From 10a1eaac97ae3e67b06807b4d7e331a9a5957371 Mon Sep 17 00:00:00 2001 From: BillionClaw <267901332+BillionClaw@users.noreply.github.com> Date: Wed, 18 Mar 2026 17:51:11 +0800 Subject: [PATCH] fix(janus): handle None values in image generation mode Fixes #44792 - Add null check for generation_config.num_return_sequences (default to 1) - Add null check for generation_config.max_length (default to 0) --- src/transformers/models/janus/modeling_janus.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/transformers/models/janus/modeling_janus.py b/src/transformers/models/janus/modeling_janus.py index c329983b7a6a..3135902d3955 100644 --- a/src/transformers/models/janus/modeling_janus.py +++ b/src/transformers/models/janus/modeling_janus.py @@ -1285,7 +1285,7 @@ def generate( input_ids, model_kwargs = self._expand_inputs_for_generation( input_ids=input_ids, attention_mask=attention_mask, - expand_size=generation_config.num_return_sequences, + expand_size=generation_config.num_return_sequences or 1, **model_kwargs, ) @@ -1315,7 +1315,7 @@ def generate( # batch_size should account for both conditional/unconditional input; hence multiplied by 2. batch_size=batch_size * 2, # we should have at least a cache len of seq_len + num_image_tokens. - max_cache_len=max(generation_config.max_length, num_image_tokens + seq_len), + max_cache_len=max(generation_config.max_length or 0, num_image_tokens + seq_len), model_kwargs=model_kwargs, )