test: add tests for truncation with prefixes

Signed-off-by: Travis Johnson <[email protected]>

Author: tjohnson31415

integration_tests/test_cases_bloom560m.yaml

@@ -58,6 +58,24 @@
         stopReason: EOS_TOKEN
         text: ' positive'
 
+- name: Greedy with tuned prompt prefix and truncation
+  # Prompt prefixes with multi-shard not yet supported
+  singleShardOnly: true
+  request:
+    prefixId: bloom_sentiment_1
+    params:
+      truncateInputTokens: 10
+      method: GREEDY
+      stopping: {"maxNewTokens": 20}
+    requests:
+      - {"text": "[This will be truncated] What a wonderful day!:\nSentiment:"}
+  response:
+    responses:
+      - generatedTokenCount: 2
+        inputTokenCount: 10
+        stopReason: EOS_TOKEN
+        text: ' positive'
+
 
 # Prompt prefix with nested path
 - name: Greedy with tuned prompt prefix with nested path (id)

integration_tests/test_cases_mt0small.yaml

@@ -90,6 +90,22 @@
         stopReason: EOS_TOKEN
         text: They are a wonderful day.
 
+# Prompt prefix - decoder only with truncation
+- name: Greedy with tuned prompt prefix and truncation for decoder only
+  request:
+    prefixId: mt0_sentiment_decoder_only
+    params:
+      method: GREEDY
+      truncateInputTokens: 12
+      stopping: {"maxNewTokens": 20}
+    requests:
+      - {"text": "[This will be truncated] What a wonderful day!:\nSentiment:"}
+  response:
+    responses:
+      - generatedTokenCount: 10
+        inputTokenCount: 12
+        stopReason: EOS_TOKEN
+        text: They are a wonderful day.
 
 # Prompt prefix returning input and generated tokens
 - name: Greedy with tuned prompt prefix and returned tokens

integration_tests/test_cases_tinystarcoderpy.yaml

@@ -33,7 +33,7 @@
         text: "\nclass Shape(object):\n    '''Shape class'''\n\n    def __init__(self, x,"
 
 # Basic Greedy (explicit)
-- name: Basic Greedy, max new tokens (implicit)
+- name: Basic Greedy, max new tokens (explicit)
   request:
     params:
       method: GREEDY
@@ -140,6 +140,24 @@
         stopReason: EOS_TOKEN
         text: "(\"Hello World!\")\n\nhello_world()\n"
 
+- name: Greedy with tuned prompt prefix and truncation
+  # Prompt prefixes with multi-shard not yet supported
+  singleShardOnly: true
+  request:
+    # Prefix is "def hello_world():\n"
+    prefixId: tiny_starcoder
+    params:
+      method: GREEDY
+      truncateInputTokens: 2
+    requests:
+      - {"text": "[this will be truncated]\tprint"}
+  response:
+    responses:
+      - generatedTokenCount: 12
+        inputTokenCount: 2
+        stopReason: EOS_TOKEN
+        text: "(\"Hello World!\")\n\nhello_world()\n"
+
 
 # Prompt prefix returning input and generated tokens
 - name: Greedy with tuned prompt prefix and returned tokens