update V1.0 benchmark manifest (#822)

Co-authored-by: Zhenzhong1 <zhenzhong.xu@intel.com>
opea-project · Sep 18, 2024 · e5affb9 · e5affb9
1 parent e2a74f7
commit e5affb9
Show file tree

Hide file tree

Showing 22 changed files with 38 additions and 70 deletions.
diff --git a/...chmark/oob_no_wrapper/with_rerank/eight_gaudi/no_wrapper_oob_eight_gaudi_with_rerank.yaml b/...chmark/oob_no_wrapper/with_rerank/eight_gaudi/no_wrapper_oob_eight_gaudi_with_rerank.yaml
@@ -134,7 +134,7 @@ metadata:
   name: embedding-dependency-deploy
   namespace: default
 spec:
-  replicas: 1
+  replicas: 8
   selector:
     matchLabels:
       app: embedding-dependency-deploy
@@ -223,10 +223,6 @@ spec:
         - '2048'
         - --max-total-tokens
         - '4096'
-        - --max-batch-total-tokens
-        - '65536'
-        - --max-batch-prefill-tokens
-        - '4096'
         env:
         - name: OMPI_MCA_btl_vader_single_copy_mechanism
           value: none

diff --git a/...enchmark/oob_no_wrapper/with_rerank/four_gaudi/no_wrapper_oob_four_gaudi_with_rerank.yaml b/...enchmark/oob_no_wrapper/with_rerank/four_gaudi/no_wrapper_oob_four_gaudi_with_rerank.yaml
@@ -134,7 +134,7 @@ metadata:
   name: embedding-dependency-deploy
   namespace: default
 spec:
-  replicas: 1
+  replicas: 4
   selector:
     matchLabels:
       app: embedding-dependency-deploy
@@ -223,10 +223,6 @@ spec:
         - '2048'
         - --max-total-tokens
         - '4096'
-        - --max-batch-total-tokens
-        - '65536'
-        - --max-batch-prefill-tokens
-        - '4096'
         env:
         - name: OMPI_MCA_btl_vader_single_copy_mechanism
           value: none

diff --git a/...mark/oob_no_wrapper/with_rerank/single_gaudi/no_wrapper_oob_single_gaudi_with_rerank.yaml b/...mark/oob_no_wrapper/with_rerank/single_gaudi/no_wrapper_oob_single_gaudi_with_rerank.yaml
@@ -223,10 +223,6 @@ spec:
         - '2048'
         - --max-total-tokens
         - '4096'
-        - --max-batch-total-tokens
-        - '65536'
-        - --max-batch-prefill-tokens
-        - '4096'
         env:
         - name: OMPI_MCA_btl_vader_single_copy_mechanism
           value: none

diff --git a/.../benchmark/oob_no_wrapper/with_rerank/two_gaudi/no_wrapper_oob_two_gaudi_with_rerank.yaml b/.../benchmark/oob_no_wrapper/with_rerank/two_gaudi/no_wrapper_oob_two_gaudi_with_rerank.yaml
@@ -134,7 +134,7 @@ metadata:
   name: embedding-dependency-deploy
   namespace: default
 spec:
-  replicas: 1
+  replicas: 2
   selector:
     matchLabels:
       app: embedding-dependency-deploy
@@ -223,10 +223,6 @@ spec:
         - '2048'
         - --max-total-tokens
         - '4096'
-        - --max-batch-total-tokens
-        - '65536'
-        - --max-batch-prefill-tokens
-        - '4096'
         env:
         - name: OMPI_MCA_btl_vader_single_copy_mechanism
           value: none

diff --git a/.../oob_no_wrapper/without_rerank/eight_gaudi/no_wrapper_oob_eight_gaudi_without_rerank.yaml b/.../oob_no_wrapper/without_rerank/eight_gaudi/no_wrapper_oob_eight_gaudi_without_rerank.yaml
@@ -134,7 +134,7 @@ metadata:
   name: embedding-dependency-deploy
   namespace: default
 spec:
-  replicas: 1
+  replicas: 8
   selector:
     matchLabels:
       app: embedding-dependency-deploy
@@ -223,10 +223,6 @@ spec:
         - '2048'
         - --max-total-tokens
         - '4096'
-        - --max-batch-total-tokens
-        - '65536'
-        - --max-batch-prefill-tokens
-        - '4096'
         env:
         - name: OMPI_MCA_btl_vader_single_copy_mechanism
           value: none

diff --git a/...rk/oob_no_wrapper/without_rerank/four_gaudi/no_wrapper_oob_four_gaudi_without_rerank.yaml b/...rk/oob_no_wrapper/without_rerank/four_gaudi/no_wrapper_oob_four_gaudi_without_rerank.yaml
@@ -134,7 +134,7 @@ metadata:
   name: embedding-dependency-deploy
   namespace: default
 spec:
-  replicas: 1
+  replicas: 4
   selector:
     matchLabels:
       app: embedding-dependency-deploy
@@ -223,10 +223,6 @@ spec:
         - '2048'
         - --max-total-tokens
         - '4096'
-        - --max-batch-total-tokens
-        - '65536'
-        - --max-batch-prefill-tokens
-        - '4096'
         env:
         - name: OMPI_MCA_btl_vader_single_copy_mechanism
           value: none

diff --git a/...ob_no_wrapper/without_rerank/single_gaudi/no_wrapper_oob_single_gaudi_without_rerank.yaml b/...ob_no_wrapper/without_rerank/single_gaudi/no_wrapper_oob_single_gaudi_without_rerank.yaml
@@ -223,10 +223,6 @@ spec:
         - '2048'
         - --max-total-tokens
         - '4096'
-        - --max-batch-total-tokens
-        - '65536'
-        - --max-batch-prefill-tokens
-        - '4096'
         env:
         - name: OMPI_MCA_btl_vader_single_copy_mechanism
           value: none

diff --git a/...mark/oob_no_wrapper/without_rerank/two_gaudi/no_wrapper_oob_two_gaudi_without_rerank.yaml b/...mark/oob_no_wrapper/without_rerank/two_gaudi/no_wrapper_oob_two_gaudi_without_rerank.yaml
@@ -134,7 +134,7 @@ metadata:
   name: embedding-dependency-deploy
   namespace: default
 spec:
-  replicas: 1
+  replicas: 2
   selector:
     matchLabels:
       app: embedding-dependency-deploy
@@ -223,10 +223,6 @@ spec:
         - '2048'
         - --max-total-tokens
         - '4096'
-        - --max-batch-total-tokens
-        - '65536'
-        - --max-batch-prefill-tokens
-        - '4096'
         env:
         - name: OMPI_MCA_btl_vader_single_copy_mechanism
           value: none

diff --git a/ChatQnA/benchmark/tuned/with_rerank/four_gaudi/tuned_four_gaudi_with_rerank.yaml b/ChatQnA/benchmark/tuned/with_rerank/four_gaudi/tuned_four_gaudi_with_rerank.yaml
@@ -167,10 +167,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
         volumeMounts:
         - mountPath: /data

diff --git a/ChatQnA/benchmark/tuned/with_rerank/single_gaudi/tuned_single_gaudi_with_rerank.yaml b/ChatQnA/benchmark/tuned/with_rerank/single_gaudi/tuned_single_gaudi_with_rerank.yaml
@@ -167,10 +167,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
         volumeMounts:
         - mountPath: /data

diff --git a/ChatQnA/benchmark/tuned/with_rerank/two_gaudi/tuned_two_gaudi_with_rerank.yaml b/ChatQnA/benchmark/tuned/with_rerank/two_gaudi/tuned_two_gaudi_with_rerank.yaml
@@ -167,10 +167,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
         volumeMounts:
         - mountPath: /data

diff --git a/ChatQnA/benchmark/tuned/without_rerank/four_gaudi/tuned_four_gaudi_without_rerank.yaml b/ChatQnA/benchmark/tuned/without_rerank/four_gaudi/tuned_four_gaudi_without_rerank.yaml
@@ -216,10 +216,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
       serviceAccountName: default
       volumes:

diff --git a/ChatQnA/benchmark/tuned/without_rerank/single_gaudi/tuned_single_gaudi_without_rerank.yaml b/ChatQnA/benchmark/tuned/without_rerank/single_gaudi/tuned_single_gaudi_without_rerank.yaml
@@ -216,10 +216,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
       serviceAccountName: default
       volumes:

diff --git a/ChatQnA/benchmark/tuned/without_rerank/two_gaudi/tuned_two_gaudi_without_rerank.yaml b/ChatQnA/benchmark/tuned/without_rerank/two_gaudi/tuned_two_gaudi_without_rerank.yaml
@@ -216,10 +216,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
       serviceAccountName: default
       volumes:

diff --git a/...rk/tuned_no_wrapper/with_rerank/eight_gaudi/no_wrapper_tuned_eight_gaudi_with_rerank.yaml b/...rk/tuned_no_wrapper/with_rerank/eight_gaudi/no_wrapper_tuned_eight_gaudi_with_rerank.yaml
@@ -167,10 +167,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
         volumeMounts:
         - mountPath: /data

diff --git a/...mark/tuned_no_wrapper/with_rerank/four_gaudi/no_wrapper_tuned_four_gaudi_with_rerank.yaml b/...mark/tuned_no_wrapper/with_rerank/four_gaudi/no_wrapper_tuned_four_gaudi_with_rerank.yaml
@@ -167,10 +167,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
         volumeMounts:
         - mountPath: /data

diff --git a/.../tuned_no_wrapper/with_rerank/single_gaudi/no_wrapper_tuned_single_gaudi_with_rerank.yaml b/.../tuned_no_wrapper/with_rerank/single_gaudi/no_wrapper_tuned_single_gaudi_with_rerank.yaml
@@ -167,10 +167,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
         volumeMounts:
         - mountPath: /data
@@ -399,7 +399,7 @@ metadata:
   name: retriever-deploy
   namespace: default
 spec:
-  replicas: 1
+  replicas: 2
   selector:
     matchLabels:
       app: retriever-deploy

diff --git a/...chmark/tuned_no_wrapper/with_rerank/two_gaudi/no_wrapper_tuned_two_gaudi_with_rerank.yaml b/...chmark/tuned_no_wrapper/with_rerank/two_gaudi/no_wrapper_tuned_two_gaudi_with_rerank.yaml
@@ -167,10 +167,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
         volumeMounts:
         - mountPath: /data

diff --git a/...ed_no_wrapper/without_rerank/eight_gaudi/no_wrapper_tuned_eight_gaudi_without_rerank.yaml b/...ed_no_wrapper/without_rerank/eight_gaudi/no_wrapper_tuned_eight_gaudi_without_rerank.yaml
@@ -29,7 +29,7 @@ metadata:
   name: chatqna-backend-server-deploy
   namespace: default
 spec:
-  replicas: 4
+  replicas: 8
   selector:
     matchLabels:
       app: chatqna-backend-server-deploy
@@ -141,7 +141,7 @@ metadata:
   name: embedding-dependency-deploy
   namespace: default
 spec:
-  replicas: 4
+  replicas: 8
   selector:
     matchLabels:
       app: embedding-dependency-deploy
@@ -167,10 +167,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
         volumeMounts:
         - mountPath: /data
@@ -313,7 +313,7 @@ metadata:
   name: retriever-deploy
   namespace: default
 spec:
-  replicas: 4
+  replicas: 8
   selector:
     matchLabels:
       app: retriever-deploy

diff --git a/...uned_no_wrapper/without_rerank/four_gaudi/no_wrapper_tuned_four_gaudi_without_rerank.yaml b/...uned_no_wrapper/without_rerank/four_gaudi/no_wrapper_tuned_four_gaudi_without_rerank.yaml
@@ -167,10 +167,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
         volumeMounts:
         - mountPath: /data

diff --git a/..._no_wrapper/without_rerank/single_gaudi/no_wrapper_tuned_single_gaudi_without_rerank.yaml b/..._no_wrapper/without_rerank/single_gaudi/no_wrapper_tuned_single_gaudi_without_rerank.yaml
@@ -167,10 +167,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
         volumeMounts:
         - mountPath: /data

diff --git a/.../tuned_no_wrapper/without_rerank/two_gaudi/no_wrapper_tuned_two_gaudi_without_rerank.yaml b/.../tuned_no_wrapper/without_rerank/two_gaudi/no_wrapper_tuned_two_gaudi_without_rerank.yaml
@@ -167,10 +167,10 @@ spec:
         - containerPort: 80
         resources:
           limits:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
           requests:
-            cpu: 80
+            cpu: 76
             memory: 20000Mi
         volumeMounts:
         - mountPath: /data