From 6c5872efc5fca7a53f5f3aa589e8cb7896d98618 Mon Sep 17 00:00:00 2001 From: Martin Kukla Date: Mon, 15 Jun 2026 10:31:57 +0100 Subject: [PATCH] [Bugfix] Unset HF's default max_new_tokens for DiffusionGemma (#45417) Signed-off-by: Martin Kukla --- vllm/model_executor/models/config.py | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/vllm/model_executor/models/config.py b/vllm/model_executor/models/config.py index 7354771764..6b21ef8308 100644 --- a/vllm/model_executor/models/config.py +++ b/vllm/model_executor/models/config.py @@ -158,6 +158,20 @@ class DiffusionGemmaModelForBlockDiffusionConfig(VerifyAndUpdateConfig): if sc is not None and sc.max_num_seqs >= SchedulerConfig.DEFAULT_MAX_NUM_SEQS: sc.max_num_seqs = 8 + # Remove the model's generation_config.json cap on max_new_tokens + # (256) so DiffusionGemma behaves like every other model: no + # server-wide limit, each request controls its own output length + # via max_tokens. Setting to None causes get_diff_sampling_param + # to skip this key entirely. + model_config = vllm_config.model_config + if "max_new_tokens" not in model_config.override_generation_config: + model_config.override_generation_config["max_new_tokens"] = None + logger.info( + "DiffusionGemma: removing server-wide max_new_tokens cap " + "from generation_config.json (use " + "--override-generation-config to set a custom limit).", + ) + class DeepseekV4ForCausalLMConfig(VerifyAndUpdateConfig): @staticmethod