU: 0
 worker_nodes:
 - name: gpu_worker
   required_resources:
     CPU: 4
-    memory: 16Gi
+    memory: 17179869184
     GPU: 1
   required_labels:
     ray.io/accelerator-type: T4
diff --git a/scripts/hooks/validate-build-yaml.py b/scripts/hooks/validate-build-yaml.py
index ed3a93d15..2d709d410 100755
--- a/scripts/hooks/validate-build-yaml.py
+++ b/scripts/hooks/validate-build-yaml.py
@@ -463,6 +463,22 @@ def check_k8s_configs_declarative(entries: list[Entry]) -> list[str]:
                     f"read by the launch-time GPU validation — use "
                     f"`required_labels: {{ray.io/accelerator-type: ...}}`"
                 )
+            # memory is bytes, never a Kubernetes quantity string. `rayapp
+            # build` copies required_resources verbatim into the published
+            # bundle, and the console clone path parses that bundle into the
+            # backend's PhysicalResources, whose `memory` is an int — "8Gi"
+            # there is a 422. Only the SDK (`compute-config create -f`)
+            # converts suffixes, so a string survives local testing and fails
+            # at launch; nothing else in the pipeline catches it.
+            mem = rr.get("memory")
+            if mem is not None and (isinstance(mem, bool) or not isinstance(mem, int)):
+                errors.append(
+                    f"{path}: {loc}.required_resources.memory must be an "
+                    f"integer number of bytes, not {mem!r} — the published "
+                    f"bundle goes to the backend unconverted "
+                    f"(8Gi = 8589934592, 16Gi = 17179869184, "
+                    f"32Gi = 34359738368)"
+                )
             # Mirror the backend's check_gpu_accelerator_consistency: GPU
             # count and accelerator-type label come together (TPU labels are
             # paired with TPU fields backend-side instead).