# Copyright 2025 The ZhipuAI Inc. team and HuggingFace Inc. team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import itertools from collections.abc import Callable import numpy as np import torch 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 from ...feature_extraction_utils import BatchFeature 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_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, logging, torch_compilable_check, ) from ...utils.generic import maybe_autocast, merge_with_config_defaults from ...utils.output_capturing import capture_outputs 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 ( Qwen2_5_VisionPatchEmbed, Qwen2_5_VisionRotaryEmbedding, Qwen2_5_VLCausalLMOutputWithPast, Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLMLP, Qwen2_5_VLModelOutputWithPast, Qwen2_5_VLPreTrainedModel, Qwen2_5_VLTextModel, Qwen2_5_VLVisionAttention, Qwen2_5_VLVisionBlock, ) from ..qwen2_vl.modeling_qwen2_vl import Qwen2VLModel from ..qwen2_vl.processing_qwen2_vl import ( Qwen2VLProcessor, Qwen2VLProcessorKwargs, ) logger = logging.get_logger(__name__) class Glm4vVisionConfig(PreTrainedConfig): r""" This is the configuration class to store the configuration of a [`Glm4vVisionModel`]. It is used to instantiate an Glm4vVisionModel model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of GLM-4.1V-9B-Thinking [THUDM/GLM-4.1V-9B-Thinking](https://huggingface.co/THUDM/GLM-4.1V-9B-Thinking). Args: depth (`int`, *optional*, defaults to 24): Number of layers (depth) in the model. hidden_size (`int`, *optional*, defaults to 1536): Dimensionality of the encoder layers and the pooler layer. hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"` are supported. attention_bias (`bool`, *optional*, defaults to `False`): Whether to add a bias to the queries, keys and values. attention_dropout (`float`, *optional*, defaults to 0.0): Dropout probability for attention weights. num_heads (``, *optional*, defaults to 12): in_channels (``, *optional*, defaults to 3): image_size (`int` or `list[int]`, *optional*, defaults to 336): The size (resolution) of each image. patch_size (`int`, *optional*, defaults to 14): The size (resolution) of each patch. rms_norm_eps (`float`, *optional*, defaults to 1e-05): The epsilon used by the rms normalization layers. spatial_merge_size (`int`, *optional*, defaults to 2): The size used for merging spatial dimensions. temporal_patch_size (`int`, *optional*, defaults to 2): The size used for patches along the temporal dimension. out_hidden_size (`int`, *optional*, defaults to 4096): The output hidden size of the vision model. intermediate_size (`int`, *optional*, defaults to 13696): Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. Example: ```python >>> from transformers import Glm4vVisionConfig, Glm4vVisionModel >>> # Initializing a Glm4vVisionConfig GLM-4.1V-9B style configuration >>> configuration = Glm4vVisionConfig() >>> # Initializing a model (with random weights) from the GLM-4.1V-9B configuration >>> model = Glm4vVisionModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "glm4v_vision" base_config_key = "vision_config" def __init__( self, depth=24, hidden_size=1536, hidden_act="silu", attention_bias=False, attention_dropout=0.0, num_heads=12, in_channels=3, image_size=336, patch_size=14, rms_norm_eps=1e-05, spatial_merge_size=2, temporal_patch_size=2, out_hidden_size=4096, intermediate_size=13696, initializer_range=0.02, **kwargs, ): super().__init__(**kwargs) self.depth = depth self.hidden_size = hidden_size self.hidden_act = hidden_act self.num_heads = num_heads self.in_channels = in_channels self.image_size = image_size self.patch_size = patch_size self.spatial_merge_size = spatial_merge_size self.temporal_patch_size = temporal_patch_size self.out_hidden_size = out_hidden_size self.intermediate_size = intermediate_size self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.attention_bias = attention_bias self.attention_dropout = attention_dropout class Glm4vTextConfig(PreTrainedConfig): r""" This is the configuration class to store the configuration of a [`Glm4vModel`]. It is used to instantiate a GLM-4.1V model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of GLM-4.1V-9B-Thinking [THUDM/GLM-4.1V-9B-Thinking](https://huggingface.co/THUDM/GLM-4.1V-9B-Thinking). Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PreTrainedConfig`] for more information. Args: vocab_size (`int`, *optional*, defaults to 151552): Vocabulary size of the Glm4v model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`Glm4vModel`] hidden_size (`int`, *optional*, defaults to 4096): Dimension of the hidden representations. intermediate_size (`int`, *optional*, defaults to 13696): Dimension of the MLP representations. num_hidden_layers (`int`, *optional*, defaults to 40): Number of hidden layers in the Transformer encoder. num_attention_heads (`int`, *optional*, defaults to 32): Number of attention heads for each attention layer in the Transformer encoder. num_key_value_heads (`int`, *optional*, defaults to 2): This is the number of key_value heads that should be used to implement Grouped Query Attention. If `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details checkout [this 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): 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. rms_norm_eps (`float`, *optional*, defaults to 1e-05): The epsilon used by the rms normalization layers. 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`. 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 >>> # Initializing a GLM-4.1V style configuration >>> configuration = Glm4vConfig() >>> # Initializing a model from the GLM-4.1V style configuration >>> model = Glm4vTextModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "glm4v_text" base_config_key = "text_config" keys_to_ignore_at_inference = ["past_key_values"] # Default tensor parallel plan for base model `Glm4v` base_model_tp_plan = { "layers.*.self_attn.q_proj": "colwise", "layers.*.self_attn.k_proj": "colwise", "layers.*.self_attn.v_proj": "colwise", "layers.*.self_attn.o_proj": "rowwise", "layers.*.mlp.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"]), "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), "norm": (["hidden_states"], ["hidden_states"]), } 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, pad_token_id: int | None = None, **kwargs, ): self.vocab_size = vocab_size self.max_position_embeddings = max_position_embeddings self.hidden_size = hidden_size self.intermediate_size = intermediate_size self.num_hidden_layers = num_hidden_layers self.num_attention_heads = num_attention_heads # for backward compatibility if num_key_value_heads is None: num_key_value_heads = num_attention_heads self.num_key_value_heads = num_key_value_heads self.hidden_act = hidden_act self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.attention_dropout = attention_dropout self.rope_parameters = rope_parameters self.pad_token_id = pad_token_id super().__init__(ignore_keys_at_rope_validation={"mrope_section"}, **kwargs) class Glm4vConfig(PreTrainedConfig): r""" This is the configuration class to store the configuration of a [`Glm4vModel`]. It is used to instantiate a GLM-4.1V model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of GLM-4.1V-9B-Thinking [THUDM/GLM-4.1V-9B-Thinking](https://huggingface.co/THUDM/GLM-4.1V-9B-Thinking). Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PreTrainedConfig`] for more information. Args: text_config (`Union[PreTrainedConfig, dict]`, *optional*, defaults to `Glm4vTextConfig`): The config object or dictionary of the text backbone. vision_config (`Union[PreTrainedConfig, dict]`, *optional*, defaults to `Glm4vVisionConfig`): The config object or dictionary of the vision backbone. image_token_id (`int`, *optional*, defaults to 151343): The image token index to encode the image prompt. video_token_id (`int`, *optional*, defaults to 151344): The video token index to encode the image prompt. image_start_token_id (`int`, *optional*, defaults to 151339): The image start token index to encode the start of image. image_end_token_id (`int`, *optional*, defaults to 151340): The image end token index to encode the end of image. video_start_token_id (`int`, *optional*, defaults to 151341): 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 >>> # Initializing a GLM-4.1V style configuration >>> configuration = Glm4vConfig() >>> # Initializing a model from the GLM-4.1V style configuration >>> model = Glm4vForConditionalGeneration(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```""" model_type = "glm4v" sub_configs = {"vision_config": Glm4vVisionConfig, "text_config": Glm4vTextConfig} keys_to_ignore_at_inference = ["past_key_values"] def __init__( self, text_config=None, vision_config=None, image_token_id=151343, video_token_id=151344, image_start_token_id=151339, 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): self.vision_config = self.sub_configs["vision_config"](**vision_config) elif vision_config is None: self.vision_config = self.sub_configs["vision_config"]() if isinstance(text_config, dict): self.text_config = self.sub_configs["text_config"](**text_config) elif text_config is None: self.text_config = self.sub_configs["text_config"](**kwargs) self.image_token_id = image_token_id self.video_token_id = video_token_id self.video_start_token_id = video_start_token_id 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) # Will be used for both Text and Vision modalities class Glm4vRMSNorm(Glm4RMSNorm): pass class Glm4VisionMlp(Qwen2_5_VLMLP): def __init__(self, config, bias: bool = False): super().__init__(config, bias) self.intermediate_size = config.out_hidden_size class Glm4vVisionPatchEmbed(Qwen2_5_VisionPatchEmbed): def __init__(self, config: Glm4vVisionConfig) -> None: nn.Module.__init__(self) self.patch_size = config.patch_size self.temporal_patch_size = config.temporal_patch_size self.in_channels = config.in_channels self.embed_dim = config.hidden_size kernel_size = [self.temporal_patch_size, self.patch_size, self.patch_size] self.proj = nn.Conv3d(self.in_channels, self.embed_dim, kernel_size=kernel_size, stride=kernel_size) class Glm4vVisionRotaryEmbedding(Qwen2_5_VisionRotaryEmbedding): pass class Glm4vVisionPatchMerger(nn.Module): def __init__(self, dim: int, context_dim: int, hidden_act: str, bias: bool = False) -> None: super().__init__() self.proj = nn.Linear(dim, dim, bias=bias) self.post_projection_norm = LayerNorm(dim) self.gate_proj = nn.Linear(dim, context_dim, bias=bias) self.up_proj = nn.Linear(dim, context_dim, bias=bias) self.down_proj = nn.Linear(context_dim, dim, bias=bias) self.act1 = nn.GELU() self.act_fn = ACT2FN[hidden_act] def forward(self, hidden_state: torch.Tensor) -> torch.Tensor: hidden_state = self.proj(hidden_state) hidden_state = self.act1(self.post_projection_norm(hidden_state)) return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state)) class Glm4vVisionEmbeddings(nn.Module): def __init__(self, config: Glm4vVisionConfig): super().__init__() self.config = config self.embed_dim = config.hidden_size self.image_size = config.image_size self.patch_size = config.patch_size 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" def forward(self, embeddings, lengths, image_shapes, h_coords, w_coords) -> torch.Tensor: """ Forward pass with integrated position encoding adaptation using 2D interpolation. Args: embeddings: Input embeddings tensor lengths (torch.Tensor): Sequence lengths for each image in the batch. image_shapes (torch.Tensor): Tensor of shape [batch_size, 3] representing the image shapes (t, h, w). h_coords (torch.Tensor): Tensor of shape [total_seq] representing the h coordinate for each patch. w_coords (torch.Tensor): Tensor of shape [total_seq] representing the w coordinate for each patch. Returns: torch.Tensor: Embeddings with adapted position encoding added. """ # Get position embedding parameters pos_embed_weight = self.position_embedding.weight hidden_size = pos_embed_weight.shape[1] 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) ) # 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 # 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" ) # 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 return embeddings class Glm4vVisionAttention(Qwen2_5_VLVisionAttention): def __init__(self, config: Glm4vVisionConfig) -> None: super().__init__(config) self.attention_dropout = config.attention_dropout self.qkv = nn.Linear(config.hidden_size, config.hidden_size * 3, bias=config.attention_bias) self.proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False) class Glm4vVisionBlock(Qwen2_5_VLVisionBlock): def __init__(self, config) -> None: super().__init__(config) self.norm1 = Glm4vRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.norm2 = Glm4vRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.attn = Glm4vVisionAttention(config) self.mlp = Glm4VisionMlp(config, bias=False) class Glm4vTextRotaryEmbedding(Glm4RotaryEmbedding): def __init__(self, config: Glm4vTextConfig, device=None): super().__init__() self.mrope_section = config.rope_parameters.get("mrope_section", [8, 12, 12]) def forward(self, x, position_ids): # In contrast to other models, GLM-V has 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 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.""" x1 = x[..., 0::2] x2 = x[..., 1::2] 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. 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. 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 that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. 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) # Interleave them instead of usual shape cos = cos[..., : cos.shape[-1] // 2].repeat_interleave(2, dim=-1) sin = sin[..., : sin.shape[-1] // 2].repeat_interleave(2, dim=-1) # Keep half or full tensor for later concatenation rotary_dim = cos.shape[-1] q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:] k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:] # Apply rotary embeddings on the first half or full tensor q_embed = (q_rot * cos) + (rotate_half_llm(q_rot) * sin) k_embed = (k_rot * cos) + (rotate_half_llm(k_rot) * sin) # 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 class Glm4vTextAttention(nn.Module): """ Multi-headed attention from 'Attention Is All You Need' paper. and "Generating Long Sequences with Sparse Transformers". """ def __init__(self, config: Glm4vTextConfig, layer_idx: int | None = None): super().__init__() self.config = config self.layer_idx = layer_idx self.hidden_size = config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = self.hidden_size // self.num_heads self.num_key_value_heads = config.num_key_value_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.is_causal = True self.attention_dropout = config.attention_dropout self.rope_parameters = config.rope_parameters self.scaling = self.head_dim**-0.5 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) 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, **kwargs: Unpack[FlashAttentionKwargs], ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]: bsz, q_len, _ = hidden_states.size() query_states = self.q_proj(hidden_states) key_states = self.k_proj(hidden_states) value_states = self.v_proj(hidden_states) query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2) 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) if past_key_values is not None: 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 = ALL_ATTENTION_FUNCTIONS.get_interface( self.config._attn_implementation, eager_attention_forward ) attn_output, attn_weights = attention_interface( self, query_states, key_states, value_states, attention_mask, dropout=0.0 if not self.training else self.attention_dropout, scaling=self.scaling, **kwargs, ) attn_output = attn_output.reshape(bsz, q_len, -1).contiguous() attn_output = self.o_proj(attn_output) return attn_output, attn_weights class Glm4vTextMLP(Glm4MLP): pass class Glm4vTextDecoderLayer(GradientCheckpointingLayer): def __init__(self, config: Glm4vTextConfig, layer_idx: int): super().__init__() self.hidden_size = config.hidden_size self.self_attn = Glm4vTextAttention(config, layer_idx) self.mlp = Glm4vTextMLP(config) self.input_layernorm = Glm4vRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.post_attention_layernorm = Glm4vRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.post_self_attn_layernorm = Glm4vRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.post_mlp_layernorm = Glm4vRMSNorm(config.hidden_size, eps=config.rms_norm_eps) @auto_docstring 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, **kwargs, ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]: residual = hidden_states hidden_states = self.input_layernorm(hidden_states) # Self Attention hidden_states, _ = self.self_attn( hidden_states=hidden_states, position_embeddings=position_embeddings, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, use_cache=use_cache, cache_position=cache_position, **kwargs, ) hidden_states = self.post_self_attn_layernorm(hidden_states) hidden_states = residual + hidden_states # Fully Connected residual = hidden_states hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = self.post_mlp_layernorm(hidden_states) hidden_states = residual + hidden_states return hidden_states class Glm4vModelOutputWithPast(Qwen2_5_VLModelOutputWithPast): pass class Glm4vPreTrainedModel(Qwen2_5_VLPreTrainedModel): _no_split_modules = ["Glm4vTextDecoderLayer", "Glm4vVisionBlock"] _can_record_outputs = { "hidden_states": Glm4vTextDecoderLayer, "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) self.spatial_merge_size = config.spatial_merge_size self.patch_size = config.patch_size self.embeddings = Glm4vVisionEmbeddings(config) self.patch_embed = Glm4vVisionPatchEmbed(config) head_dim = config.hidden_size // config.num_heads self.rotary_pos_emb = Glm4vVisionRotaryEmbedding(head_dim // 2) self.blocks = nn.ModuleList([Glm4vVisionBlock(config) for _ in range(config.depth)]) self.merger = Glm4vVisionPatchMerger( dim=config.out_hidden_size, context_dim=config.intermediate_size, hidden_act=config.hidden_act ) self.post_conv_layernorm = Glm4vRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.downsample = nn.Conv2d( in_channels=config.hidden_size, out_channels=config.out_hidden_size, kernel_size=config.spatial_merge_size, stride=config.spatial_merge_size, ) self.post_layernorm = Glm4vRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.gradient_checkpointing = False self.post_init() def rot_pos_emb(self, grid_thw): pos_ids = [] for t, h, w in grid_thw: hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w) hpos_ids = hpos_ids.reshape( h // self.spatial_merge_size, self.spatial_merge_size, w // self.spatial_merge_size, self.spatial_merge_size, ) hpos_ids = hpos_ids.permute(0, 2, 1, 3) hpos_ids = hpos_ids.flatten() wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1) wpos_ids = wpos_ids.reshape( h // self.spatial_merge_size, self.spatial_merge_size, w // self.spatial_merge_size, self.spatial_merge_size, ) wpos_ids = wpos_ids.permute(0, 2, 1, 3) wpos_ids = wpos_ids.flatten() pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1)) pos_ids = torch.cat(pos_ids, dim=0) max_grid_size = grid_thw[:, 1:].max() rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size) rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1) return rotary_pos_emb, pos_ids @merge_with_config_defaults @capture_outputs @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. """ 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()) cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum( dim=0, # Select dtype based on the following factors: # - FA2 requires that cu_seqlens_q must have dtype int32 # - torch.onnx.export requires that cu_seqlens_q must have same dtype as grid_thw # See https://github.com/huggingface/transformers/pull/34852 for more information dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32, ) 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), ) 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) hidden_states = hidden_states.view( -1, self.spatial_merge_size, self.spatial_merge_size, hidden_states.shape[-1] ) 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, ) class Glm4vTextModel(Qwen2_5_VLTextModel): def __init__(self, config: Glm4vTextConfig): super().__init__(config) self.layers = nn.ModuleList( [Glm4vTextDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] ) self.norm = Glm4vRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.rotary_emb = Glm4vTextRotaryEmbedding(config=config) del self._attn_implementation del self.has_sliding_layers @auto_docstring @merge_with_config_defaults @capture_outputs 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, **kwargs: Unpack[FlashAttentionKwargs], ) -> 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") # torch.jit.trace() doesn't support cache objects in the output if use_cache and past_key_values is None and not torch.jit.is_tracing(): past_key_values = DynamicCache(config=self.config) if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 cache_position = torch.arange( past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) # the hard coded `3` is for temporal, height and width. if position_ids is None: position_ids = cache_position.view(1, 1, -1).expand(3, inputs_embeds.shape[0], -1) elif position_ids.ndim == 2: position_ids = position_ids[None, ...].expand(3, position_ids.shape[0], -1) # NOTE: we need to pass text position ids for packing. Qwen2-VL uses 3D positions # where each dim indicates visual spatial positions for temporal/height/width grids. # There are two scenarios when FA2-like packed masking might be activated. # 1. User specifically passed packed `position_ids` and no attention mask. # In this case we expect the useer to create correct position ids for all 3 grids # and prepend text-only position ids to it. The final tensor will be [4, bs, seq-len] # 2. User runs forward with no attention mask and no position ids. In this case, position ids # are prepared by the model (`get_rope_index`) as `[4, bs, seq-len]` tensor. Text-only positions are # prepended by us when creating positions so that the mask is constructed correctly. NOTE: failing to pass # text-only positions will cause incorrect mask construction, do not change `prepare_input_for_generation` if position_ids.ndim == 3 and position_ids.shape[0] == 4: text_position_ids = position_ids[0] position_ids = position_ids[1:] else: # If inputs are not packed (usual 3D positions), do not prepare mask from position_ids text_position_ids = None mask_kwargs = { "config": self.config, "inputs_embeds": inputs_embeds, "attention_mask": attention_mask, "cache_position": cache_position, "past_key_values": past_key_values, "position_ids": text_position_ids, } # Create the masks causal_mask = create_causal_mask(**mask_kwargs) hidden_states = inputs_embeds position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids) for decoder_layer in self.layers: layer_outputs = decoder_layer( hidden_states, attention_mask=causal_mask, position_ids=text_position_ids, past_key_values=past_key_values, cache_position=cache_position, position_embeddings=position_embeddings, **kwargs, ) hidden_states = layer_outputs hidden_states = self.norm(hidden_states) return BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=past_key_values, ) class Glm4vModel(Qwen2VLModel): _checkpoint_conversion_mapping = {} _no_split_modules = ["Glm4vTextDecoderLayer", "Glm4vVisionBlock"] def __init__(self, config): super().__init__(config) self.visual = Glm4vVisionModel._from_config(config.vision_config) @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. """ pixel_values_videos = pixel_values_videos.type(self.visual.dtype) # reshape video_grid_thw -> [b, 3] -> [1, h, w] * frames temp_frames_hw = [] video_grid_thw_list = video_grid_thw.tolist() for t, h, w in video_grid_thw_list: repeated_row = torch.tensor([1, h, w]).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 ) 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 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, ): """ Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is equal to the length of multimodal features. If the lengths are different, an error is raised. """ if input_ids is None: special_image_mask = inputs_embeds == self.get_input_embeddings()( torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device) ) special_image_mask = special_image_mask.all(-1) special_video_mask = inputs_embeds == self.get_input_embeddings()( torch.tensor(self.config.video_token_id, dtype=torch.long, device=inputs_embeds.device) ) special_video_mask = special_video_mask.all(-1) else: # GLM-4.1V and GLM-4.5V special_video_mask is special_image_mask special_image_mask = input_ids == self.config.image_token_id special_video_mask = input_ids == self.config.image_token_id 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]}", ) 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]}", ) return special_image_mask, special_video_mask def get_rope_index( self, input_ids: torch.LongTensor, mm_token_type_ids: torch.IntTensor, image_grid_thw: torch.LongTensor | None = None, video_grid_thw: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, **kwargs, ) -> tuple[torch.Tensor, torch.Tensor]: """ Calculate the 3D rope index based on image and video's sizes. The utility expects a `vision + text` sequence and will error out otherwise. For pure text sequence, please rely on model's auto-inferred position ids. In a mixed vision + text sequence, vision tokens use 3D RoPE (temporal, height, width) while text tokens use standard 1D RoPE. Example: Temporal patches: 3; Height patches: 2; Width patches: 2 Each vision input results in (temporal x height × width) positions. Here: 3 x 2 × 2 = 12 positions total. Temporal position IDs are spaced by: `interval = tokens_per_second * temporal_patch_size / fps` If fps = 1; tokens_per_second = 25; temporal_patch_size = 2, temporal IDs increase by 50 for each temporal patch: `[0, 0, 0, 0, 50, 50, 50, 50, 100, 100, 100, 100]` Height IDs repeat per row: `[0, 0, 1, 1, ...]` Width IDs alternate per column: `[0, 1, 0, 1, ...]` Text tokens follow standard 1D RoPE and the position IDs grow consequently with a step of `1` Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. mm_token_type_ids (`torch.IntTensor` of shape `(batch_size, sequence_length)`): Token type ids matching each modality to a different value in the input sequence, i.e. text (0), image (1), video (2). image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*): The temporal, height and width of feature shape of each image in LLM. video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*): The temporal, height and width of feature shape of each video in LLM. attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. Returns: position_ids (`torch.LongTensor` of shape `(3, batch_size, sequence_length)`) mrope_position_deltas (`torch.Tensor` of shape `(batch_size)`) """ spatial_merge_size = self.config.vision_config.spatial_merge_size mrope_position_deltas = [] position_ids = torch.zeros( 3, input_ids.shape[0], input_ids.shape[1], dtype=input_ids.dtype, device=input_ids.device, ) grid_iters = { 1: iter(image_grid_thw) if image_grid_thw is not None else None, 2: iter(video_grid_thw) if video_grid_thw is not None else None, } for batch_idx, current_input_ids in enumerate(input_ids): input_token_type = mm_token_type_ids[batch_idx] if attention_mask is not None: current_input_ids = current_input_ids[attention_mask[batch_idx].bool()] input_token_type = input_token_type[attention_mask[batch_idx].bool()] input_type_group = [] for key, group in itertools.groupby(enumerate(input_token_type.tolist()), lambda x: x[1]): group = list(group) start_index = group[0][0] end_index = group[-1][0] + 1 input_type_group.append((key, start_index, end_index)) current_pos = 0 video_group_index = 0 llm_pos_ids_list = [] for modality_type, start_idx, end_idx in input_type_group: # text == 0 if modality_type == 0: text_len = end_idx - start_idx llm_pos_ids_list.append( torch.arange(text_len, device=input_ids.device).view(1, -1).expand(3, -1) + current_pos ) current_pos += text_len # image == 1, video == 2 else: # GLM4V splits video into segments per frame but there's only one `grid_thw` # per whole video. We can't exhaus the iterator and have to re-use the grid # while processing the same video! if modality_type == 2: if video_group_index == 0: grid_thw = next(grid_iters[modality_type]) video_group_index += 1 video_group_index = 0 if video_group_index >= grid_thw[0] else video_group_index else: grid_thw = next(grid_iters[modality_type]) # Videos are processed per frame separately, each temporal grid is always `1` temp_merge_size = grid_thw[0] vision_position_ids = self.get_vision_position_ids( current_pos, grid_thw, temp_merge_size, spatial_merge_size, device=input_ids.device ) llm_pos_ids_list.append(vision_position_ids) current_pos += max(grid_thw[1], grid_thw[2]) // spatial_merge_size llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1) if attention_mask is not None: position_ids[:, batch_idx, attention_mask[batch_idx].bool()] = llm_positions.to(position_ids.device) else: position_ids[:, batch_idx] = llm_positions.to(position_ids.device) mrope_position_deltas.append(llm_positions.max() + 1 - len(current_input_ids)) mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1) return position_ids, mrope_position_deltas @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, mm_token_type_ids: torch.IntTensor | None = None, cache_position: torch.LongTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> 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. video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*): The temporal, height and width of feature shape of each video in LLM. rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*): The rope index difference between sequence length and multimodal rope. """ if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") if inputs_embeds is None: 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 = 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 = 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) if position_ids is None: position_ids = self.compute_3d_position_ids( input_ids=input_ids, image_grid_thw=image_grid_thw, video_grid_thw=video_grid_thw, inputs_embeds=inputs_embeds, attention_mask=attention_mask, past_key_values=past_key_values, mm_token_type_ids=mm_token_type_ids, ) outputs = self.language_model( input_ids=None, position_ids=position_ids, attention_mask=attention_mask, past_key_values=past_key_values, inputs_embeds=inputs_embeds, cache_position=cache_position, **kwargs, ) return Glm4vModelOutputWithPast( **outputs, rope_deltas=self.rope_deltas, ) class Glm4vCausalLMOutputWithPast(Qwen2_5_VLCausalLMOutputWithPast): pass class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration): _checkpoint_conversion_mapping = {} 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, mm_token_type_ids: torch.IntTensor | None = None, cache_position: torch.LongTensor | None = None, logits_to_keep: int | torch.Tensor = 0, **kwargs: Unpack[TransformersKwargs], ) -> 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, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*): The temporal, height and width of feature shape of each image in LLM. video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*): The temporal, height and width of feature shape of each video in LLM. Example: ```python >>> from PIL import Image >>> import httpx >>> from io import BytesIO >>> from transformers import AutoProcessor, Glm4vForConditionalGeneration >>> 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", "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" >>> 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]) >>> # Generate >>> generate_ids = model.generate(inputs.input_ids, max_length=30) >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..." ```""" outputs = self.model( input_ids=input_ids, pixel_values=pixel_values, pixel_values_videos=pixel_values_videos, image_grid_thw=image_grid_thw, video_grid_thw=video_grid_thw, mm_token_type_ids=mm_token_type_ids, position_ids=position_ids, attention_mask=attention_mask, past_key_values=past_key_values, inputs_embeds=inputs_embeds, cache_position=cache_position, **kwargs, ) hidden_states = outputs[0] # Only compute necessary logits, and do not upcast them to float if we are not computing the loss slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep logits = self.lm_head(hidden_states[:, slice_indices, :]) loss = None if labels is not None: loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size) return Glm4vCausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, rope_deltas=outputs.rope_deltas, ) def prepare_inputs_for_generation( self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, cache_position=None, position_ids=None, use_cache=True, pixel_values=None, 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 model_inputs = super().prepare_inputs_for_generation( input_ids, past_key_values=past_key_values, attention_mask=attention_mask, inputs_embeds=inputs_embeds, cache_position=cache_position, position_ids=position_ids, pixel_values=pixel_values, pixel_values_videos=pixel_values_videos, image_grid_thw=image_grid_thw, video_grid_thw=video_grid_thw, use_cache=use_cache, is_first_iteration=is_first_iteration, **kwargs, ) if not is_first_iteration and use_cache: model_inputs["pixel_values"] = None model_inputs["pixel_values_videos"] = None return model_inputs def _get_image_nums_and_video_nums( self, input_ids: torch.LongTensor | None, inputs_embeds: torch.Tensor | None = 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. These parameters are not passed through the processor to avoid unpredictable impacts from interface modifications. Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Returns: image_nums (`torch.LongTensor` of shape `(batch_size, num_images_sample)`) video_nums (`torch.LongTensor` of shape `(batch_size, num_videos_sample)`) """ if inputs_embeds is not None: is_image = ( inputs_embeds == self.get_input_embeddings()( torch.tensor(self.config.image_start_token_id, dtype=torch.long, device=inputs_embeds.device) ) )[..., 0] is_video_start = ( inputs_embeds == self.get_input_embeddings()( torch.tensor(self.config.video_start_token_id, dtype=torch.long, device=inputs_embeds.device) ) )[..., 0] is_video_end = ( inputs_embeds == self.get_input_embeddings()( torch.tensor(self.config.video_end_token_id, dtype=torch.long, device=inputs_embeds.device) ) )[..., 0] else: is_image = input_ids == self.config.image_start_token_id is_video_start = input_ids == self.config.video_start_token_id is_video_end = input_ids == self.config.video_end_token_id # Cumulative sum to track if we're inside a video span # We'll assume well-formed video tags (i.e. matching starts and ends) video_level = torch.cumsum(is_video_start.int() - is_video_end.int(), dim=1) inside_video = video_level > 0 # shape (batch_size, seq_length) # Mask out image tokens that are inside video spans standalone_images = is_image & (~inside_video) # Count per batch image_counts = standalone_images.sum(dim=1) video_counts = is_video_start.sum(dim=1) return image_counts, video_counts class Glm4vProcessorKwargs(Qwen2VLProcessorKwargs): _defaults = { "text_kwargs": { "padding": False, "return_token_type_ids": False, "return_mm_token_type_ids": True, }, "videos_kwargs": {"return_metadata": True}, } class Glm4vProcessor(Qwen2VLProcessor): 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 self.video_token = "<|video|>" if not hasattr(tokenizer, "video_token") else tokenizer.video_token self.video_start_id = tokenizer.convert_tokens_to_ids("<|begin_of_video|>") self.video_end_id = tokenizer.convert_tokens_to_ids("<|end_of_video|>") def __call__( self, images: ImageInput | None = None, text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] = None, videos: VideoInput | None = None, **kwargs: Unpack[Glm4vProcessorKwargs], ) -> BatchFeature: r""" Returns: [`BatchFeature`]: A [`BatchFeature`] with the following fields: - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`. - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not `None`). - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`. - **pixel_values_videos** -- Pixel values of videos to be fed to a model. Returned when `videos` is not `None`. - **image_grid_thw** -- List of image 3D grid in LLM. Returned when `images` is not `None`. - **video_grid_thw** -- List of video 3D grid in LLM. Returned when `videos` is not `None`. """ output_kwargs = self._merge_kwargs( Glm4vProcessorKwargs, tokenizer_init_kwargs=self.tokenizer.init_kwargs, **kwargs, ) if images is not None: image_inputs = self.image_processor(images=images, **output_kwargs["images_kwargs"]) image_grid_thw = image_inputs["image_grid_thw"] else: image_inputs = {} image_grid_thw = None if videos is not None: videos_inputs = self.video_processor(videos=videos, **output_kwargs["videos_kwargs"]) # If user has not requested video metadata, pop it if not kwargs.get("return_metadata"): video_metadata = videos_inputs.pop("video_metadata") else: video_metadata = videos_inputs["video_metadata"] video_grid_thw = videos_inputs["video_grid_thw"] else: videos_inputs = {} video_grid_thw = None if not isinstance(text, list): text = [text] text = text.copy() # below lines change text in-place if image_grid_thw is not None: merge_length = self.image_processor.merge_size**2 index = 0 for i in range(len(text)): while self.image_token in text[i]: num_image_tokens = image_grid_thw[index].prod() // merge_length text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1) index += 1 text[i] = text[i].replace("<|placeholder|>", self.image_token) if video_grid_thw is not None: merge_length = self.video_processor.merge_size**2 video_index = 0 for i in range(len(text)): while self.video_token in text[i]: num_frames = video_grid_thw[video_index][0] video_structure = "" metadata = video_metadata[video_index] if metadata.fps is None: logger.warning_once( "SmolVLM requires frame timestamps to construct prompts, but the `fps` of the input video could not be inferred. " "Probably `video_metadata` was missing from inputs and you passed pre-sampled frames. " "Defaulting to `fps=24`. Please provide `video_metadata` for more accurate results." ) metadata.fps = 24 if metadata.fps is None else metadata.fps timestamps = metadata.timestamps[::2] # mrope unique_timestamps = [] for idx in range(0, len(timestamps)): unique_timestamps.append(timestamps[idx]) selected_timestamps = unique_timestamps[:num_frames] while len(selected_timestamps) < num_frames: selected_timestamps.append(selected_timestamps[-1] if selected_timestamps else 0) for frame_idx in range(num_frames): timestamp_sec = selected_timestamps[frame_idx] frame_structure = self.replace_frame_token_id(timestamp_sec) video_structure += frame_structure text[i] = text[i].replace(self.video_token, video_structure, 1) num_image_tokens = ( video_grid_thw[video_index].prod() // merge_length // video_grid_thw[video_index][0] ) for frame_idx in range(num_frames): if self.image_token in text[i]: text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1) video_index += 1 text[i] = text[i].replace("<|placeholder|>", self.image_token) return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None) return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", False) text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"]) self._check_special_mm_tokens(text, text_inputs, modalities=["image", "video"]) if return_mm_token_type_ids: array_ids = np.array(text_inputs["input_ids"]) mm_token_type_ids = np.zeros_like(text_inputs["input_ids"]) # Replace 0 -> 2 only inside video segments because GLM4v # uses the same special token to denote images and video # Otherwise replace 0 -> 1 for image modality starts = np.cumsum(array_ids == self.video_start_id, axis=1) ends = np.cumsum(array_ids == self.video_end_id, axis=1) is_video_modality = starts > ends mm_token_type_ids[(array_ids == self.image_token_id) & is_video_modality] = 2 mm_token_type_ids[(array_ids == self.image_token_id) & (~is_video_modality)] = 1 text_inputs["mm_token_type_ids"] = mm_token_type_ids.tolist() return BatchFeature(data={**text_inputs, **image_inputs, **videos_inputs}, tensor_type=return_tensors) def replace_frame_token_id(self, timestamp_sec): return f"<|begin_of_image|>{self.image_token}<|end_of_image|>{int(timestamp_sec)}" __all__ = [ "Glm4vConfig", "Glm4vTextConfig", "Glm4vVisionConfig", "Glm4vForConditionalGeneration", "Glm4vModel", "Glm4vPreTrainedModel", "Glm4vProcessor", "Glm4vTextModel", "Glm4vVisionModel", ]