Skip to content

fix TPU node pool scale to zero #75

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
divyashreepathihalli merged 3 commits into from
Mar 6, 2026
Merged

fix TPU node pool scale to zero #75

Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
eca05f0
fix TPU node pool scale to zero
divyashreepathihalli Mar 6, 2026
e8e5977
code reformat
divyashreepathihalli Mar 6, 2026
505a851
address gemini comment
divyashreepathihalli Mar 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 19 additions & 3 deletions keras_remote/backend/gke_client.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -437,22 +437,38 @@ def _check_node_pool_exists_cached(selector_items) -> bool:
pool_labels = config_dict.get("labels", {}).copy()

# Map GKE injected node labels for accelerators mapping
accelerators = config_dict.get("accelerators", [])
if accelerators:
accel_type = accelerators[0].get("acceleratorType", "")
accel_config_list = config_dict.get("accelerators", [])
if accel_config_list:
accel_type = accel_config_list[0].get("acceleratorType", "")
if accel_type.startswith("tpu-"):
pool_labels["cloud.google.com/gke-tpu-accelerator"] = accel_type
else:
pool_labels["cloud.google.com/gke-accelerator"] = accel_type

# TPU mapping fallback
machine_type = config_dict.get("machineType", "")

# Check resource labels for TPU type (common in v5e/v5litepod)
resource_labels = config_dict.get("resourceLabels", {})
if "goog-gke-accelerator-type" in resource_labels:
pool_labels["cloud.google.com/gke-tpu-accelerator"] = resource_labels[
"goog-gke-accelerator-type"
]

if machine_type.startswith("ct"):
# We roughly map TPU topology presence for preflight
pool_labels["cloud.google.com/gke-tpu-topology"] = selector.get(
"cloud.google.com/gke-tpu-topology", ""
)

# Infer accelerator count from machine type using registry
# This is robust because it uses the same source of truth as the Pod spec generation
for tpu_spec in accelerators.TPUS.values():
for chips, topo_spec in tpu_spec.topologies.items():
if topo_spec.machine_type == machine_type:
pool_labels["cloud.google.com/gke-accelerator-count"] = str(chips)
break

if all(pool_labels.get(k) == str(v) for k, v in selector.items()):
return True
return False
Expand Down
Loading