diff --git a/vllm/entrypoints/pooling/embed/io_processor.py b/vllm/entrypoints/pooling/embed/io_processor.py index 3c5b75cee6..1b23a77ec6 100644 --- a/vllm/entrypoints/pooling/embed/io_processor.py +++ b/vllm/entrypoints/pooling/embed/io_processor.py @@ -142,13 +142,16 @@ class EmbedIOProcessor(PoolingIOProcessor): ) prompt_token_ids = cast(list[int], token_ids) + cache_salt = cast(str | None, engine_input["prompts"].get("cache_salt")) for chunk_idx, chunk_tokens in enumerate( chunk_list(prompt_token_ids, max_model_len) ): chunked_engine_inputs.append( PoolingEngineInput( - prompts=tokens_input(prompt_token_ids=chunk_tokens), + prompts=tokens_input( + prompt_token_ids=chunk_tokens, cache_salt=cache_salt + ), params=engine_input["params"], lora_requests=engine_input["lora_requests"], priorities=engine_input["priorities"],