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

Server: fix seed for multiple slots (#6835)

Browse source 
* Server: add tests for consistent results

* sampling: separate rng per sampling context
This commit is contained in:
Johannes Gäßler 2024-04-24 11:08:36 +02:00 • committed by GitHub
parent c0d1b3e03e
commit 28103f4832
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
11 changed files with 145 additions and 30 deletions
Download patch file Download diff file Expand all files Collapse all files

57
examples/server/tests/features/results.feature Normal file
View file

@ -0,0 +1,57 @@
@llama.cpp
@results
Feature: Results
Background: Server startup
Given a server listening on localhost:8080
And a model file tinyllamas/split/stories15M-00001-of-00003.gguf from HF repo ggml-org/models
And a model file test-model-00001-of-00003.gguf
And 128 as batch size
And 256 KV cache size
And 128 max tokens to predict
Scenario Outline: Multi users completion
Given <n_slots> slots
And continuous batching
Then the server is starting
Then the server is healthy
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given concurrent completion requests
Then the server is busy
Then the server is idle
And all slots are idle
Then all predictions are equal
Examples:
| n_slots |
| 1 |
| 2 |

34
examples/server/tests/features/steps/steps.py
View file

@ -61,6 +61,7 @@ def step_server_config(context, server_fqdn, server_port):
context.server_metrics = False
context.server_process = None
context.seed = None
context.draft = None
context.server_seed = None
context.user_api_key = None
context.response_format = None
@ -107,6 +108,11 @@ def step_n_gpu_layer(context, ngl):
context.n_gpu_layer = ngl
@step('{draft:d} as draft')
def step_draft(context, draft):
context.draft = draft
@step('{n_ctx:d} KV cache size')
def step_n_ctx(context, n_ctx):
context.n_ctx = n_ctx
@ -254,6 +260,15 @@ def step_n_tokens_predicted(context, predicted_n):
assert_n_tokens_predicted(context.completion, predicted_n)
@step('all predictions are equal')
@async_run_until_complete
async def step_predictions_equal(context):
n_completions = await gather_tasks_results(context)
assert n_completions >= 2, "need at least 2 completions"
assert_all_predictions_equal(context.tasks_result)
context.tasks_result = []
@step('the completion is truncated')
def step_assert_completion_truncated(context):
step_assert_completion_truncated(context, '')
@ -1020,6 +1035,23 @@ def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re
assert n_predicted == expected_predicted_n, (f'invalid number of tokens predicted:'
f' {n_predicted} <> {expected_predicted_n}')
def assert_all_predictions_equal(completion_responses):
content_0 = completion_responses[0]['content']
if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
print(f"content 0: {content_0}")
i = 1
for response in completion_responses[1:]:
content = response['content']
if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
print(f"content {i}: {content}")
assert content == content_0, "contents not equal"
i += 1
async def gather_tasks_results(context):
n_tasks = len(context.concurrent_tasks)
@ -1148,6 +1180,8 @@ def start_server_background(context):
server_args.extend(['--ubatch-size', context.n_ubatch])
if context.n_gpu_layer:
server_args.extend(['--n-gpu-layers', context.n_gpu_layer])
if context.draft is not None:
server_args.extend(['--draft', context.draft])
if context.server_continuous_batching:
server_args.append('--cont-batching')
if context.server_embeddings: