[Bugfix] Unset HF's default max_new_tokens for DiffusionGemma (#45417)

Signed-off-by: Martin Kukla <martin.kukla@cantab.net>

Martin Kukla committed Jun 15, 2026 at 10:31 UTC 6c5872efc5fca7a53f5f3aa589e8cb7896d98618

1 file changed +14

vllm/model_executor/models/config.py

+14

@@ -158,6 +158,20 @@ class DiffusionGemmaModelForBlockDiffusionConfig(VerifyAndUpdateConfig):

158 if sc is not None and sc.max_num_seqs >= SchedulerConfig.DEFAULT_MAX_NUM_SEQS:

159 sc.max_num_seqs = 8

160

161 + # Remove the model's generation_config.json cap on max_new_tokens

162 + # (256) so DiffusionGemma behaves like every other model: no

163 + # server-wide limit, each request controls its own output length

164 + # via max_tokens. Setting to None causes get_diff_sampling_param

165 + # to skip this key entirely.

166 + model_config = vllm_config.model_config

167 + if "max_new_tokens" not in model_config.override_generation_config:

168 + model_config.override_generation_config["max_new_tokens"] = None

169 + logger.info(

170 + "DiffusionGemma: removing server-wide max_new_tokens cap "

171 + "from generation_config.json (use "

172 + "--override-generation-config to set a custom limit).",

173 + )

174 +

175

176 class DeepseekV4ForCausalLMConfig(VerifyAndUpdateConfig):

177 @staticmethod