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

transformers/models/glm4v/image_processing_glm4v_fast.py

@@ -16,7 +16,7 @@
 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,
@@ -87,7 +87,7 @@ class Glm4vImageProcessorFast(BaseImageProcessorFast):
         images: 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/glm4v/modeling_glm4v.py

@@ -35,12 +35,18 @@ from ...integrations import use_kernel_forward_from_hub
 from ...masking_utils import create_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, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_torchdynamo_compiling
-from ...utils.generic import check_model_inputs, maybe_autocast
+from ...utils import (
+    TransformersKwargs,
+    auto_docstring,
+    can_return_tuple,
+    is_torchdynamo_compiling,
+    torch_compilable_check,
+)
+from ...utils.generic import check_model_inputs, is_flash_attention_requested, maybe_autocast
 from .configuration_glm4v import Glm4vConfig, Glm4vTextConfig, Glm4vVisionConfig
 
 
@@ -299,11 +305,11 @@ class Glm4vVisionAttention(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(
@@ -552,9 +558,9 @@ class Glm4vTextAttention(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,
@@ -695,6 +701,10 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
     config: Glm4vVisionConfig
     input_modalities = ("image", "video")
     _no_split_modules = ["Glm4vVisionBlock"]
+    _can_record_outputs = {
+        "hidden_states": Glm4vVisionBlock,
+        "attentions": Glm4vVisionAttention,
+    }
 
     def __init__(self, config) -> None:
         super().__init__(config)
@@ -753,13 +763,16 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
         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.
@@ -793,6 +806,7 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
                 hidden_states,
                 cu_seqlens=cu_seqlens,
                 position_embeddings=position_embeddings,
+                **kwargs,
             )
 
         hidden_states = self.post_layernorm(hidden_states)
@@ -803,8 +817,12 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
         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,
+        )
 
 
 @auto_docstring
@@ -1128,17 +1146,19 @@ class Glm4vModel(Glm4vPreTrainedModel):
 
             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
@@ -1147,26 +1167,36 @@ class Glm4vModel(Glm4vPreTrainedModel):
             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,
@@ -1195,18 +1225,19 @@ class Glm4vModel(Glm4vPreTrainedModel):
 
         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
@@ -1241,13 +1272,13 @@ class Glm4vModel(Glm4vPreTrainedModel):
             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)
@@ -1365,13 +1396,37 @@ class Glm4vForConditionalGeneration(Glm4vPreTrainedModel, 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
@@ -1405,23 +1460,25 @@ class Glm4vForConditionalGeneration(Glm4vPreTrainedModel, GenerationMixin):
 
         ```python
         >>> from PIL import Image
-        >>> import requests
+        >>> import httpx
+        >>> from io import BytesIO
         >>> from transformers import AutoProcessor, Glm4vForConditionalGeneration
 
-        >>> model = Glm4vForConditionalGeneration.from_pretrained("THUDM/GLM-4.1V-9B-Thinking")
-        >>> processor = AutoProcessor.from_pretrained("THUDM/GLM-4.1V-9B-Thinking")
+        >>> model = Glm4vForConditionalGeneration.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])
@@ -1498,7 +1555,7 @@ class Glm4vForConditionalGeneration(Glm4vPreTrainedModel, 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/modular_glm4v.py

@@ -29,12 +29,19 @@ from ...image_utils import ImageInput
 from ...masking_utils import create_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
-from ...modeling_outputs import BaseModelOutputWithPast
+from ...modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling
 from ...modeling_rope_utils import RopeParameters
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...tokenization_utils_base import PreTokenizedInput, TextInput
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_torchdynamo_compiling, logging
+from ...utils import (
+    TransformersKwargs,
+    auto_docstring,
+    can_return_tuple,
+    is_torchdynamo_compiling,
+    logging,
+    torch_compilable_check,
+)
 from ...utils.generic import check_model_inputs, maybe_autocast
 from ...video_utils import VideoInput
 from ..glm4.modeling_glm4 import Glm4MLP, Glm4RMSNorm, Glm4RotaryEmbedding, eager_attention_forward
@@ -193,14 +200,15 @@ class Glm4vTextConfig(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*):
+            The id of the padding token.
+
 
     ```python
     >>> from transformers import Glm4vTextModel, Glm4vConfig
@@ -224,8 +232,8 @@ class Glm4vTextConfig(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"]),
@@ -246,9 +254,9 @@ class Glm4vTextConfig(PreTrainedConfig):
         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,
+        pad_token_id: int | None = None,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -269,10 +277,9 @@ class Glm4vTextConfig(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)
 
 
 class Glm4vConfig(PreTrainedConfig):
@@ -303,6 +310,8 @@ class Glm4vConfig(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 Glm4vForConditionalGeneration, Glm4vConfig
@@ -331,6 +340,7 @@ class Glm4vConfig(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):
@@ -349,6 +359,7 @@ class Glm4vConfig(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)
 
@@ -616,9 +627,9 @@ class Glm4vTextAttention(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,
@@ -714,6 +725,10 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
     config: Glm4vVisionConfig
     input_modalities = ("image", "video")
     _no_split_modules = ["Glm4vVisionBlock"]
+    _can_record_outputs = {
+        "hidden_states": Glm4vVisionBlock,
+        "attentions": Glm4vVisionAttention,
+    }
 
     def __init__(self, config) -> None:
         super().__init__(config)
@@ -772,13 +787,16 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
         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.
@@ -812,6 +830,7 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
                 hidden_states,
                 cu_seqlens=cu_seqlens,
                 position_embeddings=position_embeddings,
+                **kwargs,
             )
 
         hidden_states = self.post_layernorm(hidden_states)
@@ -822,8 +841,12 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
         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 Glm4vTextModel(Qwen2_5_VLTextModel):
@@ -1121,17 +1144,19 @@ class Glm4vModel(Qwen2_5_VLModel):
 
             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
@@ -1140,10 +1165,14 @@ class Glm4vModel(Qwen2_5_VLModel):
             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
+
+        return vision_outputs
 
     def get_placeholder_mask(
         self,
@@ -1172,18 +1201,19 @@ class Glm4vModel(Qwen2_5_VLModel):
 
         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
@@ -1218,13 +1248,13 @@ class Glm4vModel(Qwen2_5_VLModel):
             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)
@@ -1331,23 +1361,25 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
 
         ```python
         >>> from PIL import Image
-        >>> import requests
+        >>> import httpx
+        >>> from io import BytesIO
         >>> from transformers import AutoProcessor, Glm4vForConditionalGeneration
 
-        >>> model = Glm4vForConditionalGeneration.from_pretrained("THUDM/GLM-4.1V-9B-Thinking")
-        >>> processor = AutoProcessor.from_pretrained("THUDM/GLM-4.1V-9B-Thinking")
+        >>> model = Glm4vForConditionalGeneration.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])
@@ -1424,7 +1456,7 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
             **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: