Various minor updates to deployment manifests

Author: njhill

deployment/models/flan-ul2-tp/kustomization.yaml

@@ -13,6 +13,7 @@ patchesStrategicMerge:
   - ../../base/patches/gpus/2-gpu.yaml
   - ../../base/patches/gpus/a100-80g.yaml
   - ../../base/patches/limits/t5-xxl.yaml
+  - ../../base/patches/shm.yaml
   - |-
    apiVersion: apps/v1
    kind: Deployment

deployment/models/gpt-neox/kustomization.yaml

@@ -33,7 +33,7 @@ patchesStrategicMerge:
              - name: MAX_CONCURRENT_REQUESTS
                value: "320"
              - name: MAX_BATCH_WEIGHT
-               value: "9900"
+               value: "9200"
              - name: MAX_SEQUENCE_LENGTH
                value: "8192"
              - name: MAX_PREFILL_WEIGHT

deployment/models/ibm-mpt-7b-instruct/kustomization.yaml

@@ -40,11 +40,3 @@ patchesStrategicMerge:
                value: "2000000"
              - name: MAX_BATCH_WEIGHT
                value: "200000000"
-
-             startupProbe:
-               httpGet:
-                 port: http
-                 path: /health
-               # MPT-7B currently takes more than 15 mins to load :-(
-               failureThreshold: 52
-               periodSeconds: 30

deployment/models/mpt-7b/kustomization.yaml

@@ -42,11 +42,3 @@ patchesStrategicMerge:
                value: "2000000"
              - name: MAX_BATCH_WEIGHT
                value: "200000000"
-
-             startupProbe:
-               httpGet:
-                 port: http
-                 path: /health
-               # MPT-7B currently takes more than 15 mins to load :-(
-               failureThreshold: 52
-               periodSeconds: 30