mirror of
https://github.com/BerriAI/litellm.git
synced 2026-11-08 08:08:37 +00:00
ab78459097
Response transformation crashed with TypeError when counting tokens for multimodal (dict-shaped) embedding inputs, since encoding.encode() was called directly on the dict instead of its text parts. Also removed a dead, equally-buggy duplicate of the same loop in _transform_response. Separately, the embeddings guardrail handler silently skipped multimodal dict inputs instead of scanning them, letting prohibited text/image content bypass the configured guardrail. It now extracts and scans the text parts while leaving image parts untouched.
44 lines
1.1 KiB
Python
44 lines
1.1 KiB
Python
from __future__ import annotations
|
|
|
|
import os
|
|
import sys
|
|
|
|
import redis
|
|
|
|
from tests._vcr_redis_persister import CASSETTE_REDIS_URL_ENV, _redis_url_from_env
|
|
|
|
PREFIX = "litellm:vcr:cassette:"
|
|
SCAN_BATCH = 500
|
|
|
|
|
|
def _client() -> redis.Redis:
|
|
url = _redis_url_from_env()
|
|
if not url:
|
|
sys.exit(f"Set {CASSETTE_REDIS_URL_ENV} to flush the VCR cache")
|
|
return redis.Redis.from_url(
|
|
url,
|
|
socket_timeout=5,
|
|
socket_connect_timeout=5,
|
|
decode_responses=False,
|
|
)
|
|
|
|
|
|
def main() -> None:
|
|
client = _client()
|
|
deleted = 0
|
|
pipeline = client.pipeline(transaction=False)
|
|
pending = 0
|
|
for key in client.scan_iter(match=f"{PREFIX}*", count=SCAN_BATCH):
|
|
pipeline.delete(key)
|
|
pending += 1
|
|
if pending >= SCAN_BATCH:
|
|
deleted += sum(pipeline.execute())
|
|
pipeline = client.pipeline(transaction=False)
|
|
pending = 0
|
|
if pending:
|
|
deleted += sum(pipeline.execute())
|
|
print(f"Deleted {deleted} VCR cassette key(s) under {PREFIX!r}")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|