From 6eac8e0070c574fe1b8d6c276422ea301ddfda4c Mon Sep 17 00:00:00 2001 From: Taneem Ibrahim Date: Sat, 4 Jul 2026 07:14:13 -0500 Subject: [PATCH] [Misc] Preserve cross-encoder pooling extra kwargs (#47082) Signed-off-by: Taneem Ibrahim Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> --- vllm/entrypoints/pooling/scoring/io_processor.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/vllm/entrypoints/pooling/scoring/io_processor.py b/vllm/entrypoints/pooling/scoring/io_processor.py index 7c58e7e9ab..811e69e0ee 100644 --- a/vllm/entrypoints/pooling/scoring/io_processor.py +++ b/vllm/entrypoints/pooling/scoring/io_processor.py @@ -463,7 +463,10 @@ class CrossEncoderIOProcessor(ScoringIOProcessor): if token_type_ids := engine_prompt.pop("token_type_ids", None): params = pooling_params.clone() compressed = compress_token_type_ids(token_type_ids) - params.extra_kwargs = {"compressed_token_type_ids": compressed} + params.extra_kwargs = { + **(params.extra_kwargs or {}), + "compressed_token_type_ids": compressed, + } pooling_params_list.append(params) else: pooling_params_list.append(pooling_params)