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

Signed-off-by: Martin Kukla <martin.kukla@cantab.net>
This commit is contained in:
Martin Kukla
2026-06-15 17:31:57 +08:00
committed by GitHub
parent 1d88c4dadd
commit 6c5872efc5
+14
View File
@@ -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