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

transformers/models/glm4v_moe/configuration_glm4v_moe.py

@@ -61,8 +61,6 @@ class Glm4vMoeTextConfig(PreTrainedConfig):
         use_cache (`bool`, *optional*, defaults to `True`):
             Whether or not the model should return the last key/values attentions (not used by all models). Only
             relevant if `config.is_decoder=True`.
-        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
-            Whether the model's input and output word embeddings should be tied.
         rope_parameters (`RopeParameters`, *optional*):
             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
@@ -90,8 +88,15 @@ class Glm4vMoeTextConfig(PreTrainedConfig):
                                                                     \--k dense layers--/
         norm_topk_prob (`bool`, *optional*, defaults to `True`):
             Whether to normalize the topk probabilities.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        eos_token_id (`int`, *optional*):
+            End of stream token id.
+        bos_token_id (`int`, *optional*):
+            Beginning of stream token id.
         router_aux_loss_coef (`float`, *optional*, defaults to 0.0001):
             The aux loss factor for the loss.
+
     ```python
     >>> from transformers import Glm4vMoeTextModel, Glm4vMoeConfig
 
@@ -140,7 +145,6 @@ class Glm4vMoeTextConfig(PreTrainedConfig):
         initializer_range: float | None = 0.02,
         rms_norm_eps: int | None = 1e-5,
         use_cache: bool | None = True,
-        tie_word_embeddings: bool | None = False,
         rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
         attention_bias: bool | None = True,
         attention_dropout: float | None = 0.0,
@@ -153,9 +157,15 @@ class Glm4vMoeTextConfig(PreTrainedConfig):
         topk_group: int | None = 1,
         first_k_dense_replace: int | None = 1,
         norm_topk_prob: bool | None = True,
+        pad_token_id: int | None = None,
+        eos_token_id: int | None = None,
+        bos_token_id: int | None = None,
         router_aux_loss_coef: float | None = 0.0001,
         **kwargs,
     ):
+        self.pad_token_id = pad_token_id
+        self.eos_token_id = eos_token_id
+        self.bos_token_id = bos_token_id
         self.vocab_size = vocab_size
         self.max_position_embeddings = max_position_embeddings
         self.hidden_size = hidden_size
@@ -184,9 +194,7 @@ class Glm4vMoeTextConfig(PreTrainedConfig):
         self.first_k_dense_replace = first_k_dense_replace
         self.norm_topk_prob = norm_topk_prob
         self.router_aux_loss_coef = router_aux_loss_coef
-        super().__init__(
-            tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope_section"}, **kwargs
-        )
+        super().__init__(ignore_keys_at_rope_validation={"mrope_section"}, **kwargs)
 
 
 class Glm4vMoeVisionConfig(PreTrainedConfig):
@@ -310,6 +318,8 @@ class Glm4vMoeConfig(PreTrainedConfig):
             The video start token index to encode the start of video.
         video_end_token_id (`int`, *optional*, defaults to 151342):
             The video end token index to encode the end of video.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
 
     ```python
     >>> from transformers import Glm4vMoeForConditionalGeneration, Glm4vMoeConfig
@@ -338,6 +348,7 @@ class Glm4vMoeConfig(PreTrainedConfig):
         image_end_token_id=151340,
         video_start_token_id=151341,
         video_end_token_id=151342,
+        tie_word_embeddings=False,
         **kwargs,
     ):
         if isinstance(vision_config, dict):
@@ -356,6 +367,7 @@ class Glm4vMoeConfig(PreTrainedConfig):
         self.video_end_token_id = video_end_token_id
         self.image_start_token_id = image_start_token_id
         self.image_end_token_id = image_end_token_id
+        self.tie_word_embeddings = tie_word_embeddings
 
         super().__init__(**kwargs)
 

transformers/models/glm4v_moe/modeling_glm4v_moe.py

@@ -35,7 +35,7 @@ from ...integrations import use_experts_implementation, use_kernel_forward_from_
 from ...masking_utils import create_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
-from ...modeling_outputs import ModelOutput, MoeModelOutputWithPast
+from ...modeling_outputs import BaseModelOutputWithPooling, ModelOutput, MoeModelOutputWithPast
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
@@ -45,8 +45,9 @@ from ...utils import (
     can_return_tuple,
     is_grouped_mm_available,
     is_torchdynamo_compiling,
+    torch_compilable_check,
 )
-from ...utils.generic import OutputRecorder, check_model_inputs, maybe_autocast
+from ...utils.generic import OutputRecorder, check_model_inputs, is_flash_attention_requested, maybe_autocast
 from .configuration_glm4v_moe import Glm4vMoeConfig, Glm4vMoeTextConfig, Glm4vMoeVisionConfig
 
 
@@ -185,9 +186,9 @@ class Glm4vMoeTextAttention(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,
@@ -684,11 +685,11 @@ class Glm4vMoeVisionAttention(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(
@@ -766,6 +767,10 @@ class Glm4vMoeVisionModel(Glm4vMoePreTrainedModel):
     config: Glm4vMoeVisionConfig
     input_modalities = ("image", "video")
     _no_split_modules = ["Glm4vMoeVisionBlock"]
+    _can_record_outputs = {
+        "hidden_states": Glm4vMoeVisionBlock,
+        "attentions": Glm4vMoeVisionAttention,
+    }
 
     def __init__(self, config) -> None:
         super().__init__(config)
@@ -824,13 +829,16 @@ class Glm4vMoeVisionModel(Glm4vMoePreTrainedModel):
         rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
         return rotary_pos_emb, pos_ids
 
-    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
-        """
-        Args:
-            hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
-                The final hidden states of the model.
-            grid_thw (`torch.Tensor` of shape `(num_images_or_videos, 3)`):
-                The temporal, height and width of feature shape of each image in LLM.
+    @check_model_inputs
+    @auto_docstring
+    def forward(
+        self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs: Unpack[TransformersKwargs]
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
+            The final hidden states of the model.
+        grid_thw (`torch.Tensor` of shape `(num_images_or_videos, 3)`):
+            The temporal, height and width of feature shape of each image in LLM.
 
         Returns:
             `torch.Tensor`: hidden_states.
@@ -864,6 +872,7 @@ class Glm4vMoeVisionModel(Glm4vMoePreTrainedModel):
                 hidden_states,
                 cu_seqlens=cu_seqlens,
                 position_embeddings=position_embeddings,
+                **kwargs,
             )
 
         hidden_states = self.post_layernorm(hidden_states)
@@ -874,8 +883,12 @@ class Glm4vMoeVisionModel(Glm4vMoePreTrainedModel):
         hidden_states = hidden_states.permute(0, 3, 1, 2)
         hidden_states = self.downsample(hidden_states).view(-1, self.config.out_hidden_size)
 
-        hidden_states = self.merger(hidden_states)
-        return hidden_states
+        merged_hidden_states = self.merger(hidden_states)
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=hidden_states,
+            pooler_output=merged_hidden_states,
+        )
 
 
 class Glm4vMoeTextRotaryEmbedding(nn.Module):
@@ -988,7 +1001,7 @@ class Glm4vMoeTextModel(Glm4vMoePreTrainedModel):
         use_cache: bool | None = None,
         cache_position: torch.LongTensor | None = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> MoeModelOutputWithPast:
+    ) -> tuple | MoeModelOutputWithPast:
         if (input_ids is None) ^ (inputs_embeds is not None):
             raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
 
@@ -1302,17 +1315,19 @@ class Glm4vMoeModel(Glm4vMoePreTrainedModel):
 
             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)
         # reshape video_grid_thw -> [b, 3] -> [1, h, w] * frames
@@ -1321,26 +1336,36 @@ class Glm4vMoeModel(Glm4vMoePreTrainedModel):
             repeated_row = torch.tensor([1, h.item(), w.item()]).unsqueeze(0).repeat(t, 1)
             temp_frames_hw.append(repeated_row)
         flattened_video_grid_thw = torch.cat(temp_frames_hw, dim=0)
-        video_embeds = self.visual(pixel_values_videos, grid_thw=flattened_video_grid_thw)
+        vision_outputs = self.visual(
+            pixel_values_videos, grid_thw=flattened_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,
@@ -1369,18 +1394,19 @@ class Glm4vMoeModel(Glm4vMoePreTrainedModel):
 
         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
@@ -1415,13 +1441,13 @@ class Glm4vMoeModel(Glm4vMoePreTrainedModel):
             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, image_features=image_embeds)
             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, video_features=video_embeds)
             inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
@@ -1592,13 +1618,37 @@ class Glm4vMoeForConditionalGeneration(Glm4vMoePreTrainedModel, 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)
 
     @auto_docstring
     @check_model_inputs
@@ -1632,23 +1682,25 @@ class Glm4vMoeForConditionalGeneration(Glm4vMoePreTrainedModel, GenerationMixin)
 
         ```python
         >>> from PIL import Image
-        >>> import requests
+        >>> import httpx
+        >>> from io import BytesIO
         >>> from transformers import AutoProcessor, Glm4vMoeForConditionalGeneration
 
-        >>> model = Glm4vMoeForConditionalGeneration.from_pretrained("THUDM/GLM-4.1V-9B-Thinking")
-        >>> processor = AutoProcessor.from_pretrained("THUDM/GLM-4.1V-9B-Thinking")
+        >>> model = Glm4vMoeForConditionalGeneration.from_pretrained("zai-org/GLM-4.1V-9B-Thinking")
+        >>> processor = AutoProcessor.from_pretrained("zai-org/GLM-4.1V-9B-Thinking")
 
         >>> messages = [
             {
                 "role": "user",
                 "content": [
-                    {"type": "image"},
+                    {"type": "image", "url": "https://www.ilankelman.org/stopsigns/australia.jpg"},
                     {"type": "text", "text": "What is shown in this image?"},
                 ],
             },
         ]
         >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
-        >>> image = Image.open(requests.get(url, stream=True).raw)
+        >>> with httpx.stream("GET", url) as response:
+        ...     image = Image.open(BytesIO(response.read()))
 
         >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
         >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])
@@ -1738,7 +1790,7 @@ class Glm4vMoeForConditionalGeneration(Glm4vMoePreTrainedModel, GenerationMixin)
             **kwargs,
         )
 
-        # GLM-4.1V position_ids are prepareed with rope_deltas in forward
+        # GLM-V position_ids are prepared with rope_deltas in forward
         model_inputs["position_ids"] = None
 
         if not is_first_iteration and use_cache:

transformers/models/glm4v_moe/modular_glm4v_moe.py

@@ -22,7 +22,7 @@ from ...configuration_utils import PreTrainedConfig
 from ...masking_utils import create_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_outputs import MoeModelOutputWithPast
-from ...modeling_rope_utils import RopeParameters, RotaryEmbeddingConfigMixin
+from ...modeling_rope_utils import RopeParameters
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, logging
@@ -55,7 +55,7 @@ from ..qwen3_vl_moe.modeling_qwen3_vl_moe import (
 logger = logging.get_logger(__name__)
 
 
-class Glm4vMoeTextConfig(Glm4MoeConfig, RotaryEmbeddingConfigMixin):
+class Glm4vMoeTextConfig(Glm4MoeConfig):
     r"""
     This is the configuration class to store the configuration of a [`Glm4vMoeModel`]. It is used to instantiate a
     GLM-4.5V model according to the specified arguments, defining the model architecture. Instantiating a
@@ -95,8 +95,6 @@ class Glm4vMoeTextConfig(Glm4MoeConfig, RotaryEmbeddingConfigMixin):
         use_cache (`bool`, *optional*, defaults to `True`):
             Whether or not the model should return the last key/values attentions (not used by all models). Only
             relevant if `config.is_decoder=True`.
-        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
-            Whether the model's input and output word embeddings should be tied.
         rope_parameters (`RopeParameters`, *optional*):
             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
@@ -124,8 +122,15 @@ class Glm4vMoeTextConfig(Glm4MoeConfig, RotaryEmbeddingConfigMixin):
                                                                     \--k dense layers--/
         norm_topk_prob (`bool`, *optional*, defaults to `True`):
             Whether to normalize the topk probabilities.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        eos_token_id (`int`, *optional*):
+            End of stream token id.
+        bos_token_id (`int`, *optional*):
+            Beginning of stream token id.
         router_aux_loss_coef (`float`, *optional*, defaults to 0.0001):
             The aux loss factor for the loss.
+
     ```python
     >>> from transformers import Glm4vMoeTextModel, Glm4vMoeConfig
 
@@ -171,7 +176,6 @@ class Glm4vMoeTextConfig(Glm4MoeConfig, RotaryEmbeddingConfigMixin):
         initializer_range: float | None = 0.02,
         rms_norm_eps: int | None = 1e-5,
         use_cache: bool | None = True,
-        tie_word_embeddings: bool | None = False,
         rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
         attention_bias: bool | None = True,
         attention_dropout: float | None = 0.0,
@@ -184,9 +188,15 @@ class Glm4vMoeTextConfig(Glm4MoeConfig, RotaryEmbeddingConfigMixin):
         topk_group: int | None = 1,
         first_k_dense_replace: int | None = 1,
         norm_topk_prob: bool | None = True,
+        pad_token_id: int | None = None,
+        eos_token_id: int | None = None,
+        bos_token_id: int | None = None,
         router_aux_loss_coef: float | None = 0.0001,
         **kwargs,
     ):
+        self.pad_token_id = pad_token_id
+        self.eos_token_id = eos_token_id
+        self.bos_token_id = bos_token_id
         self.vocab_size = vocab_size
         self.max_position_embeddings = max_position_embeddings
         self.hidden_size = hidden_size
@@ -215,9 +225,7 @@ class Glm4vMoeTextConfig(Glm4MoeConfig, RotaryEmbeddingConfigMixin):
         self.first_k_dense_replace = first_k_dense_replace
         self.norm_topk_prob = norm_topk_prob
         self.router_aux_loss_coef = router_aux_loss_coef
-        PreTrainedConfig.__init__(
-            self, tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope_section"}, **kwargs
-        )
+        PreTrainedConfig.__init__(self, ignore_keys_at_rope_validation={"mrope_section"}, **kwargs)
 
 
 class Glm4vMoeConfig(Glm4vConfig):
@@ -248,6 +256,8 @@ class Glm4vMoeConfig(Glm4vConfig):
             The video start token index to encode the start of video.
         video_end_token_id (`int`, *optional*, defaults to 151342):
             The video end token index to encode the end of video.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
 
     ```python
     >>> from transformers import Glm4vMoeForConditionalGeneration, Glm4vMoeConfig
@@ -272,6 +282,7 @@ class Glm4vMoeConfig(Glm4vConfig):
         image_end_token_id=151340,
         video_start_token_id=151341,
         video_end_token_id=151342,
+        tie_word_embeddings=False,
         **kwargs,
     ):
         super().__init__()
@@ -310,9 +321,9 @@ class Glm4vMoeTextAttention(Glm4Attention):
             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,
@@ -404,7 +415,7 @@ class Glm4vMoeTextModel(Glm4vTextModel):
         use_cache: bool | None = None,
         cache_position: torch.LongTensor | None = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> MoeModelOutputWithPast:
+    ) -> tuple | MoeModelOutputWithPast:
         if (input_ids is None) ^ (inputs_embeds is not None):
             raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
 

transformers/models/glm_image/configuration_glm_image.py

@@ -150,8 +150,8 @@ class GlmImageTextConfig(PreTrainedConfig):
 
     Args:
         vocab_size (`int`, *optional*, defaults to 168064):
-            Vocabulary size of the GlmImage model. Defines the number of different tokens that can be represented by the
-            `inputs_ids` passed when calling [`GlmImageModel`]
+            Vocabulary size of the GlmImage model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`GlmImageModel`]
         hidden_size (`int`, *optional*, defaults to 4096):
             Dimension of the hidden representations.
         intermediate_size (`int`, *optional*, defaults to 13696):
@@ -169,7 +169,7 @@ class GlmImageTextConfig(PreTrainedConfig):
             paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`.
         hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
             The non-linear activation function (function or string) in the decoder.
-        max_position_embeddings (`int`, *optional*, defaults to 32768):
+        max_position_embeddings (`int`, *optional*, defaults to 131072):
             The maximum sequence length that this model might ever be used with.
         initializer_range (`float`, *optional*, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
@@ -178,19 +178,21 @@ class GlmImageTextConfig(PreTrainedConfig):
         use_cache (`bool`, *optional*, defaults to `True`):
             Whether or not the model should return the last key/values attentions (not used by all models). Only
             relevant if `config.is_decoder=True`.
-        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
-            Whether the model's input and output word embeddings should be tied.
         attention_dropout (`float`, *optional*, defaults to 0.0):
             The dropout ratio for the attention probabilities.
         rope_parameters (`RopeParameters`, *optional*):
             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`.
+        pad_token_id (`int`, *optional*, defaults to 167841):
+            The id of the padding token.
         vision_vocab_size (`int`, *optional*, defaults to 16512):
-            Vision vocabulary size of the GlmImage model. Defines the number of different tokens that can be represented
-            by the `inputs_ids` passed when calling [`GlmImageVisionModel`]
+            Vision vocabulary size of the GlmImage model. Defines the number of different tokens that can be
+            represented by the `inputs_ids` passed when calling [`GlmImageVisionModel`]
         attention_bias (`bool`, *optional*, defaults to `True`):
             Whether to add a bias to the queries, keys and values.
+        eos_token_id (`int`, *optional*, defaults to 16385):
+            The id of the end of sequence token.
 
     ```python
     >>> from transformers import GlmImageTextModel, GlmImageConfig
@@ -214,8 +216,8 @@ class GlmImageTextConfig(PreTrainedConfig):
         "layers.*.self_attn.k_proj": "colwise",
         "layers.*.self_attn.v_proj": "colwise",
         "layers.*.self_attn.o_proj": "rowwise",
-        "layers.*.mlp.gate_up_proj": "colwise_rep",  # we need to replicate here due to the `chunk` operation
-        "layers.*.mlp.down_proj": "rowwise_rep",  # we need to replicate here due to the `chunk` operation
+        "layers.*.mlp.gate_up_proj": "colwise_gather_output",  # we need to replicate here due to the `chunk` operation
+        "layers.*.mlp.down_proj": "rowwise_split_input",  # input is replicated due to the `chunk` operation
     }
     base_model_pp_plan = {
         "embed_tokens": (["input_ids"], ["inputs_embeds"]),
@@ -225,27 +227,25 @@ class GlmImageTextConfig(PreTrainedConfig):
 
     def __init__(
         self,
-        vocab_size: int | None = 168064,
+        vocab_size: int = 168064,
         hidden_size: int | None = 4096,
         intermediate_size: int | None = 13696,
         num_hidden_layers: int | None = 40,
         num_attention_heads: int | None = 32,
         num_key_value_heads: int | None = 2,
         hidden_act: str | None = "silu",
-        max_position_embeddings: int | None = 32768,
+        max_position_embeddings: int = 131072,
         initializer_range: float | None = 0.02,
         rms_norm_eps: int | None = 1e-05,
         use_cache: bool | None = True,
-        tie_word_embeddings: bool | None = False,
         attention_dropout: float | None = 0.0,
         rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
-        vision_vocab_size: int | None = 16512,
-        attention_bias: bool | None = True,
+        pad_token_id: int = 167841,
+        vision_vocab_size: int = 16512,
+        attention_bias: bool = True,
+        eos_token_id: int = 16385,
         **kwargs,
     ):
-        self.vocab_size = vocab_size
-        self.vision_vocab_size = vision_vocab_size
-        self.attention_bias = attention_bias
         self.vocab_size = vocab_size
         self.max_position_embeddings = max_position_embeddings
         self.hidden_size = hidden_size
@@ -264,10 +264,12 @@ class GlmImageTextConfig(PreTrainedConfig):
         self.use_cache = use_cache
         self.attention_dropout = attention_dropout
         self.rope_parameters = rope_parameters
+        self.pad_token_id = pad_token_id
 
-        super().__init__(
-            tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope_section"}, **kwargs
-        )
+        super().__init__(ignore_keys_at_rope_validation={"mrope_section"}, **kwargs)
+        self.vision_vocab_size = vision_vocab_size
+        self.attention_bias = attention_bias
+        self.eos_token_id = eos_token_id
 
 
 class GlmImageConfig(PreTrainedConfig):
@@ -293,6 +295,8 @@ class GlmImageConfig(PreTrainedConfig):
             The image start token index to encode the start of image.
         image_end_token_id (`int`, *optional*, defaults to 16385):
             The image end token index to encode the end of image.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
 
     ```python
     >>> from transformers import Glm4vForConditionalGeneration, Glm4vConfig
@@ -323,6 +327,7 @@ class GlmImageConfig(PreTrainedConfig):
         image_token_id=167855,
         image_start_token_id=16384,
         image_end_token_id=16385,
+        tie_word_embeddings: bool | None = False,
         **kwargs,
     ):
         if isinstance(vision_config, dict):
@@ -346,6 +351,7 @@ class GlmImageConfig(PreTrainedConfig):
         self.text_config = text_config
         self.vision_config = vision_config
         self.vq_config = vq_config
+        self.tie_word_embeddings = tie_word_embeddings
         super().__init__(**kwargs)
 
 

transformers/models/glm_image/image_processing_glm_image.py

@@ -144,7 +144,7 @@ class GlmImageImageProcessor(BaseImageProcessor):
             The merge size of the vision encoder to llm encoder.
     """
 
-    model_input_names = ["pixel_values", "image_grid_thw"]
+    model_input_names = ["pixel_values", "image_grid_thw", "images_per_sample"]
     valid_kwargs = GlmImageImageProcessorKwargs
 
     def __init__(