transformers 5.0.0__py3-none-any.whl → 5.0.0rc0__py3-none-any.whl

transformers/models/glm4v/modular_glm4v.py

@@ -1,3 +1,4 @@
+# coding=utf-8
 # Copyright 2025 The ZhipuAI Inc. team and HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -13,6 +14,7 @@
 # limitations under the License.
 import itertools
 from collections.abc import Callable
+from typing import Optional, Union
 
 import numpy as np
 import torch
@@ -20,7 +22,6 @@ import torch.nn as nn
 import torch.nn.functional as F
 from torch.nn import LayerNorm
 
-from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...configuration_utils import PreTrainedConfig
@@ -29,20 +30,13 @@ 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, BaseModelOutputWithPooling
+from ...modeling_outputs import BaseModelOutputWithPast
 from ...modeling_rope_utils import RopeParameters
-from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
 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,
-    torch_compilable_check,
-)
-from ...utils.generic import check_model_inputs, maybe_autocast
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_torchdynamo_compiling, logging
+from ...utils.generic import check_model_inputs
 from ...video_utils import VideoInput
 from ..glm4.modeling_glm4 import Glm4MLP, Glm4RMSNorm, Glm4RotaryEmbedding, eager_attention_forward
 from ..qwen2_5_vl.modeling_qwen2_5_vl import (
@@ -200,6 +194,8 @@ 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*):
@@ -240,19 +236,20 @@ class Glm4vTextConfig(PreTrainedConfig):
 
     def __init__(
         self,
-        vocab_size: int | None = 151552,
-        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,
-        initializer_range: float | None = 0.02,
-        rms_norm_eps: int | None = 1e-05,
-        use_cache: bool | None = True,
-        attention_dropout: float | None = 0.0,
-        rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
+        vocab_size: Optional[int] = 151552,
+        hidden_size: Optional[int] = 4096,
+        intermediate_size: Optional[int] = 13696,
+        num_hidden_layers: Optional[int] = 40,
+        num_attention_heads: Optional[int] = 32,
+        num_key_value_heads: Optional[int] = 2,
+        hidden_act: Optional[str] = "silu",
+        max_position_embeddings: Optional[int] = 32768,
+        initializer_range: Optional[float] = 0.02,
+        rms_norm_eps: Optional[int] = 1e-05,
+        use_cache: Optional[bool] = True,
+        tie_word_embeddings: Optional[bool] = False,
+        attention_dropout: Optional[float] = 0.0,
+        rope_parameters: Optional[RopeParameters | dict[str, RopeParameters]] = None,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -274,7 +271,7 @@ class Glm4vTextConfig(PreTrainedConfig):
         self.attention_dropout = attention_dropout
         self.rope_parameters = rope_parameters
 
-        super().__init__(ignore_keys_at_rope_validation={"mrope_section"}, **kwargs)
+        super().__init__(tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope"}, **kwargs)
 
 
 class Glm4vConfig(PreTrainedConfig):
@@ -305,8 +302,6 @@ 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
@@ -335,7 +330,6 @@ 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):
@@ -354,7 +348,6 @@ 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)
 
@@ -414,7 +407,7 @@ class Glm4vVisionEmbeddings(nn.Module):
         self.num_patches = (self.image_size // self.patch_size) ** 2
         self.num_positions = self.num_patches
         self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
-        self.interpolated_method = "bicubic"
+        self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
 
     def forward(self, embeddings, lengths, image_shapes, h_coords, w_coords) -> torch.Tensor:
         """
@@ -433,45 +426,57 @@ class Glm4vVisionEmbeddings(nn.Module):
         # Get position embedding parameters
         pos_embed_weight = self.position_embedding.weight
         hidden_size = pos_embed_weight.shape[1]
+        total_seq = h_coords.shape[0]
         device = pos_embed_weight.device
 
-        # Convert inputs to tensors if needed
-        if isinstance(lengths, list):
-            lengths = torch.tensor(lengths, device=device, dtype=torch.long)
-
-        # Prepare 2D position embedding
-        orig_size_sq = pos_embed_weight.shape[0]
-        orig_size = int(orig_size_sq**0.5)
-        pos_embed_2d = (
-            pos_embed_weight.view(orig_size, orig_size, hidden_size)
-            .permute(2, 0, 1)
-            .unsqueeze(0)
-            .to(device=device, dtype=torch.float32)
-        )
+        # Move coordinates to correct device
+        h_coords, w_coords = h_coords.to(device), w_coords.to(device)
 
-        # Calculate target dimensions for each patch
-        target_h = torch.cat([image_shapes[i, 1].repeat(lengths[i]) for i in range(len(lengths))]).to(
-            device=device, dtype=torch.float32
-        )
-        target_w = torch.cat([image_shapes[i, 2].repeat(lengths[i]) for i in range(len(lengths))]).to(
-            device=device, dtype=torch.float32
-        )
+        # Handle empty sequence case
+        if total_seq == 0:
+            adapted_pos_embed = torch.empty(0, hidden_size, device=device, dtype=pos_embed_weight.dtype)
+        else:
+            # Convert inputs to tensors if needed
+            if isinstance(lengths, list):
+                lengths = torch.tensor(lengths, device=device, dtype=torch.long)
+            if not isinstance(image_shapes, torch.Tensor):
+                image_shapes = torch.tensor(image_shapes, device=device, dtype=torch.long)
+
+            # Prepare 2D position embedding
+            orig_size_sq = pos_embed_weight.shape[0]
+            orig_size = int(orig_size_sq**0.5)
+            pos_embed_2d = (
+                pos_embed_weight.view(orig_size, orig_size, hidden_size)
+                .permute(2, 0, 1)
+                .unsqueeze(0)
+                .to(device=device, dtype=torch.float32)
+            )
+
+            # Calculate target dimensions for each patch
+            target_h = torch.cat([image_shapes[i, 1].repeat(lengths[i]) for i in range(len(lengths))]).to(
+                device=device, dtype=torch.float32
+            )
+            target_w = torch.cat([image_shapes[i, 2].repeat(lengths[i]) for i in range(len(lengths))]).to(
+                device=device, dtype=torch.float32
+            )
 
-        # Normalize coordinates to [-1, 1] range for grid_sample
-        norm_w = ((w_coords + 0.5) / target_w) * 2 - 1
-        norm_h = ((h_coords + 0.5) / target_h) * 2 - 1
+            # Normalize coordinates to [-1, 1] range for grid_sample
+            h_coords = h_coords.to(device=device, dtype=torch.float32)
+            w_coords = w_coords.to(device=device, dtype=torch.float32)
+            norm_w = ((w_coords + 0.5) / target_w) * 2 - 1
+            norm_h = ((h_coords + 0.5) / target_h) * 2 - 1
 
-        # Create sampling grid
-        grid = torch.stack((norm_w, norm_h), dim=-1).unsqueeze(0).unsqueeze(2)
+            # Create sampling grid
+            grid = torch.stack((norm_w, norm_h), dim=-1).unsqueeze(0).unsqueeze(2)
 
-        # Perform bicubic interpolation
-        interpolated_embed_fp32 = F.grid_sample(
-            pos_embed_2d, grid, mode=self.interpolated_method, align_corners=False, padding_mode="border"
-        )
+            # Perform bicubic interpolation
+            interpolated_embed_fp32 = F.grid_sample(
+                pos_embed_2d, grid, mode="bicubic", align_corners=False, padding_mode="border"
+            )
 
-        # Reshape and convert back to original dtype
-        adapted_pos_embed_fp32 = interpolated_embed_fp32.squeeze(0).squeeze(-1).permute(1, 0)
-        adapted_pos_embed = adapted_pos_embed_fp32.to(pos_embed_weight.dtype).to(embeddings.device)
+            # Reshape and convert back to original dtype
+            adapted_pos_embed_fp32 = interpolated_embed_fp32.squeeze(0).squeeze(-1).permute(1, 0)
+            adapted_pos_embed = adapted_pos_embed_fp32.to(pos_embed_weight.dtype).to(embeddings.device)
 
         # Add adapted position encoding to embeddings
         embeddings = embeddings + adapted_pos_embed
@@ -496,32 +501,22 @@ class Glm4vVisionBlock(Qwen2_5_VLVisionBlock):
 
 
 class Glm4vTextRotaryEmbedding(Glm4RotaryEmbedding):
-    def __init__(self, config: Glm4vTextConfig, device=None):
-        super().__init__()
-        self.mrope_section = config.rope_parameters.get("mrope_section", [8, 12, 12])
-
+    # Ignore copy
     def forward(self, x, position_ids):
-        # In contrast to other models, GLM-V has different position ids for the grids
+        # In contrast to other models, GLM4V different position ids for the grids
         # So we expand the inv_freq to shape (3, ...)
         inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1)
         position_ids_expanded = position_ids[:, :, None, :].float()  # shape (3, bs, 1, positions)
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
-            freqs = self.apply_mrope(freqs, self.mrope_section)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
             sin = emb.sin() * self.attention_scaling
 
         return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
 
-    def apply_mrope(self, freqs, mrope_section):
-        section = mrope_section
-        chunks = freqs.split(section, dim=-1)
-        result = torch.cat([chunk[i % 3] for i, chunk in enumerate(chunks)], dim=-1)
-        return result
-
 
 def rotate_half_llm(x):
     """Rotates half the hidden dims of the input."""
@@ -530,14 +525,25 @@ def rotate_half_llm(x):
     return torch.stack((-x2, x1), dim=-1).flatten(-2)
 
 
-def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
-    """Applies Rotary Position Embedding to the query and key tensors.
+def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/).
+
+    Explanation:
+        Multimodal 3D rotary position embedding is an extension to 1D rotary position embedding. The input embedding
+        sequence contains vision (images / videos) embedding and text embedding or just contains text embedding. For
+        vision embedding part, we apply rotary position embedding on temporal, height and width dimension separately.
+        Here we split the channel dimension to 3 chunks for the temporal, height and width rotary position embedding.
+        For text embedding part, we just apply 1D rotary position embedding. The three rotary position index (temporal,
+        height and width) of text embedding is always the same, so the text embedding rotary position embedding has no
+        difference with modern LLMs.
 
     Args:
         q (`torch.Tensor`): The query tensor.
         k (`torch.Tensor`): The key tensor.
         cos (`torch.Tensor`): The cosine part of the rotary embedding.
         sin (`torch.Tensor`): The sine part of the rotary embedding.
+        mrope_section(`List(int)`):
+            Multimodal rope section is for channel dimension of temporal, height and width in rope calculation.
         unsqueeze_dim (`int`, *optional*, defaults to 1):
             The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
             sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
@@ -548,8 +554,13 @@ def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
     Returns:
         `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
     """
-    cos = cos.unsqueeze(unsqueeze_dim)
-    sin = sin.unsqueeze(unsqueeze_dim)
+    mrope_section = mrope_section * 2
+    cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
+        unsqueeze_dim
+    )
+    sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
+        unsqueeze_dim
+    )
 
     # Interleave them instead of usual shape
     cos = cos[..., : cos.shape[-1] // 2].repeat_interleave(2, dim=-1)
@@ -567,6 +578,7 @@ def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
     # Concatenate back to full shape
     q_embed = torch.cat([q_embed, q_pass], dim=-1)
     k_embed = torch.cat([k_embed, k_pass], dim=-1)
+
     return q_embed, k_embed
 
 
@@ -576,7 +588,7 @@ class Glm4vTextAttention(nn.Module):
     and "Generating Long Sequences with Sparse Transformers".
     """
 
-    def __init__(self, config: Glm4vTextConfig, layer_idx: int | None = None):
+    def __init__(self, config: Glm4vTextConfig, layer_idx: Optional[int] = None):
         super().__init__()
         self.config = config
         self.layer_idx = layer_idx
@@ -599,12 +611,12 @@ class Glm4vTextAttention(nn.Module):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
-        attention_mask: torch.Tensor | None = None,
-        past_key_values: Cache | None = None,
-        cache_position: torch.LongTensor | None = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
         bsz, q_len, _ = hidden_states.size()
 
         query_states = self.q_proj(hidden_states)
@@ -616,7 +628,9 @@ class Glm4vTextAttention(nn.Module):
         value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
 
         cos, sin = position_embeddings
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+        query_states, key_states = apply_multimodal_rotary_pos_emb(  # diff with Llama
+            query_states, key_states, cos, sin, self.rope_parameters["mrope_section"]
+        )
 
         if past_key_values is not None:
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
@@ -661,14 +675,14 @@ class Glm4vTextDecoderLayer(GradientCheckpointingLayer):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        use_cache: bool | None = False,
-        cache_position: torch.LongTensor | None = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
         **kwargs,
-    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
+    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
         residual = hidden_states
 
         hidden_states = self.input_layernorm(hidden_states)
@@ -709,21 +723,11 @@ class Glm4vPreTrainedModel(Qwen2_5_VLPreTrainedModel):
         "attentions": Glm4vTextAttention,
     }
 
-    def _init_weights(self, module):
-        PreTrainedModel._init_weights(self, module)
-        if isinstance(module, Glm4vVisionRotaryEmbedding):
-            inv_freq = 1.0 / (module.theta ** (torch.arange(0, module.dim, 2, dtype=torch.float) / module.dim))
-            init.copy_(module.inv_freq, inv_freq)
-
 
 class 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)
@@ -782,22 +786,20 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
         rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
         return rotary_pos_emb, pos_ids
 
-    @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.
+    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor) -> 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.
 
         Returns:
             `torch.Tensor`: hidden_states.
         """
         hidden_states = self.patch_embed(hidden_states)
         hidden_states = self.post_conv_layernorm(hidden_states)
+
         rotary_pos_emb, image_type_ids = self.rot_pos_emb(grid_thw)
         emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
         position_embeddings = (emb.cos(), emb.sin())
@@ -812,20 +814,13 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
         )
         cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
         seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
-        hidden_states = self.embeddings(
-            hidden_states,
-            seqlens,
-            grid_thw,
-            image_type_ids[:, 0].to(hidden_states.device),
-            image_type_ids[:, 1].to(hidden_states.device),
-        )
+        hidden_states = self.embeddings(hidden_states, seqlens, grid_thw, image_type_ids[:, 0], image_type_ids[:, 1])
 
         for blk in self.blocks:
             hidden_states = blk(
                 hidden_states,
                 cu_seqlens=cu_seqlens,
                 position_embeddings=position_embeddings,
-                **kwargs,
             )
 
         hidden_states = self.post_layernorm(hidden_states)
@@ -836,12 +831,8 @@ 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)
 
-        merged_hidden_states = self.merger(hidden_states)
-
-        return BaseModelOutputWithPooling(
-            last_hidden_state=hidden_states,
-            pooler_output=merged_hidden_states,
-        )
+        hidden_states = self.merger(hidden_states)
+        return hidden_states
 
 
 class Glm4vTextModel(Qwen2_5_VLTextModel):
@@ -859,15 +850,15 @@ class Glm4vTextModel(Qwen2_5_VLTextModel):
     @check_model_inputs
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        use_cache: bool | None = None,
-        cache_position: torch.LongTensor | None = None,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple | BaseModelOutputWithPast:
+    ) -> Union[tuple, BaseModelOutputWithPast]:
         if (input_ids is None) ^ (inputs_embeds is not None):
             raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
 
@@ -925,7 +916,7 @@ class Glm4vTextModel(Qwen2_5_VLTextModel):
             layer_outputs = decoder_layer(
                 hidden_states,
                 attention_mask=causal_mask,
-                position_ids=text_position_ids,
+                position_ids=position_ids,
                 past_key_values=past_key_values,
                 cache_position=cache_position,
                 position_embeddings=position_embeddings,
@@ -951,10 +942,10 @@ class Glm4vModel(Qwen2_5_VLModel):
 
     def get_rope_index(
         self,
-        input_ids: torch.LongTensor | None = None,
-        image_grid_thw: torch.LongTensor | None = None,
-        video_grid_thw: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
+        input_ids: Optional[torch.LongTensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        video_grid_thw: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """
         Calculate the 3D rope index based on image and video's temporal, height and width in LLM.
@@ -1139,19 +1130,17 @@ 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,
-        **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.
+        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = 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.
         """
         pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
         # reshape video_grid_thw -> [b, 3] -> [1, h, w] * frames
@@ -1160,21 +1149,17 @@ 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)
-        vision_outputs = self.visual(
-            pixel_values_videos, grid_thw=flattened_video_grid_thw, return_dict=True, **kwargs
-        )
+        video_embeds = self.visual(pixel_values_videos, grid_thw=flattened_video_grid_thw)
         split_sizes = (video_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
-        video_embeds = torch.split(vision_outputs.pooler_output, split_sizes)
-        vision_outputs.pooler_output = video_embeds
-
-        return vision_outputs
+        video_embeds = torch.split(video_embeds, split_sizes)
+        return video_embeds
 
     def get_placeholder_mask(
         self,
         input_ids: torch.LongTensor,
         inputs_embeds: torch.FloatTensor,
-        image_features: torch.FloatTensor | None = None,
-        video_features: torch.FloatTensor | None = None,
+        image_features: Optional[torch.FloatTensor] = None,
+        video_features: Optional[torch.FloatTensor] = None,
     ):
         """
         Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
@@ -1196,38 +1181,37 @@ 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:
-            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]}",
+        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]}"
             )
 
         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:
-            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]}",
+        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]}"
             )
+
         return special_image_mask, special_video_mask
 
     @auto_docstring
     @can_return_tuple
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        pixel_values: torch.Tensor | None = None,
-        pixel_values_videos: torch.FloatTensor | None = None,
-        image_grid_thw: torch.LongTensor | None = None,
-        video_grid_thw: torch.LongTensor | None = None,
-        rope_deltas: torch.LongTensor | None = None,
-        cache_position: torch.LongTensor | None = None,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        pixel_values_videos: Optional[torch.FloatTensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        video_grid_thw: Optional[torch.LongTensor] = None,
+        rope_deltas: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
         **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | Glm4vModelOutputWithPast:
+    ) -> Union[tuple, Glm4vModelOutputWithPast]:
         r"""
         image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
             The temporal, height and width of feature shape of each image in LLM.
@@ -1243,13 +1227,13 @@ 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, return_dict=True).pooler_output
+            image_embeds = self.get_image_features(pixel_values, image_grid_thw)
             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, return_dict=True).pooler_output
+            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw)
             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)
@@ -1328,20 +1312,20 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
 
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        labels: torch.LongTensor | None = None,
-        pixel_values: torch.Tensor | None = None,
-        pixel_values_videos: torch.FloatTensor | None = None,
-        image_grid_thw: torch.LongTensor | None = None,
-        video_grid_thw: torch.LongTensor | None = None,
-        cache_position: torch.LongTensor | None = None,
-        logits_to_keep: int | torch.Tensor = 0,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        pixel_values_videos: Optional[torch.FloatTensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        video_grid_thw: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
         **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | Glm4vCausalLMOutputWithPast:
+    ) -> Union[tuple, Glm4vCausalLMOutputWithPast]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
@@ -1356,8 +1340,7 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
 
         ```python
         >>> from PIL import Image
-        >>> import httpx
-        >>> from io import BytesIO
+        >>> import requests
         >>> from transformers import AutoProcessor, Glm4vForConditionalGeneration
 
         >>> model = Glm4vForConditionalGeneration.from_pretrained("THUDM/GLM-4.1V-9B-Thinking")
@@ -1373,8 +1356,7 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
             },
         ]
         >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
-        >>> with httpx.stream("GET", url) as response:
-        ...     image = Image.open(BytesIO(response.read()))
+        >>> image = Image.open(requests.get(url, stream=True).raw)
 
         >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
         >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])
@@ -1430,7 +1412,6 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
         pixel_values_videos=None,
         image_grid_thw=None,
         video_grid_thw=None,
-        is_first_iteration=False,
         **kwargs,
     ):
         # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
@@ -1447,14 +1428,13 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
             image_grid_thw=image_grid_thw,
             video_grid_thw=video_grid_thw,
             use_cache=use_cache,
-            is_first_iteration=is_first_iteration,
             **kwargs,
         )
 
         # GLM-4.1V position_ids are prepareed with rope_deltas in forward
         model_inputs["position_ids"] = None
 
-        if not is_first_iteration and use_cache:
+        if cache_position[0] != 0:
             model_inputs["pixel_values"] = None
             model_inputs["pixel_values_videos"] = None
 
@@ -1462,8 +1442,8 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
 
     def _get_image_nums_and_video_nums(
         self,
-        input_ids: torch.LongTensor | None,
-        inputs_embeds: torch.Tensor | None = None,
+        input_ids: Optional[torch.LongTensor],
+        inputs_embeds: Optional[torch.Tensor] = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """
         Get the number of images and videos for each sample to calculate the separation length of the sample tensor.
@@ -1529,6 +1509,20 @@ class Glm4vProcessorKwargs(Qwen2VLProcessorKwargs):
 
 
 class Glm4vProcessor(Qwen2VLProcessor):
+    r"""
+    Constructs a GLM-4V processor which wraps a GLM-4V image processor and a GLM-4 tokenizer into a single processor.
+    [`~Glm4vProcessor.__call__`] and [`~Glm4vProcessor.decode`] for more information.
+    Args:
+        image_processor ([`Glm4vProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`PreTrainedTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+        video_processor ([`Glm4vVideoProcessor`], *optional*):
+            The video processor is a required input.
+        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+            in a chat into a tokenizable string.
+    """
+
     def __init__(self, image_processor=None, tokenizer=None, video_processor=None, chat_template=None, **kwargs):
         super().__init__(image_processor, tokenizer, video_processor, chat_template=chat_template)
         self.image_token = "<|image|>" if not hasattr(tokenizer, "image_token") else tokenizer.image_token
@@ -1536,12 +1530,32 @@ class Glm4vProcessor(Qwen2VLProcessor):
 
     def __call__(
         self,
-        images: ImageInput | None = None,
-        text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] = None,
-        videos: VideoInput | None = None,
+        images: Optional[ImageInput] = None,
+        text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None,
+        videos: Optional[VideoInput] = None,
         **kwargs: Unpack[Glm4vProcessorKwargs],
     ) -> BatchFeature:
-        r"""
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to PreTrainedTokenizerFast's [`~PreTrainedTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text.
+
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            videos (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
+                tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported.
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+
         Returns:
             [`BatchFeature`]: A [`BatchFeature`] with the following fields: