phoenix/litellm-mirror
0
0
Fork 1
mirror of https://github.com/BerriAI/litellm.git synced 2025-04-26 11:14:04 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Use max_input_token for trim_messages

Browse source
This commit is contained in:
Chen Wang 2024-04-16 13:36:25 +01:00
parent 7ffd3d40fa
commit ebc889d77a
No known key found for this signature in database
GPG key ID: 4B04331CFE3E3BF5
2 changed files with 24 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

16
litellm/tests/test_utils.py
View file

@ -173,6 +173,22 @@ def test_trimming_should_not_change_original_messages():
assert messages == messages_copy
@pytest.mark.parametrize("model", ["gpt-4-0125-preview", "claude-3-opus-20240229"])
def test_trimming_with_model_cost_max_input_tokens(model):
messages = [
{"role": "system", "content": "This is a normal system message"},
{
"role": "user",
"content": "This is a sentence" * 100000,
},
]
trimmed_messages = trim_messages(messages, model=model)
assert (
get_token_count(trimmed_messages, model=model)
< litellm.model_cost[model]["max_input_tokens"]
)
def test_get_valid_models():
old_environ = os.environ
os.environ = {"OPENAI_API_KEY": "temp"} # mock set only openai key in environ

13
litellm/utils.py
View file

@ -10577,16 +10577,19 @@ def trim_messages(
messages = copy.deepcopy(messages)
try:
print_verbose(f"trimming messages")
if max_tokens == None:
if max_tokens is None:
# Check if model is valid
if model in litellm.model_cost:
max_tokens_for_model = litellm.model_cost[model]["max_tokens"]
if (
model in litellm.model_cost
and "max_input_tokens" in litellm.model_cost[model]
):
max_tokens_for_model = litellm.model_cost[model]["max_input_tokens"]
max_tokens = int(max_tokens_for_model * trim_ratio)
else:
# if user did not specify max tokens
# if user did not specify max input tokens
# or passed an llm litellm does not know
# do nothing, just return messages
return
return messages
system_message = ""
for message in messages: