vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

server: tests: add infinite loop

Browse source
This commit is contained in:
Pierrick HYMBERT 2024-02-20 23:05:27 +01:00
parent 0772884b06
commit 6b9dc4f291
2 changed files with 30 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

43
examples/server/tests/features/server.feature
View file

@ -36,24 +36,13 @@ Feature: llama.cpp server
Scenario: Multi users Scenario: Multi users
Given a prompt: Given a prompt:
""" """
Write a formal complaint email to Air France about my delayed Write a very long story about AI.
baggage from my flight on Tuesday, January 17th, from Paris to Toulouse. Be verbose.
""" """
And a prompt: And a prompt:
""" """
Translate the following War & Peace chapter into Russian: WELL, PRINCE, Write another very long music lyrics.
Genoa and Lucca are now no more than private estates of the Bonaparte
family. No, I warn you, that if you do not tell me we are at war,
if you again allow yourself to palliate all the infamies and atrocities
of this Antichrist (upon my word, I believe he is), I don’t know you
in future, you are no longer my friend, no longer my faithful slave,
as you say. There, how do you do, how do you do? I see I’m scaring you,
sit down and talk to me.” These words were uttered in July 1805 by
Anna Pavlovna Scherer, a distinguished lady of the court,
and confidential maid-of-honour to the Empress Marya Fyodorovna.
It was her greeting to Prince Vassily, a man high in rank
and office, who was the first to arrive at her soirée.
""" """
And 512 max tokens to predict
Given concurrent completion requests Given concurrent completion requests
Then the server is busy Then the server is busy
And all slots are busy And all slots are busy
@ -65,8 +54,6 @@ Feature: llama.cpp server
Scenario: Multi users OAI Compatibility Scenario: Multi users OAI Compatibility
Given a system prompt "You are an AI assistant." Given a system prompt "You are an AI assistant."
And a model tinyllama-2 And a model tinyllama-2
And 1024 max tokens to predict
And streaming is enabled
Given a prompt: Given a prompt:
""" """
Write a very long story about AI. Write a very long story about AI.
@ -75,6 +62,8 @@ Feature: llama.cpp server
""" """
Write another very long music lyrics. Write another very long music lyrics.
""" """
And 512 max tokens to predict
And streaming is enabled
Given concurrent OAI completions requests Given concurrent OAI completions requests
Then the server is busy Then the server is busy
And all slots are busy And all slots are busy
@ -82,3 +71,25 @@ Feature: llama.cpp server
And all slots are idle And all slots are idle
Then all prompts are predicted Then all prompts are predicted
# FIXME: infinite loop on the CI, not locally, if n_prompt * n_predict > kv_size
Scenario: Multi users with total number of tokens to predict exceeds the KV Cache size
Given a prompt:
"""
Write a very long story about AI.
"""
And a prompt:
"""
Write another very long music lyrics.
"""
And a prompt:
"""
Write a very long poem.
"""
And 1024 max tokens to predict
Given concurrent completion requests
Then the server is busy
And all slots are busy
Then the server is idle
And all slots are idle
Then all prompts are predicted

6
examples/server/tests/features/steps/steps.py
View file

@ -105,7 +105,7 @@ def step_model(context, model):
@step(u'{max_tokens} max tokens to predict') @step(u'{max_tokens} max tokens to predict')
def step_max_tokens(context, max_tokens): def step_max_tokens(context, max_tokens):
context.max_tokens = int(max_tokens) context.n_predict = int(max_tokens)
@step(u'streaming is {enable_streaming}') @step(u'streaming is {enable_streaming}')
@ -154,7 +154,7 @@ def concurrent_requests(context, f_completion):
def request_completion(context, prompt, n_predict=None): def request_completion(context, prompt, n_predict=None):
response = requests.post(f'{context.base_url}/completion', json={ response = requests.post(f'{context.base_url}/completion', json={
"prompt": prompt, "prompt": prompt,
"n_predict": int(n_predict) if n_predict is not None else 4096, "n_predict": int(n_predict) if n_predict is not None else context.n_predict,
"seed": context.seed "seed": context.seed
}) })
assert response.status_code == 200 assert response.status_code == 200
@ -174,7 +174,7 @@ def oai_chat_completions(context, user_prompt):
} }
], ],
model=context.model, model=context.model,
max_tokens=context.max_tokens, max_tokens=context.n_predict,
stream=context.enable_streaming, stream=context.enable_streaming,
seed = context.seed seed = context.seed
) )