llama.cpp/examples/server/tests/features/wrong_usages.feature

26 lines
883 B
Gherkin
Raw Normal View History

# run with: ./tests.sh --no-skipped --tags wrong_usage
@wrong_usage
Feature: Wrong usage of llama.cpp server
#3969 The user must always set --n-predict option
# to cap the number of tokens any completion request can generate
# or pass n_predict/max_tokens in the request.
Scenario: Infinite loop
Given a server listening on localhost:8080
And a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
And 42 as server seed
And 2048 KV cache size
# Uncomment below to fix the issue
#And 64 server max tokens to predict
Then the server is starting
Then the server is healthy
Given a prompt:
"""
Go to: infinite loop
"""
# Uncomment below to fix the issue
#And 128 max tokens to predict
Given concurrent completion requests
Then the server is idle
Then all prompts are predicted