[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/vllm/v1/engine/processor.py
+++ b/vllm/v1/engine/processor.py
@@ -317,6 +317,7 @@ class Processor:
            eos_token_id=eos_token_id,
            arrival_time=arrival_time,
            lora_request=lora_request,
+            cache_salt=decoder_inputs.get("cache_salt"),
        )

    def _validate_model_inputs(self,