[Core] Prevent side-channel attacks via cache salting (#17045)

Signed-off-by: Marko Rosenmueller <5467316+dr75@users.noreply.github.com>
2025-04-30 14:27:21 +02:00
parent a7d5b016bd
commit 77073c77bc
18 changed files with 328 additions and 126 deletions
--- a/tests/tokenization/test_detokenize.py
+++ b/tests/tokenization/test_detokenize.py
@ -60,8 +60,16 @@ def _run_incremental_decode(tokenizer,
        skip_special_tokens=skip_special_tokens,
        spaces_between_special_tokens=spaces_between_special_tokens,
    )
-    request = EngineCoreRequest("", prompt_token_ids, None, None, None, params,
-                                None, 0.0, None)
+    request = EngineCoreRequest("",
+                                prompt_token_ids,
+                                None,
+                                None,
+                                None,
+                                params,
+                                None,
+                                0.0,
+                                None,
+                                cache_salt=None)

    if fast is None:
        detokenizer = IncrementalDetokenizer.from_new_request(