Skip to content

Commit 5ef6fa4

Browse files
phymbertjordankanter
authored andcommitted
server: tests - slow inference causes timeout on the CI (ggml-org#5715)
* server: tests - longer inference timeout for CI
1 parent 9c4fabf commit 5ef6fa4

File tree

2 files changed

+4
-2
lines changed

2 files changed

+4
-2
lines changed

common/sampling.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -266,7 +266,7 @@ static llama_token llama_sampling_sample_impl(
266266
// }
267267
//}
268268

269-
LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
269+
//LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
270270
}
271271
}
272272

examples/server/tests/features/steps/steps.py

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -699,6 +699,8 @@ async def wait_for_health_status(context,
699699
if context.debug:
700700
print(f"Starting checking for health for expected_health_status={expected_health_status}")
701701
timeout = 3 # seconds
702+
if expected_health_status == 'ok':
703+
timeout = 10 # CI slow inference
702704
interval = 0.5
703705
counter = 0
704706
async with aiohttp.ClientSession() as session:
@@ -736,7 +738,7 @@ async def wait_for_health_status(context,
736738
if n_completions > 0:
737739
return
738740

739-
assert False, 'timeout exceeded'
741+
assert False, f'{expected_health_status} timeout exceeded {counter}s>={timeout}'
740742

741743

742744
def assert_embeddings(embeddings):

0 commit comments

Comments
 (0)