konduktor-nightly 0.1.0.dev20250919104536__py3-none-any.whl → 0.1.0.dev20250921104307__py3-none-any.whl

konduktor/templates/deployment.yaml.j2

@@ -12,6 +12,10 @@ metadata:
     {{ deployment_num_accelerators_label }}: "{{ num_accelerators }}"
     trainy.ai/has-autoscaler: "{{ autoscaler }}"
     trainy.ai/konduktor-managed: "true"
+    {% if autoscaler == 'true' %}
+    trainy.ai/original-min-replicas: "{{ min_replicas }}"
+    trainy.ai/original-max-replicas: "{{ max_replicas }}"
+    {% endif %}
   name: {{ name }}
   namespace: default
 spec:
@@ -61,18 +65,15 @@ spec:
     {{ model_name_label }}: {{ name }}
     {% endif %}
     {{ deployment_name_label }}: "{{ name }}"
-  {% if general %}
-  type: LoadBalancer
-  {% else %}
   type: ClusterIP
-  {% endif %}
 
-{% if not general %}
+# AIBRIX PODAUTOSCALER STUFF (KPA)
+{% if not general and autoscaler == 'true' %}
 ---
 apiVersion: autoscaling.aibrix.ai/v1alpha1
 kind: PodAutoscaler
 metadata:
-  name: {{ name }}-apa
+  name: {{ name }}-pa
   namespace: default
   labels:
     {{ model_name_label }}: {{ name }}
@@ -80,63 +81,109 @@ metadata:
     app.kubernetes.io/managed-by: kustomize
     {{ deployment_name_label }}: "{{ name }}"
     {{ deployment_user_label }}: "{{ user }}"
-  annotations:
-    autoscaling.aibrix.ai/up-fluctuation-tolerance: '0.1'
-    autoscaling.aibrix.ai/down-fluctuation-tolerance: '0.2'
-    apa.autoscaling.aibrix.ai/window: 30s
 spec:
-  scalingStrategy: APA
+  scalingStrategy: KPA
   minReplicas: {{ min_replicas }}
   maxReplicas: {{ max_replicas }}
   metricsSources:
-    - metricSourceType: pod
+    - metricSourceType: domain
       protocolType: http
-      port: "{{ ports }}"
-      path: metrics
-      targetMetric: gpu_cache_usage_perc
-      targetValue: '0.5'
+      endpoint: aibrix-activator.aibrix-activator.svc.cluster.local:8080
+      path: /metrics/default/{{ name }}
+      targetMetric: vllm:deployment_replicas
+      targetValue: "1"
   scaleTargetRef:
     apiVersion: apps/v1
     kind: Deployment
     name: {{ name }}
 {% endif %}
 
+# KEDA HTTP ADD-ON STUFF (1 per deployment)
 {% if general %}
+{% if autoscaler == 'true' %}
+# HTTPScaledObject (1 per deployment) - only when autoscaling enabled
 ---
-apiVersion: autoscaling/v2
-kind: HorizontalPodAutoscaler
+apiVersion: http.keda.sh/v1alpha1
+kind: HTTPScaledObject
 metadata:
-  name: {{ name }}-hpa
+  name: {{ name }}-httpscaledobject
   namespace: default
   labels:
     {{ deployment_name_label }}: "{{ name }}"
     {{ deployment_user_label }}: "{{ user }}"
-    trainy.ai/has-autoscaler: "{{ autoscaler }}"
 spec:
+  hosts:
+    - {{ name }}
+  pathPrefixes:
+    - "/"
+    {% if probe_path %}
+    - "{{ probe_path }}"
+    {% endif %}
   scaleTargetRef:
-    apiVersion: apps/v1
+    name: "{{ name }}"
     kind: Deployment
-    name: {{ name }}
-  minReplicas: {{ min_replicas }}
-  maxReplicas: {{ max_replicas }}
-  metrics:
-    - type: Resource
-      resource:
-        name: cpu
-        target:
-          type: Utilization
-          averageUtilization: 50
-  behavior:
-    scaleDown:
-      stabilizationWindowSeconds: 60
-      policies:
-      - type: Percent
-        value: 100
-        periodSeconds: 15
-    scaleUp:
-      stabilizationWindowSeconds: 20
-      policies:
-      - type: Percent
-        value: 100
-        periodSeconds: 15
+    apiVersion: apps/v1
+    service: "{{ name }}"
+    port: {{ ports }}
+  replicas:
+    min: {{ min_replicas }}
+    max: {{ max_replicas }}
+  scaledownPeriod: 1200 # 20 minutes
+  scalingMetric:
+    requestRate:
+      targetValue: 4
+      granularity: "1s"
+      window: "30s"
+{% endif %}
+
+# INGRESS (1 per deployment)
+---
+apiVersion: networking.k8s.io/v1
+kind: Ingress
+metadata:
+  name: {{ name }}-ingress
+  annotations:
+    nginx.ingress.kubernetes.io/use-regex: "true"
+    nginx.ingress.kubernetes.io/rewrite-target: /$1
+    {% if autoscaler == 'true' %}
+    nginx.ingress.kubernetes.io/upstream-vhost: "{{ name }}"
+    {% endif %}
+spec:
+  ingressClassName: nginx
+  rules:
+  - host: {{ general_base_host }}
+    http:
+      paths:
+      - path: /{{ name }}(.*)
+        pathType: ImplementationSpecific
+        backend:
+          service:
+            {% if autoscaler == 'true' %}
+            # Use KEDA interceptor for autoscaling
+            name: keda-proxy
+            port:
+              number: 8080
+            {% else %}
+            # Direct to app service for fixed replicas
+            name: {{ name }}
+            port:
+              number: {{ ports }}
+            {% endif %}
+  # Direct access convenience rule (via LB IP + Host: {{ name }})
+  - host: {{ name }}
+    http:
+      paths:
+      - path: /(.*)
+        pathType: ImplementationSpecific
+        backend:
+          service:
+            {% if autoscaler == 'true' %}
+            name: keda-proxy
+            port:
+              number: 8080
+            {% else %}
+            name: {{ name }}
+            port:
+              number: {{ ports }}
+            {% endif %}
 {% endif %}

konduktor/utils/schemas.py

@@ -389,7 +389,7 @@ def get_serving_schema():
         'properties': {
             'min_replicas': {
                 'type': 'integer',
-                'minimum': 1,
+                'minimum': 0,
                 'description': 'Minimum number of replicas for autoscaling.',
             },
             'max_replicas': {

{konduktor_nightly-0.1.0.dev20250919104536.dist-info → konduktor_nightly-0.1.0.dev20250921104307.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: konduktor-nightly
-Version: 0.1.0.dev20250919104536
+Version: 0.1.0.dev20250921104307
 Summary: GPU Cluster Health Management
 Author: Andrew Aikawa
 Author-email: asai@berkeley.edu

{konduktor_nightly-0.1.0.dev20250919104536.dist-info → konduktor_nightly-0.1.0.dev20250921104307.dist-info}/RECORD

@@ -1,4 +1,4 @@
-konduktor/__init__.py,sha256=jxyJKjEi7MB605Q6QUD49c52B7T2f4EpMyNrkDcMnao,1574
+konduktor/__init__.py,sha256=i-3wiQO7jODykrQYHnkmj7qoFXpZYAp1NRSWS2HzVA4,1574
 konduktor/adaptors/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 konduktor/adaptors/aws.py,sha256=s47Ra-GaqCQibzVfmD0pmwEWHif1EGO5opMbwkLxTCU,8244
 konduktor/adaptors/common.py,sha256=ZIqzjx77PIHUwpjfAQ1uX8B2aX78YMuGj4Bppd-MdyM,4183
@@ -7,13 +7,13 @@ konduktor/authentication.py,sha256=_mVy3eqoKohicHostFiGwG1-2ybxP-l7ouofQ0LRlCY,4
 konduktor/backends/__init__.py,sha256=usWJ8HdZJEyg7MIsN8Zcz9rk9e2Lq5dWJ8dv6hCN3ys,199
 konduktor/backends/backend.py,sha256=qh0bp94lzoTYZkzyQv2-CVrB5l91FkG2vclXg24UFC0,2910
 konduktor/backends/constants.py,sha256=nt9G9AmFCOMwO4GuKgRQSzJJuKapOmaROp4_Y0tMF5A,732
-konduktor/backends/deployment.py,sha256=fswN9hX_7NwcEogYmo1xn3WgWF8XCcGDvV5yx54_CA0,5860
-konduktor/backends/deployment_utils.py,sha256=qcuoLPeMvEVqgD_h71hQZXAp4ZCdXsFeSBxhtXW6pAA,39846
+konduktor/backends/deployment.py,sha256=VXz0GMvFzqq8kd1-pXaNaTsEtSyMmVO2OSd-_-T9nIE,6171
+konduktor/backends/deployment_utils.py,sha256=rfVpiM0KlXma8SxA029TczrpbY-gRG3T0pj2eeARkE8,49257
 konduktor/backends/jobset.py,sha256=drt8Gc0iYQx18JWXBU6XfhUvC2xCKd8szSJ2JC4O20Q,8640
 konduktor/backends/jobset_utils.py,sha256=If4pv5peB_yXrJJwjkySgVzbjcxDEDWfkOQxUkwSlOk,26386
 konduktor/backends/pod_utils.py,sha256=KP_PAgsdNHFgt4Od-5gAtpifAKIL7DMBg7NJ44uqikg,14885
 konduktor/check.py,sha256=JennyWoaqSKhdyfUldd266KwVXTPJpcYQa4EED4a_BA,7569
-konduktor/cli.py,sha256=OwVdT4ibAQoAJO79YzlThQv_VKlHpsD-CHRwOzehGQ8,57613
+konduktor/cli.py,sha256=HOL_O7BpYCl7uF5mKkMvXLjFNMTeAh2hYkraaJ6zi3w,57674
 konduktor/config.py,sha256=9upqgCCYvcu6fKw7tovEYC1MWTkAAir0_WHPdayylbI,15536
 konduktor/constants.py,sha256=T3AeXXxuQHINW_bAWyztvDeS8r4g8kXBGIwIq13cys0,1814
 konduktor/controller/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -68,17 +68,17 @@ konduktor/data/storage_utils.py,sha256=n4GivkN0KMqmyOTDznF0Z-hzsJvm7KCEh5i5HgFAT
 konduktor/execution.py,sha256=d0EP79iSrW2uFsoqn0YV_4kgIupPIqpMOParXx0y3kg,18519
 konduktor/kube_client.py,sha256=HtM3d-_GigHnfGINRANchApR9_OigqczBgeYJ6Dj4j0,8504
 konduktor/logging.py,sha256=xtcCdnecmC3rqMTyunK-klQRINojI7NI4Apag78i9jM,3221
-konduktor/manifests/apoxy-setup.yaml,sha256=HDZu7Evm_siIpK1E4tNZ9WVTnFV2LhBXwjJlFOYSTcU,3319
-konduktor/manifests/apoxy-setup2.yaml,sha256=BhXsgcVrLBruLXnF7xlj0Ej6YVJFYMABJIpYtwakQMo,731
+konduktor/manifests/aibrix-setup.yaml,sha256=cUi7ToRRasRt2bXq53htV4QoCDEcxp2pvnCi-esJ4yk,9109
+konduktor/manifests/apoxy-setup.yaml,sha256=EipknCq33aBdxu9BIo6y5novjO0B_d_DCWqY44zYNuU,4262
+konduktor/manifests/apoxy-setup2.yaml,sha256=Benmg-io8audt_vi5T_T7vK8MYnaTnpX-bNDgRXM5t0,2512
 konduktor/manifests/controller_deployment.yaml,sha256=6p3oSLkEVONZsvKZGqVop0Dhn4bo3lrigRmhf8NXBHE,1730
 konduktor/manifests/dashboard_deployment.yaml,sha256=xJLd4FbPMAosI0fIv5_8y7dV9bw0Vsf81l-w4MB_aU8,2837
 konduktor/manifests/dmesg_daemonset.yaml,sha256=pSWt7YOeTYjS0l0iki1fvHOs7MhY-sH-RQfVW6JJyno,1391
 konduktor/manifests/pod_cleanup_controller.yaml,sha256=hziL1Ka1kCAEL9R7Tjvpb80iw1vcq9_3gwHCu75Bi0A,3939
 konduktor/resource.py,sha256=kfdhnUR_9kDLSm2sUAkv1sLQXyAkI08p3wIzCz1p7-M,20791
-konduktor/serving.py,sha256=sh8TPAUXg23Bkt0ByatIMdxFFqzRm18HJTEkt3wHzdo,5147
-konduktor/task.py,sha256=oFRHdMevg7lGYkHugCHl89FUREfq9M-l0Qd3N-rjHMA,37727
-konduktor/templates/apoxy-deployment.yaml.j2,sha256=_EdT7w0rBK3if1INHT6GGUEugy0mOkRfYOWRgBcKLdo,942
-konduktor/templates/deployment.yaml.j2,sha256=uXFjDQaimbpFdAn2RJGaIvS_PzDY136cw_L3QMjz3ZA,3452
+konduktor/serving.py,sha256=4s8cQhsVjf-HByZF65pbMxuqaV319hUSQE9pC8gP4Sg,5405
+konduktor/task.py,sha256=LcTblvTNCSFNBHNUIOGN-i43whR8B6ccBm3HZTDoCMs,38050
+konduktor/templates/deployment.yaml.j2,sha256=ByA33oaBgWnSx_QZ6yKeaMcCjktphcwF8R5KiOlM4pE,4859
 konduktor/templates/jobset.yaml.j2,sha256=NQcVeRNsTLLmTnJRnkL1vr45mSeth-b11YShXn_RoSg,1323
 konduktor/templates/pod.yaml.j2,sha256=p9yE-AQkCF2Tgjd1QQiOLtzgI6Gbpps-MZYJZ9fQWIs,19159
 konduktor/usage/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -96,12 +96,12 @@ konduktor/utils/kubernetes_utils.py,sha256=XleYxzG64hciZb-CjzBDjX8BOMhFATIIHZlXD
 konduktor/utils/log_utils.py,sha256=VUyTtN819BJnSwm33-73-h8aaD51Y5Gawt6ek2kU1tk,18181
 konduktor/utils/loki_utils.py,sha256=eOGiD7dZNuwzmyXKiifyqz00EVh2nwcUPFSiPkac9y0,4050
 konduktor/utils/rich_utils.py,sha256=ycADW6Ij3wX3uT8ou7T8qxX519RxlkJivsLvUahQaJo,3583
-konduktor/utils/schemas.py,sha256=X2q-Nuk71EfMQXl4QFOtFWlQgd8tC_jFAo5dScmlEQc,19067
+konduktor/utils/schemas.py,sha256=cr39nEAgjluhXoUYnvIwCwLBH8rLds37MBsF1uQv1rw,19067
 konduktor/utils/subprocess_utils.py,sha256=WoFkoFhGecPR8-rF8WJxbIe-YtV94LXz9UG64SDhCY4,9448
 konduktor/utils/ux_utils.py,sha256=LSH4b5lckD157qDF4keThxtkGdxNrAfGKmH1ewhZkm4,8646
 konduktor/utils/validator.py,sha256=gCB5v9Up9bCWD_92fS5ChfRRXj_m56Ky9uzd_77wXGI,16927
-konduktor_nightly-0.1.0.dev20250919104536.dist-info/LICENSE,sha256=MuuqTZbHvmqXR_aNKAXzggdV45ANd3wQ5YI7tnpZhm0,6586
-konduktor_nightly-0.1.0.dev20250919104536.dist-info/METADATA,sha256=Iu3P8ik5D0tOYXoP5LMArmXo-JqSjbRsw2EMWiVRPRs,4247
-konduktor_nightly-0.1.0.dev20250919104536.dist-info/WHEEL,sha256=XbeZDeTWKc1w7CSIyre5aMDU_-PohRwTQceYnisIYYY,88
-konduktor_nightly-0.1.0.dev20250919104536.dist-info/entry_points.txt,sha256=k3nG5wDFIJhNqsZWrHk4d0irIB2Ns9s47cjRWYsTCT8,48
-konduktor_nightly-0.1.0.dev20250919104536.dist-info/RECORD,,
+konduktor_nightly-0.1.0.dev20250921104307.dist-info/LICENSE,sha256=MuuqTZbHvmqXR_aNKAXzggdV45ANd3wQ5YI7tnpZhm0,6586
+konduktor_nightly-0.1.0.dev20250921104307.dist-info/METADATA,sha256=-KRFEgnGTyI9F7gXRZmm7Yu1fsXMigW2Uo0JotzLwAw,4247
+konduktor_nightly-0.1.0.dev20250921104307.dist-info/WHEEL,sha256=XbeZDeTWKc1w7CSIyre5aMDU_-PohRwTQceYnisIYYY,88
+konduktor_nightly-0.1.0.dev20250921104307.dist-info/entry_points.txt,sha256=k3nG5wDFIJhNqsZWrHk4d0irIB2Ns9s47cjRWYsTCT8,48
+konduktor_nightly-0.1.0.dev20250921104307.dist-info/RECORD,,

konduktor/templates/apoxy-deployment.yaml.j2

@@ -1,33 +0,0 @@
----
-# Apoxy Backend for general deployment
-apiVersion: core.apoxy.dev/v1alpha
-kind: Backend
-metadata:
-  name: {{ unique_cluster_name }}-backend-{{ deployment_number }}
-  labels:
-    task_name: {{ name }}
-    endpoint_name: {{ cluster_name }}-{{ deployment_number }}.trainy.us
-spec:
-  endpoints:
-    - fqdn: {{ name }}.default.{{ unique_cluster_name }}.tun.apoxy.net
----
-# Apoxy Route for general deployment
-apiVersion: gateway.apoxy.dev/v1
-kind: HTTPRoute
-metadata:
-  name: {{ unique_cluster_name }}-route-{{ deployment_number }}
-  labels:
-    task_name: {{ name }}
-    endpoint_name: {{ cluster_name }}-{{ deployment_number }}.trainy.us
-spec:
-  parentRefs:
-    - name: default 
-      kind: Gateway
-      port: 443
-  hostnames:
-    - '{{ cluster_name }}-{{ deployment_number }}.trainy.us'
-  rules:
-    - backendRefs:
-        - kind: Backend
-          name: {{ unique_cluster_name }}-backend-{{ deployment_number }}
-          port: {{ ports }}