qwen-tts 0.0.5__tar.gz → 0.1.0__tar.gz

{qwen_tts-0.0.5 → qwen_tts-0.1.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: qwen-tts
-Version: 0.0.5
+Version: 0.1.0
 Summary: Qwen-TTS python package
 Author: Alibaba Qwen Team
 License: Apache-2.0
@@ -1380,4 +1380,10 @@ If you find our paper and code useful in your research, please consider giving a
 }
 ```
 
+
+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=QwenLM/Qwen3-TTS&type=Date)](https://star-history.com/#QwenLM/Qwen3-TTS&Date)
+
+
 <br>

{qwen_tts-0.0.5 → qwen_tts-0.1.0}/README.md

@@ -1352,4 +1352,10 @@ If you find our paper and code useful in your research, please consider giving a
 }
 ```
 
+
+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=QwenLM/Qwen3-TTS&type=Date)](https://star-history.com/#QwenLM/Qwen3-TTS&Date)
+
+
 <br>

{qwen_tts-0.0.5 → qwen_tts-0.1.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "qwen-tts"
-version = "0.0.5"
+version = "0.1.0"
 description = "Qwen-TTS python package"
 readme = "README.md"
 requires-python = ">=3.9"

{qwen_tts-0.0.5 → qwen_tts-0.1.0}/qwen_tts/core/models/modeling_qwen3_tts.py

@@ -470,7 +470,7 @@ class Qwen3TTSPreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = ["Qwen3TTSDecoderLayer"]
     _skip_keys_device_placement = "past_key_values"
-    _supports_flash_attn_2 = True
+    _supports_flash_attn = True
     _supports_sdpa = True
     _supports_cache_class = True
     _supports_static_cache = False
@@ -501,8 +501,7 @@ class Qwen3TTSTalkerTextPreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = []
     _skip_keys_device_placement = ["past_key_values"]
-    _supports_flash_attn_3 = True
-    _supports_flash_attn_2 = True
+    _supports_flash_attn = True
     _supports_sdpa = True
     _supports_flex_attn = True
     _supports_cache_class = True
@@ -1869,6 +1868,11 @@ class Qwen3TTSForConditionalGeneration(Qwen3TTSPreTrainedModel, GenerationMixin)
         weights_only=True,
         **kwargs,
     ):
+        # Hotfix to enable passing the correct attn implementation which is stored in the config but not in kwargs
+        requested_attn_implementation = kwargs.pop("attn_implementation", None)
+        if requested_attn_implementation is None and config and config._attn_implementation:
+            requested_attn_implementation = config._attn_implementation
+
         model = super().from_pretrained(
             pretrained_model_name_or_path,
             *model_args,
@@ -1881,6 +1885,7 @@ class Qwen3TTSForConditionalGeneration(Qwen3TTSPreTrainedModel, GenerationMixin)
             revision=revision,
             use_safetensors=use_safetensors,
             weights_only=weights_only,
+            attn_implementation=requested_attn_implementation,
             **kwargs,
         )
         if not local_files_only and not os.path.isdir(pretrained_model_name_or_path):

{qwen_tts-0.0.5 → qwen_tts-0.1.0}/qwen_tts/core/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py

@@ -198,12 +198,13 @@ 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()