transformers 5.0.0rc3__py3-none-any.whl → 5.1.0__py3-none-any.whl

transformers/models/qwen2_vl/modeling_qwen2_vl.py

@@ -34,7 +34,7 @@ from ...generation import GenerationMixin
 from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
-from ...modeling_outputs import BaseModelOutputWithPast, ModelOutput
+from ...modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling, ModelOutput
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
@@ -43,11 +43,10 @@ from ...utils import (
     auto_docstring,
     can_return_tuple,
     logging,
+    torch_compilable_check,
 )
-from ...utils.generic import maybe_autocast
-from ..qwen2.modeling_qwen2 import (
-    Qwen2RMSNorm,
-)
+from ...utils.generic import check_model_inputs, is_flash_attention_requested, maybe_autocast
+from ..qwen2.modeling_qwen2 import Qwen2RMSNorm
 from .configuration_qwen2_vl import Qwen2VLConfig, Qwen2VLTextConfig, Qwen2VLVisionConfig
 
 
@@ -382,11 +381,11 @@ class VisionAttention(nn.Module):
         key_states = key_states.transpose(0, 1).unsqueeze(0)
         value_states = value_states.transpose(0, 1).unsqueeze(0)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
-        if "flash" in self.config._attn_implementation:
+        if is_flash_attention_requested(self.config):
             # Flash Attention: Use cu_seqlens for variable length attention
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
             attn_output, _ = attention_interface(
@@ -548,9 +547,9 @@ class Qwen2VLAttention(nn.Module):
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
             key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -575,7 +574,7 @@ class Qwen2VLDecoderLayer(GradientCheckpointingLayer):
         super().__init__()
         self.hidden_size = config.hidden_size
 
-        if config.use_sliding_window and config._attn_implementation != "flash_attention_2":
+        if config.use_sliding_window and not is_flash_attention_requested(config):
             logger.warning_once(
                 f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
                 "unexpected results may be encountered."
@@ -680,6 +679,10 @@ class Qwen2VisionTransformerPretrainedModel(Qwen2VLPreTrainedModel):
     input_modalities = ("image", "video")
     _no_split_modules = ["Qwen2VLVisionBlock"]
     _input_embed_layer = "patch_embed"
+    _can_record_outputs = {
+        "hidden_states": Qwen2VLVisionBlock,
+        "attentions": VisionAttention,
+    }
 
     def __init__(self, config) -> None:
         super().__init__(config)
@@ -738,12 +741,13 @@ class Qwen2VisionTransformerPretrainedModel(Qwen2VLPreTrainedModel):
         rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
         return rotary_pos_emb
 
+    @check_model_inputs
     @auto_docstring
     def forward(
         self,
         hidden_states: torch.Tensor,
         grid_thw: torch.Tensor,
-        **kwargs,
+        **kwargs: Unpack[TransformersKwargs],
     ) -> torch.Tensor:
         r"""
         grid_thw (`torch.LongTensor` of shape `(num_images, 3)`):
@@ -772,7 +776,12 @@ class Qwen2VisionTransformerPretrainedModel(Qwen2VLPreTrainedModel):
                 **kwargs,
             )
 
-        return self.merger(hidden_states)
+        merged_hidden_states = self.merger(hidden_states)
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=hidden_states,
+            pooler_output=merged_hidden_states,
+        )
 
 
 @auto_docstring
@@ -1103,39 +1112,49 @@ class Qwen2VLModel(Qwen2VLPreTrainedModel):
 
             return position_ids, mrope_position_deltas
 
+    @can_return_tuple
+    @auto_docstring
     def get_video_features(
-        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: torch.LongTensor | None = None
-    ):
-        """
-        Encodes videos into continuous embeddings that can be forwarded to the language model.
-
-        Args:
-            pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
-                The tensors corresponding to the input videos.
-            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
-                The temporal, height and width of feature shape of each video in LLM.
+        self,
+        pixel_values_videos: torch.FloatTensor,
+        video_grid_thw: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The tensors corresponding to the input videos.
+        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+            The temporal, height and width of feature shape of each video in LLM.
         """
         pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
-        video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+        vision_outputs = self.visual(pixel_values_videos, grid_thw=video_grid_thw, return_dict=True, **kwargs)
         split_sizes = (video_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
-        video_embeds = torch.split(video_embeds, split_sizes)
-        return video_embeds
+        video_embeds = torch.split(vision_outputs.pooler_output, split_sizes)
+        vision_outputs.pooler_output = video_embeds
 
-    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: torch.LongTensor | None = None):
-        """
-        Encodes images into continuous embeddings that can be forwarded to the language model.
+        return vision_outputs
 
-        Args:
-            pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
-                The tensors corresponding to the input images.
-            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
-                The temporal, height and width of feature shape of each image in LLM.
+    @can_return_tuple
+    @auto_docstring
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_grid_thw: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The tensors corresponding to the input images.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
         """
         pixel_values = pixel_values.type(self.visual.dtype)
-        image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+        vision_outputs = self.visual(pixel_values, grid_thw=image_grid_thw, return_dict=True, **kwargs)
         split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
-        image_embeds = torch.split(image_embeds, split_sizes)
-        return image_embeds
+        image_embeds = torch.split(vision_outputs.pooler_output, split_sizes)
+        vision_outputs.pooler_output = image_embeds
+
+        return vision_outputs
 
     def get_placeholder_mask(
         self,
@@ -1163,18 +1182,19 @@ class Qwen2VLModel(Qwen2VLPreTrainedModel):
 
         n_image_tokens = special_image_mask.sum()
         special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
-        if image_features is not None and inputs_embeds[special_image_mask].numel() != image_features.numel():
-            raise ValueError(
-                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {image_features.shape[0]}"
+        if image_features is not None:
+            torch_compilable_check(
+                inputs_embeds[special_image_mask].numel() == image_features.numel(),
+                f"Image features and image tokens do not match, tokens: {n_image_tokens}, features: {image_features.shape[0]}",
             )
 
         n_video_tokens = special_video_mask.sum()
         special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
-        if video_features is not None and inputs_embeds[special_video_mask].numel() != video_features.numel():
-            raise ValueError(
-                f"Videos features and video tokens do not match: tokens: {n_video_tokens}, features {video_features.shape[0]}"
+        if video_features is not None:
+            torch_compilable_check(
+                inputs_embeds[special_video_mask].numel() == video_features.numel(),
+                f"Video features and video tokens do not match, tokens: {n_video_tokens}, features: {video_features.shape[0]}",
             )
-
         return special_image_mask, special_video_mask
 
     @auto_docstring
@@ -1216,7 +1236,7 @@ class Qwen2VLModel(Qwen2VLPreTrainedModel):
             inputs_embeds = self.get_input_embeddings()(input_ids)
 
         if pixel_values is not None:
-            image_embeds = self.get_image_features(pixel_values, image_grid_thw)
+            image_embeds = self.get_image_features(pixel_values, image_grid_thw, return_dict=True).pooler_output
             image_embeds = torch.cat(image_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
             image_mask, _ = self.get_placeholder_mask(
                 input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds
@@ -1224,7 +1244,7 @@ class Qwen2VLModel(Qwen2VLPreTrainedModel):
             inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
 
         if pixel_values_videos is not None:
-            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw)
+            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw, return_dict=True).pooler_output
             video_embeds = torch.cat(video_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
             _, video_mask = self.get_placeholder_mask(
                 input_ids, inputs_embeds=inputs_embeds, video_features=video_embeds
@@ -1291,13 +1311,37 @@ class Qwen2VLForConditionalGeneration(Qwen2VLPreTrainedModel, GenerationMixin):
     def set_input_embeddings(self, value):
         self.model.set_input_embeddings(value)
 
+    @auto_docstring
     def get_video_features(
-        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: torch.LongTensor | None = None
-    ):
-        return self.model.get_video_features(pixel_values_videos, video_grid_thw)
+        self,
+        pixel_values_videos: torch.FloatTensor,
+        video_grid_thw: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The tensors corresponding to the input videos.
+        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+            The temporal, height and width of feature shape of each video in LLM.
+        """
+        return self.model.get_video_features(
+            pixel_values_videos=pixel_values_videos, video_grid_thw=video_grid_thw, **kwargs
+        )
 
-    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: torch.LongTensor | None = None):
-        return self.model.get_image_features(pixel_values, image_grid_thw)
+    @auto_docstring
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_grid_thw: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The tensors corresponding to the input images.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
+        """
+        return self.model.get_image_features(pixel_values=pixel_values, image_grid_thw=image_grid_thw, **kwargs)
 
     @can_return_tuple
     @auto_docstring

transformers/models/qwen2_vl/video_processing_qwen2_vl.py

@@ -22,7 +22,7 @@ import math
 from typing import Optional
 
 import torch
-from torchvision.transforms.v2 import functional as F
+import torchvision.transforms.v2.functional as tvF
 
 from ...image_processing_utils import BatchFeature
 from ...image_utils import (
@@ -79,8 +79,6 @@ class Qwen2VLVideoProcessor(BaseVideoProcessor):
     do_rescale = True
     do_normalize = True
     do_convert_rgb = True
-    min_pixels = 128 * 28 * 28
-    max_pixels = 28 * 28 * 768
     patch_size = 14
     temporal_patch_size = 2
     merge_size = 2
@@ -105,7 +103,7 @@ class Qwen2VLVideoProcessor(BaseVideoProcessor):
         if "shortest_edge" not in size or "longest_edge" not in size:
             raise ValueError("size must contain 'shortest_edge' and 'longest_edge' keys.")
 
-        super().__init__(size=size, min_pixels=min_pixels, max_pixels=max_pixels, **kwargs)
+        super().__init__(size=size, **kwargs)
 
     def _further_process_kwargs(
         self,
@@ -128,7 +126,7 @@ class Qwen2VLVideoProcessor(BaseVideoProcessor):
         else:
             size = {**self.size}
 
-        return super()._further_process_kwargs(size=size, min_pixels=min_pixels, max_pixels=max_pixels, **kwargs)
+        return super()._further_process_kwargs(size=size, **kwargs)
 
     def sample_frames(
         self,
@@ -205,7 +203,7 @@ class Qwen2VLVideoProcessor(BaseVideoProcessor):
         videos: list["torch.Tensor"],
         do_resize: bool,
         size: SizeDict,
-        interpolation: Optional["F.InterpolationMode"],
+        interpolation: Optional["tvF.InterpolationMode"],
         do_rescale: bool,
         rescale_factor: float,
         do_normalize: bool,

transformers/models/qwen3/configuration_qwen3.py

@@ -83,6 +83,12 @@ class Qwen3Config(PreTrainedConfig):
             Attention pattern for each layer.
         attention_dropout (`float`, *optional*, defaults to 0.0):
             The dropout ratio for the attention probabilities.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*):
+            End of stream token id.
 
     ```python
     >>> from transformers import Qwen3Model, Qwen3Config
@@ -128,7 +134,7 @@ class Qwen3Config(PreTrainedConfig):
         hidden_act: str | None = "silu",
         max_position_embeddings: int | None = 32768,
         initializer_range: float | None = 0.02,
-        rms_norm_eps: int | None = 1e-6,
+        rms_norm_eps: float | None = 1e-6,
         use_cache: bool | None = True,
         tie_word_embeddings: bool | None = False,
         rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
@@ -138,6 +144,9 @@ class Qwen3Config(PreTrainedConfig):
         max_window_layers: int | None = 28,
         layer_types: list[str] | None = None,
         attention_dropout: float | None = 0.0,
+        pad_token_id: int | None = None,
+        bos_token_id: int | None = None,
+        eos_token_id: int | None = None,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -173,12 +182,13 @@ class Qwen3Config(PreTrainedConfig):
             ]
         layer_type_validation(self.layer_types, self.num_hidden_layers)
 
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.tie_word_embeddings = tie_word_embeddings
         self.rope_parameters = rope_parameters
 
-        super().__init__(
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )
+        super().__init__(**kwargs)
 
 
 __all__ = ["Qwen3Config"]

transformers/models/qwen3/modeling_qwen3.py

@@ -273,9 +273,9 @@ class Qwen3Attention(nn.Module):
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
             key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -453,7 +453,7 @@ class Qwen3Model(Qwen3PreTrainedModel):
 @auto_docstring
 class Qwen3ForCausalLM(Qwen3PreTrainedModel, GenerationMixin):
     _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
-    _tp_plan = {"lm_head": "colwise_rep"}
+    _tp_plan = {"lm_head": "colwise_gather_output"}
     _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
 
     def __init__(self, config):

transformers/models/qwen3/modular_qwen3.py

@@ -89,9 +89,9 @@ class Qwen3Attention(LlamaAttention):
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
             key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,

transformers/models/qwen3_moe/configuration_qwen3_moe.py

@@ -95,6 +95,12 @@ class Qwen3MoeConfig(PreTrainedConfig):
             Indicate which layers use Qwen3MoeMLP rather than Qwen3MoeSparseMoeBlock
             The list contains layer index, from 0 to num_layers-1 if we have num_layers layers
             If `mlp_only_layers` is empty, `decoder_sparse_step` is used to determine the sparsity.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*):
+            End of stream token id.
 
     ```python
     >>> from transformers import Qwen3MoeModel, Qwen3MoeConfig
@@ -112,15 +118,18 @@ class Qwen3MoeConfig(PreTrainedConfig):
     model_type = "qwen3_moe"
     keys_to_ignore_at_inference = ["past_key_values"]
 
+    attribute_map = {
+        "num_experts": "num_local_experts",
+    }
+
     # Default tensor parallel plan for base model `Qwen3Moe`
     base_model_tp_plan = {
         "layers.*.self_attn.q_proj": "colwise",
         "layers.*.self_attn.k_proj": "colwise",
         "layers.*.self_attn.v_proj": "colwise",
         "layers.*.self_attn.o_proj": "rowwise",
-        "layers.*.mlp.experts.gate_up_proj": "local_rowwise",
-        "layers.*.mlp.experts.down_proj": "local_rowwise",
-        "layers.*.mlp.experts": "gather",
+        "layers.*.mlp.experts.gate_up_proj": "packed_colwise",
+        "layers.*.mlp.experts.down_proj": "rowwise",
         "layers.*.mlp.gate_proj": "colwise",
         "layers.*.mlp.up_proj": "colwise",
         "layers.*.mlp.down_proj": "rowwise",
@@ -158,6 +167,9 @@ class Qwen3MoeConfig(PreTrainedConfig):
         output_router_logits: bool | None = False,
         router_aux_loss_coef: float | None = 0.001,
         mlp_only_layers: bool | None = None,
+        pad_token_id: int | None = None,
+        bos_token_id: int | None = None,
+        eos_token_id: int | None = None,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -188,10 +200,11 @@ class Qwen3MoeConfig(PreTrainedConfig):
         self.router_aux_loss_coef = router_aux_loss_coef
         self.mlp_only_layers = [] if mlp_only_layers is None else mlp_only_layers
 
-        super().__init__(
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.tie_word_embeddings = tie_word_embeddings
+        super().__init__(**kwargs)
 
 
 __all__ = ["Qwen3MoeConfig"]

transformers/models/qwen3_moe/modeling_qwen3_moe.py

@@ -177,9 +177,9 @@ class Qwen3MoeAttention(nn.Module):
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
             key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -620,7 +620,7 @@ def load_balancing_loss_func(
 @auto_docstring
 class Qwen3MoeForCausalLM(Qwen3MoePreTrainedModel, GenerationMixin):
     _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
-    _tp_plan = {"lm_head": "colwise_rep"}
+    _tp_plan = {"lm_head": "colwise_gather_output"}
     _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
 
     def __init__(self, config):

transformers/models/qwen3_next/configuration_qwen3_next.py

@@ -107,6 +107,12 @@ class Qwen3NextConfig(PreTrainedConfig):
             If `mlp_only_layers` is empty, `decoder_sparse_step` is used to determine the sparsity.
         layer_types (`list[str]`, *optional*):
             Types of each layer (attention or linear).
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*):
+            End of stream token id.
 
     ```python
     >>> from transformers import Qwen3NextModel, Qwen3NextConfig
@@ -130,9 +136,8 @@ class Qwen3NextConfig(PreTrainedConfig):
         "layers.*.self_attn.k_proj": "colwise",
         "layers.*.self_attn.v_proj": "colwise",
         "layers.*.self_attn.o_proj": "rowwise",
-        "layers.*.mlp.experts.gate_up_proj": "local_rowwise",
-        "layers.*.mlp.experts.down_proj": "local_rowwise",
-        "layers.*.mlp.experts": "gather",
+        "layers.*.mlp.experts.gate_up_proj": "packed_colwise",
+        "layers.*.mlp.experts.down_proj": "rowwise",
         "layers.*.mlp.shared_expert.gate_proj": "colwise",
         "layers.*.mlp.shared_expert.up_proj": "colwise",
         "layers.*.mlp.shared_expert.down_proj": "rowwise",
@@ -179,8 +184,15 @@ class Qwen3NextConfig(PreTrainedConfig):
         router_aux_loss_coef: float | None = 0.001,
         mlp_only_layers: list[int] | None = [],
         layer_types: list[str] | None = None,
+        pad_token_id: int | None = None,
+        bos_token_id: int | None = None,
+        eos_token_id: int | None = None,
         **kwargs,
     ):
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.tie_word_embeddings = tie_word_embeddings
         self.vocab_size = vocab_size
         self.max_position_embeddings = max_position_embeddings
         self.hidden_size = hidden_size
@@ -224,7 +236,7 @@ class Qwen3NextConfig(PreTrainedConfig):
         self.output_router_logits = output_router_logits
         self.router_aux_loss_coef = router_aux_loss_coef
         self.mlp_only_layers = mlp_only_layers
-        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+        super().__init__(**kwargs)
 
 
 __all__ = ["Qwen3NextConfig"]

transformers/models/qwen3_next/modeling_qwen3_next.py

@@ -401,9 +401,9 @@ class Qwen3NextAttention(nn.Module):
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
             key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -975,7 +975,7 @@ class Qwen3NextPreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = ["Qwen3NextDecoderLayer"]
     _skip_keys_device_placement = "past_key_values"
-    _supports_flash_attn_2 = True
+    _supports_flash_attn = True
     _supports_sdpa = True
     _keys_to_ignore_on_load_unexpected = [r"^mtp.*"]
     _can_record_outputs = {
@@ -1176,7 +1176,7 @@ def load_balancing_loss_func(
 @auto_docstring
 class Qwen3NextForCausalLM(Qwen3NextPreTrainedModel, GenerationMixin):
     _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
-    _tp_plan = {"lm_head": "colwise_rep"}
+    _tp_plan = {"lm_head": "colwise_gather_output"}
     _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
 
     def __init__(self, config):

transformers/models/qwen3_next/modular_qwen3_next.py

@@ -256,9 +256,9 @@ class Qwen3NextAttention(Qwen3MoeAttention):
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
             key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -726,7 +726,7 @@ class Qwen3NextPreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = ["Qwen3NextDecoderLayer"]
     _skip_keys_device_placement = "past_key_values"
-    _supports_flash_attn_2 = True
+    _supports_flash_attn = True
     _supports_sdpa = True
     _keys_to_ignore_on_load_unexpected = [r"^mtp.*"]
     _can_record_outputs = {