diff --git a/vllm_omni/model_executor/models/qwen3_tts/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py b/vllm_omni/model_executor/models/qwen3_tts/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py index 31abe254b..78fc965fe 100644 --- a/vllm_omni/model_executor/models/qwen3_tts/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py +++ b/vllm_omni/model_executor/models/qwen3_tts/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py @@ -197,12 +197,13 @@ class Qwen3TTSTokenizerV2CausalTransConvNet(nn.Module): self.conv = nn.ConvTranspose1d(in_channels, out_channels, kernel_size, stride=stride) pad = kernel_size - stride - self.left_pad = math.ceil(pad) - self.right_pad = pad = self.left_pad + self.left_pad = 0 + self.right_pad = int(pad) def forward(self, hidden_state): hidden_state = self.conv(hidden_state) - hidden_state = hidden_state[..., self.left_pad : hidden_state.shape[-1] - self.right_pad] + if self.right_pad > 0: + hidden_state = hidden_state[..., : hidden_state.shape[-1] - self.right_pad] return hidden_state.contiguous()