transformers 5.0.0rc1__py3-none-any.whl → 5.0.0rc2__py3-none-any.whl

transformers/models/qwen3_omni_moe/modular_qwen3_omni_moe.py

@@ -62,7 +62,11 @@ from ..qwen2_5_omni.modeling_qwen2_5_omni import (
     Qwen2_5OmniThinkerForConditionalGeneration,
     SnakeBeta,
 )
-from ..qwen2_5_omni.processing_qwen2_5_omni import Qwen2_5OmniProcessor, Qwen2_5OmniProcessorKwargs
+from ..qwen2_5_omni.processing_qwen2_5_omni import (
+    Qwen2_5OmniProcessor,
+    Qwen2_5OmniProcessorKwargs,
+    SinusoidsPositionEmbedding,
+)
 from ..qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
 from ..qwen3.configuration_qwen3 import Qwen3Config
 from ..qwen3.modeling_qwen3 import (
@@ -91,6 +95,7 @@ from ..qwen3_vl_moe.modeling_qwen3_vl_moe import (
     Qwen3VLMoeTextRotaryEmbedding,
     Qwen3VLMoeVisionAttention,
     Qwen3VLMoeVisionModel,
+    Qwen3VLMoeVisionRotaryEmbedding,
 )
 
 
@@ -668,6 +673,7 @@ class Qwen3OmniMoeTalkerConfig(PreTrainedConfig):
         self.audio_start_token_id = audio_start_token_id
         self.vision_start_token_id = vision_start_token_id
         self.speaker_id = speaker_id
+        self.initializer_range = self.text_config.initializer_range
         super().__init__(**kwargs)
 
 
@@ -758,6 +764,7 @@ class Qwen3OmniMoeCode2WavConfig(PreTrainedConfig):
         upsampling_ratios=(2, 2),
         decoder_dim=1536,
         attention_dropout=0.0,
+        initializer_range=0.02,
         **kwargs,
     ):
         self.codebook_size = codebook_size
@@ -777,6 +784,7 @@ class Qwen3OmniMoeCode2WavConfig(PreTrainedConfig):
         self.upsampling_ratios = upsampling_ratios
         self.decoder_dim = decoder_dim
         self.attention_dropout = attention_dropout
+        self.initializer_range = initializer_range
         self.rope_parameters = rope_parameters
 
         super().__init__(**kwargs)
@@ -865,6 +873,7 @@ class Qwen3OmniMoeConfig(PreTrainedConfig):
         self.thinker_config = Qwen3OmniMoeThinkerConfig(**thinker_config)
         self.talker_config = Qwen3OmniMoeTalkerConfig(**talker_config)
         self.code2wav_config = Qwen3OmniMoeCode2WavConfig(**code2wav_config)
+        self.initializer_range = self.thinker_config.initializer_range
         self.enable_audio_output = enable_audio_output
         self.im_start_token_id = im_start_token_id
         self.im_end_token_id = im_end_token_id
@@ -900,6 +909,19 @@ class Qwen3OmniMoePreTrainedModel(Qwen2_5OmniPreTrainedModel, PreTrainedModel):
             init.normal_(module.experts.gate_up_proj, mean=0.0, std=std)
             init.normal_(module.experts.down_proj, mean=0.0, std=std)
             init.normal_(module.gate.weight, mean=0.0, std=std)
+        elif isinstance(module, Qwen3OmniMoeCode2Wav):
+            init.copy_(
+                module.code_offset,
+                torch.arange(module.config.num_quantizers).view(1, -1, 1) * module.config.codebook_size,
+            )
+        elif isinstance(module, SinusoidsPositionEmbedding):
+            log_timescale_increment = np.log(module.max_timescale) / (module.channels // 2 - 1)
+            inv_timescales = torch.exp(-log_timescale_increment * torch.arange(module.channels // 2).float())
+            scaled_time = torch.arange(module.length)[:, np.newaxis] * inv_timescales[np.newaxis, :]
+            init.copy_(module.positional_embedding, torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1))
+        elif isinstance(module, Qwen3OmniMoeVisionRotaryEmbedding):
+            inv_freq = 1.0 / (module.theta ** (torch.arange(0, module.dim, 2, dtype=torch.float) / module.dim))
+            init.copy_(module.inv_freq, inv_freq)
 
 
 class Qwen3OmniMoePreTrainedModelForConditionalGeneration(Qwen2_5OmniPreTrainedModelForConditionalGeneration):
@@ -1297,6 +1319,10 @@ class Qwen3OmniMoeVisionPatchMerger(nn.Module):
         return hidden
 
 
+class Qwen3OmniMoeVisionRotaryEmbedding(Qwen3VLMoeVisionRotaryEmbedding):
+    pass
+
+
 class Qwen3OmniMoeVisionEncoder(Qwen3VLMoeVisionModel):
     config: Qwen3OmniMoeVisionEncoderConfig
     _no_split_modules = ["Qwen3OmniMoeVisionBlock"]
@@ -1847,6 +1873,9 @@ class Qwen3OmniMoeTalkerModel(Qwen3VLMoeTextModel):
 
 
 class Qwen3OmniMoeTalkerForConditionalGeneration(Qwen3MoeForCausalLM):
+    _tied_weights_keys = {"codec_head": "model.codec_embedding.weight"}
+    _tp_plan = {"codec_head": "colwise_rep"}
+    _pp_plan = {"codec_head": (["hidden_states"], ["logits"])}
     config_class = Qwen3OmniMoeTalkerConfig
     base_model_prefix = "talker"
     _no_split_modules = ["Qwen3OmniMoeTalkerCodePredictorModelForConditionalGeneration"]
@@ -2033,18 +2062,31 @@ class Qwen3OmniMoeTalkerForConditionalGeneration(Qwen3MoeForCausalLM):
         return model_kwargs
 
     def prepare_inputs_for_generation(
-        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, cache_position=None, **kwargs
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        is_first_iteration=False,
+        **kwargs,
     ):
         hidden_states = kwargs.pop("hidden_states", None)
         inputs = super().prepare_inputs_for_generation(
-            input_ids, past_key_values, attention_mask, inputs_embeds, cache_position, **kwargs
+            input_ids,
+            past_key_values,
+            attention_mask,
+            inputs_embeds,
+            cache_position,
+            is_first_iteration=is_first_iteration,
+            **kwargs,
         )
 
         # Qwen3-Omni will prepare position ids in forward with deltas
         inputs["position_ids"] = None
 
         # TODO(raushan, gante): Refactor this part to a utility function
-        if cache_position[0] != 0:
+        if not is_first_iteration and kwargs.get("use_cache", True):
             input_ids = input_ids[:, -1:]
             generation_step = kwargs.get("generation_step")
             trailing_text_hidden = kwargs.get("trailing_text_hidden")
@@ -2337,6 +2379,8 @@ class Qwen3OmniMoeCode2WavDecoderBlock(Qwen3OmniMoePreTrainedModel):
 
         self.block = nn.ModuleList(block)
 
+        self.post_init()
+
     def forward(self, hidden, **kwargs):
         for block in self.block:
             hidden = block(hidden)

transformers/models/qwen3_vl/configuration_qwen3_vl.py

@@ -110,7 +110,7 @@ class Qwen3VLTextConfig(PreTrainedConfig):
             Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain
             a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE
             with longer `max_position_embeddings`.
-        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+        attention_bias (`bool`, *optional*, defaults to `False`):
             Whether to use a bias in the query, key, value and output projection layers during self-attention.
         attention_dropout (`float`, *optional*, defaults to 0.0):
             The dropout ratio for the attention probabilities.
@@ -197,13 +197,13 @@ class Qwen3VLConfig(PreTrainedConfig):
         vision_config (`Union[PreTrainedConfig, dict]`,  *optional*, defaults to `Qwen3VLVisionConfig`):
             The config object or dictionary of the vision backbone.
         image_token_id (`int`, *optional*, defaults to 151655):
-            The image token index to encode the image prompt.
+            The token id used as the placeholder for image inputs.
         video_token_id (`int`, *optional*, defaults to 151656):
-            The video token index to encode the image prompt.
+            The token id used as the placeholder for video inputs.
         vision_start_token_id (`int`, *optional*, defaults to 151652):
-            The start token index to encode the image prompt.
+            The token id that marks the start of a vision segment (image or video).
         vision_end_token_id (`int`, *optional*, defaults to 151653):
-            The end token index to encode the image prompt.
+            The token id that marks the end of a vision segment (image or video).
         tie_word_embeddings (`bool`, *optional*, defaults to `False`):
             Whether to tie the word embeddings.
 

transformers/models/qwen3_vl/modeling_qwen3_vl.py

@@ -27,6 +27,7 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
+from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
@@ -81,6 +82,8 @@ class Qwen3VLVisionRotaryEmbedding(nn.Module):
 
     def __init__(self, dim: int, theta: float = 10000.0) -> None:
         super().__init__()
+        self.dim = dim
+        self.theta = theta
         inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
         self.register_buffer("inv_freq", inv_freq, persistent=False)
 
@@ -202,8 +205,8 @@ class Qwen3VLVisionAttention(nn.Module):
         if self.config._attn_implementation != "eager":
             attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
-        if self.config._attn_implementation == "flash_attention_2":
-            # Flash Attention 2: Use cu_seqlens for variable length attention
+        if "flash" in self.config._attn_implementation:
+            # Flash Attention: Use cu_seqlens for variable length attention
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
             attn_output, _ = attention_interface(
                 self,
@@ -292,7 +295,7 @@ class Qwen3VLTextRotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
         self.mrope_section = config.rope_parameters.get("mrope_section", [24, 20, 20])
 
@@ -592,6 +595,12 @@ class Qwen3VLPreTrainedModel(PreTrainedModel):
         "attentions": Qwen3VLTextAttention,
     }
 
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        if isinstance(module, Qwen3VLVisionRotaryEmbedding):
+            inv_freq = 1.0 / (module.theta ** (torch.arange(0, module.dim, 2, dtype=torch.float) / module.dim))
+            init.copy_(module.inv_freq, inv_freq)
+
 
 class Qwen3VLVisionModel(Qwen3VLPreTrainedModel):
     config: Qwen3VLVisionConfig
@@ -632,6 +641,8 @@ class Qwen3VLVisionModel(Qwen3VLPreTrainedModel):
 
         self.gradient_checkpointing = False
 
+        self.post_init()
+
     def rot_pos_emb(self, grid_thw: torch.Tensor) -> torch.Tensor:
         merge_size = self.spatial_merge_size
 
@@ -1407,6 +1418,7 @@ class Qwen3VLForConditionalGeneration(Qwen3VLPreTrainedModel, GenerationMixin):
         pixel_values_videos=None,
         image_grid_thw=None,
         video_grid_thw=None,
+        is_first_iteration=False,
         **kwargs,
     ):
         # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
@@ -1423,6 +1435,7 @@ class Qwen3VLForConditionalGeneration(Qwen3VLPreTrainedModel, GenerationMixin):
             image_grid_thw=image_grid_thw,
             video_grid_thw=video_grid_thw,
             use_cache=use_cache,
+            is_first_iteration=is_first_iteration,
             **kwargs,
         )
 
@@ -1454,7 +1467,7 @@ class Qwen3VLForConditionalGeneration(Qwen3VLPreTrainedModel, GenerationMixin):
             text_positions = model_inputs["position_ids"][None, ...]
             model_inputs["position_ids"] = torch.cat([text_positions, vision_positions], dim=0)
 
-        if cache_position[0] != 0:
+        if not is_first_iteration and use_cache:
             model_inputs["pixel_values"] = None
             model_inputs["pixel_values_videos"] = None
 

transformers/models/qwen3_vl/modular_qwen3_vl.py

@@ -22,6 +22,7 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
+from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...configuration_utils import PreTrainedConfig
@@ -31,7 +32,7 @@ from ...masking_utils import create_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_outputs import BaseModelOutputWithPast
 from ...modeling_rope_utils import RopeParameters, dynamic_rope_update
-from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import ProcessingKwargs, Unpack
 from ...tokenization_utils_base import PreTokenizedInput, TextInput
 from ...utils import auto_docstring, can_return_tuple, logging
@@ -151,7 +152,7 @@ class Qwen3VLTextConfig(PreTrainedConfig):
             Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain
             a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE
             with longer `max_position_embeddings`.
-        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+        attention_bias (`bool`, *optional*, defaults to `False`):
             Whether to use a bias in the query, key, value and output projection layers during self-attention.
         attention_dropout (`float`, *optional*, defaults to 0.0):
             The dropout ratio for the attention probabilities.
@@ -238,13 +239,13 @@ class Qwen3VLConfig(PreTrainedConfig):
         vision_config (`Union[PreTrainedConfig, dict]`,  *optional*, defaults to `Qwen3VLVisionConfig`):
             The config object or dictionary of the vision backbone.
         image_token_id (`int`, *optional*, defaults to 151655):
-            The image token index to encode the image prompt.
+            The token id used as the placeholder for image inputs.
         video_token_id (`int`, *optional*, defaults to 151656):
-            The video token index to encode the image prompt.
+            The token id used as the placeholder for video inputs.
         vision_start_token_id (`int`, *optional*, defaults to 151652):
-            The start token index to encode the image prompt.
+            The token id that marks the start of a vision segment (image or video).
         vision_end_token_id (`int`, *optional*, defaults to 151653):
-            The end token index to encode the image prompt.
+            The token id that marks the end of a vision segment (image or video).
         tie_word_embeddings (`bool`, *optional*, defaults to `False`):
             Whether to tie the word embeddings.
 
@@ -488,6 +489,12 @@ class Qwen3VLPreTrainedModel(Qwen2VLPreTrainedModel):
         "attentions": Qwen3VLTextAttention,
     }
 
+    def _init_weights(self, module):
+        PreTrainedModel._init_weights(self, module)
+        if isinstance(module, Qwen3VLVisionRotaryEmbedding):
+            inv_freq = 1.0 / (module.theta ** (torch.arange(0, module.dim, 2, dtype=torch.float) / module.dim))
+            init.copy_(module.inv_freq, inv_freq)
+
 
 class Qwen3VLVisionModel(Qwen3VLPreTrainedModel):
     config: Qwen3VLVisionConfig
@@ -528,6 +535,8 @@ class Qwen3VLVisionModel(Qwen3VLPreTrainedModel):
 
         self.gradient_checkpointing = False
 
+        self.post_init()
+
     def rot_pos_emb(self, grid_thw: torch.Tensor) -> torch.Tensor:
         merge_size = self.spatial_merge_size
 
@@ -1190,6 +1199,7 @@ class Qwen3VLForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
         pixel_values_videos=None,
         image_grid_thw=None,
         video_grid_thw=None,
+        is_first_iteration=False,
         **kwargs,
     ):
         # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
@@ -1206,6 +1216,7 @@ class Qwen3VLForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
             image_grid_thw=image_grid_thw,
             video_grid_thw=video_grid_thw,
             use_cache=use_cache,
+            is_first_iteration=is_first_iteration,
             **kwargs,
         )
 
@@ -1237,7 +1248,7 @@ class Qwen3VLForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
             text_positions = model_inputs["position_ids"][None, ...]
             model_inputs["position_ids"] = torch.cat([text_positions, vision_positions], dim=0)
 
-        if cache_position[0] != 0:
+        if not is_first_iteration and use_cache:
             model_inputs["pixel_values"] = None
             model_inputs["pixel_values_videos"] = None
 
@@ -1393,9 +1404,9 @@ class Qwen3VLProcessor(Qwen2VLProcessor):
         **kwargs: Unpack[Qwen3VLProcessorKwargs],
     ) -> BatchFeature:
         """
-        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        Main method to prepare for the model one or several sequence(s) and image(s). This method forwards the `text`
         and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode
-        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwrags` arguments to
+        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwargs` arguments to
         Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.
 
         Args:
@@ -1407,7 +1418,7 @@ class Qwen3VLProcessor(Qwen2VLProcessor):
                 (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
                 `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
             videos (`np.ndarray`, `torch.Tensor`, `list[np.ndarray]`, `list[torch.Tensor]`):
-                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
+                The video or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
                 tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported.
             return_tensors (`str` or [`~utils.TensorType`], *optional*):
                 If set, will return tensors of a particular framework. Acceptable values are:

transformers/models/qwen3_vl/processing_qwen3_vl.py

@@ -99,9 +99,9 @@ class Qwen3VLProcessor(ProcessorMixin):
         **kwargs: Unpack[Qwen3VLProcessorKwargs],
     ) -> BatchFeature:
         """
-        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        Main method to prepare for the model one or several sequence(s) and image(s). This method forwards the `text`
         and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode
-        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwrags` arguments to
+        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwargs` arguments to
         Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.
 
         Args:
@@ -113,7 +113,7 @@ class Qwen3VLProcessor(ProcessorMixin):
                 (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
                 `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
             videos (`np.ndarray`, `torch.Tensor`, `list[np.ndarray]`, `list[torch.Tensor]`):
-                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
+                The video or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
                 tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported.
             return_tensors (`str` or [`~utils.TensorType`], *optional*):
                 If set, will return tensors of a particular framework. Acceptable values are: