From 8dddffa34031c8d88bf01ca9bedceffc503499f5 Mon Sep 17 00:00:00 2001 From: "sirui.li" <48873278+lsrami@users.noreply.github.com> Date: Wed, 5 Mar 2025 18:03:36 +0800 Subject: [PATCH] [data] fix qwen2audio plugin (#7166) * Update pairwise.py [data]Repair multimodal model dpo training * Update pairwise.py [data]repair multimodal model dpo training using deepcopy * Update pairwise.py * Update mm_plugin.py Former-commit-id: dff4130969bac9cb1abe66fd5dfada8c757c716f --- src/llamafactory/data/mm_plugin.py | 1 + 1 file changed, 1 insertion(+) diff --git a/src/llamafactory/data/mm_plugin.py b/src/llamafactory/data/mm_plugin.py index bc6e1afe..5597d73b 100644 --- a/src/llamafactory/data/mm_plugin.py +++ b/src/llamafactory/data/mm_plugin.py @@ -993,6 +993,7 @@ class Qwen2AudioPlugin(BasePlugin): self._validate_input(processor, images, videos, audios) bos_token: str = getattr(processor, "audio_bos_token") eos_token: str = getattr(processor, "audio_eos_token") + messages = deepcopy(messages) mm_inputs = self._get_mm_inputs([], [], audios, processor) if "feature_attention_mask" in mm_inputs: audio_lengths = mm_inputs["feature_attention_mask"].sum(-1).tolist()