From 09d81b44a788b5336672c352023cbcd7c5130639 Mon Sep 17 00:00:00 2001
From: Yuekai Zhang <zhangyuekai@foxmail.com>
Date: Mon, 21 Apr 2025 17:10:25 +0800
Subject: [PATCH] change padding side name

---
 egs/speech_llm/SPEECH2SPEECH/slam_omni/model.py | 4 ++--
 egs/speech_llm/SPEECH2SPEECH/slam_omni/train.py | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/egs/speech_llm/SPEECH2SPEECH/slam_omni/model.py b/egs/speech_llm/SPEECH2SPEECH/slam_omni/model.py
index 22f627ecc..fb3921ba3 100644
--- a/egs/speech_llm/SPEECH2SPEECH/slam_omni/model.py
+++ b/egs/speech_llm/SPEECH2SPEECH/slam_omni/model.py
@@ -59,7 +59,7 @@ class SPEECH_LLM(nn.Module):
         llm: nn.Module,
         encoder_projector: nn.Module,
         codec_lm: nn.Module = None,
-        use_flash_attention: bool = False,
+        codec_lm_padding_side: str = "left",
     ):
         super().__init__()
         self.encoder = encoder
@@ -74,7 +74,7 @@ class SPEECH_LLM(nn.Module):
                 self.codec_lm.config.hidden_size, self.codec_lm.config.vocab_size
             )
             self.loss_fct = torch.nn.CrossEntropyLoss()
-            self.codec_lm_padding_side = "left" if use_flash_attention else "right"
+            self.codec_lm_padding_side = codec_lm_padding_side
 
     def _merge_input_ids_with_speech_features(
         self, speech_features, inputs_embeds, input_ids, attention_mask, labels=None
diff --git a/egs/speech_llm/SPEECH2SPEECH/slam_omni/train.py b/egs/speech_llm/SPEECH2SPEECH/slam_omni/train.py
index 143c10c68..ef7e7a464 100755
--- a/egs/speech_llm/SPEECH2SPEECH/slam_omni/train.py
+++ b/egs/speech_llm/SPEECH2SPEECH/slam_omni/train.py
@@ -793,7 +793,7 @@ def run(rank, world_size, args):
         llm,
         encoder_projector,
         codec_lm,
-        params.use_flash_attn,
+        codec_lm_padding_side= "left" if params.use_flash_attn else "right",
     )
 
     if params.pretrained_model_path: