transformers 5.0.0rc1__py3-none-any.whl → 5.0.0rc3__py3-none-any.whl

transformers/models/glm4v_moe/modeling_glm4v_moe.py

@@ -4,7 +4,6 @@
 #             the file from the modular. If any change should be done, please apply the change to the
 #                          modular_glm4v_moe.py file directly. One of our CI enforces this.
 #                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
-# 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");
@@ -21,7 +20,7 @@
 import itertools
 from collections.abc import Callable
 from dataclasses import dataclass
-from typing import Any, Optional, Union
+from typing import Any, Optional
 
 import torch
 import torch.nn as nn
@@ -32,7 +31,7 @@ from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_kernel_forward_from_hub, use_kernelized_func
+from ...integrations import use_experts_implementation, use_kernel_forward_from_hub, use_kernelized_func
 from ...masking_utils import create_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
@@ -40,125 +39,17 @@ from ...modeling_outputs import ModelOutput, MoeModelOutputWithPast
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_torchdynamo_compiling
+from ...utils import (
+    TransformersKwargs,
+    auto_docstring,
+    can_return_tuple,
+    is_grouped_mm_available,
+    is_torchdynamo_compiling,
+)
 from ...utils.generic import OutputRecorder, check_model_inputs, maybe_autocast
 from .configuration_glm4v_moe import Glm4vMoeConfig, Glm4vMoeTextConfig, Glm4vMoeVisionConfig
 
 
-@use_kernel_forward_from_hub("RMSNorm")
-class Glm4vMoeRMSNorm(nn.Module):
-    def __init__(self, hidden_size, eps=1e-6):
-        """
-        Glm4vMoeRMSNorm is equivalent to T5LayerNorm
-        """
-        super().__init__()
-        self.weight = nn.Parameter(torch.ones(hidden_size))
-        self.variance_epsilon = eps
-
-    def forward(self, hidden_states):
-        input_dtype = hidden_states.dtype
-        hidden_states = hidden_states.to(torch.float32)
-        variance = hidden_states.pow(2).mean(-1, keepdim=True)
-        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
-        return self.weight * hidden_states.to(input_dtype)
-
-    def extra_repr(self):
-        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
-
-
-@dataclass
-@auto_docstring(
-    custom_intro="""
-    Base class for Llava outputs, with hidden states and attentions.
-    """
-)
-class Glm4vMoeModelOutputWithPast(ModelOutput):
-    r"""
-    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
-        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
-
-        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
-        `past_key_values` input) to speed up sequential decoding.
-    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
-        The rope index difference between sequence length and multimodal rope.
-    """
-
-    last_hidden_state: Optional[torch.FloatTensor] = None
-    past_key_values: Optional[Cache] = None
-    hidden_states: Optional[tuple[torch.FloatTensor]] = None
-    attentions: Optional[tuple[torch.FloatTensor]] = None
-    rope_deltas: Optional[torch.LongTensor] = None
-
-
-class Glm4vMoeTextRotaryEmbedding(nn.Module):
-    inv_freq: torch.Tensor  # fix linting for `register_buffer`
-
-    def __init__(self, config: Glm4vMoeTextConfig, device=None, layer_type=None):
-        super().__init__()
-        self.max_seq_len_cached = config.max_position_embeddings
-        self.original_max_seq_len = config.max_position_embeddings
-
-        self.config = config
-
-        self.rope_type = self.config.rope_parameters["rope_type"]
-        rope_init_fn: Callable = self.compute_default_rope_parameters
-        if self.rope_type != "default":
-            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
-        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
-
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = inv_freq
-
-    @staticmethod
-    def compute_default_rope_parameters(
-        config: Optional[Glm4vMoeTextConfig] = None,
-        device: Optional["torch.device"] = None,
-        seq_len: Optional[int] = None,
-    ) -> tuple["torch.Tensor", float]:
-        """
-        Computes the inverse frequencies according to the original RoPE implementation
-        Args:
-            config ([`~transformers.PreTrainedConfig`]):
-                The model configuration.
-            device (`torch.device`):
-                The device to use for initialization of the inverse frequencies.
-            seq_len (`int`, *optional*):
-                The current sequence length. Unused for this type of RoPE.
-        Returns:
-            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
-            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
-        """
-        base = config.rope_parameters["rope_theta"]
-        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 1.0)
-        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
-        dim = int(head_dim * partial_rotary_factor)
-
-        attention_factor = 1.0  # Unused in this type of RoPE
-
-        # Compute the inverse frequencies
-        inv_freq = 1.0 / (
-            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
-        )
-        return inv_freq, attention_factor
-
-    @torch.no_grad()
-    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
-    def forward(self, x, position_ids):
-        # In contrast to other models, GLM4V_MOE 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)
-            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 repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     """
     This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
@@ -176,7 +67,7 @@ def eager_attention_forward(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attention_mask: Optional[torch.Tensor],
+    attention_mask: torch.Tensor | None,
     scaling: float,
     dropout: float = 0.0,
     **kwargs: Unpack[TransformersKwargs],
@@ -204,7 +95,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 
     Args:
@@ -212,8 +103,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
         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.
-        position_ids (`torch.Tensor`, *optional*):
-            Deprecated and unused.
         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
@@ -227,10 +116,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     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:]
@@ -246,64 +131,11 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     return q_embed, k_embed
 
 
-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
-            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.
-    """
-    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
-    )
-
-    # 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(q_rot) * sin)
-    k_embed = (k_rot * cos) + (rotate_half(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
-
-
 @use_kernelized_func(apply_rotary_pos_emb)
 class Glm4vMoeTextAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
-    def __init__(self, config: Glm4vMoeTextConfig, layer_idx: Optional[int] = None):
+    def __init__(self, config: Glm4vMoeTextConfig, layer_idx: int | None = None):
         super().__init__()
         self.config = config
         self.layer_idx = layer_idx
@@ -329,11 +161,11 @@ class Glm4vMoeTextAttention(nn.Module):
         self,
         hidden_states: torch.Tensor,
         position_embeddings: tuple[torch.Tensor, torch.Tensor],
-        attention_mask: Optional[torch.Tensor],
-        past_key_values: Optional[Cache] = None,
-        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
         **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, self.head_dim)
 
@@ -346,9 +178,7 @@ class Glm4vMoeTextAttention(nn.Module):
         value_states = value_states.transpose(1, 2)
 
         cos, sin = position_embeddings
-        query_states, key_states = apply_multimodal_rotary_pos_emb(  # diff with Llama
-            query_states, key_states, cos, sin, self.rope_parameters["mrope_section"]
-        )
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
         if past_key_values is not None:
             # sin and cos are specific to RoPE models; position_ids needed for the static cache
@@ -395,6 +225,7 @@ class Glm4vMoeTextTopkRouter(nn.Module):
         return router_logits
 
 
+@use_experts_implementation
 class Glm4vMoeTextNaiveMoe(nn.Module):
     """Collection of expert weights stored as 3D tensors."""
 
@@ -402,7 +233,7 @@ class Glm4vMoeTextNaiveMoe(nn.Module):
         super().__init__()
         self.num_experts = config.num_local_experts
         self.hidden_dim = config.hidden_size
-        self.intermediate_dim = config.intermediate_size
+        self.intermediate_dim = config.moe_intermediate_size
         self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
         self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
         self.act_fn = ACT2FN[config.hidden_act]
@@ -545,12 +376,12 @@ class Glm4vMoeTextDecoderLayer(GradientCheckpointingLayer):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        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,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = 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: tuple[torch.Tensor, torch.Tensor] | None = None,
         **kwargs: Unpack[TransformersKwargs],
     ) -> torch.Tensor:
         residual = hidden_states
@@ -586,7 +417,9 @@ class Glm4vMoePreTrainedModel(PreTrainedModel):
     _supports_flash_attn = True
     _supports_sdpa = True
     _supports_flex_attn = True
-    _can_compile_fullgraph = False
+    _can_compile_fullgraph = (
+        is_grouped_mm_available()
+    )  # https://huggingface.co/docs/transformers/experts_interface#torchcompile
     _supports_attention_backend = True
 
     _can_record_outputs = {
@@ -602,9 +435,13 @@ class Glm4vMoePreTrainedModel(PreTrainedModel):
         super()._init_weights(module)
         if isinstance(module, Glm4vMoeTextTopkRouter):
             init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            init.zeros_(module.e_score_correction_bias)
         elif isinstance(module, Glm4vMoeTextNaiveMoe):
             init.normal_(module.gate_up_proj, mean=0.0, std=self.config.initializer_range)
             init.normal_(module.down_proj, mean=0.0, std=self.config.initializer_range)
+        if isinstance(module, Glm4vMoeVisionRotaryEmbedding):
+            inv_freq = 1.0 / (module.theta ** (torch.arange(0, module.dim, 2, dtype=torch.float) / module.dim))
+            init.copy_(module.inv_freq, inv_freq)
 
 
 @dataclass
@@ -628,13 +465,50 @@ class Glm4vMoeCausalLMOutputWithPast(ModelOutput):
         The rope index difference between sequence length and multimodal rope.
     """
 
-    loss: Optional[torch.FloatTensor] = None
-    logits: Optional[torch.FloatTensor] = None
-    past_key_values: Optional[Cache] = None
-    hidden_states: Optional[tuple[torch.FloatTensor]] = None
-    attentions: Optional[tuple[torch.FloatTensor]] = None
-    rope_deltas: Optional[torch.LongTensor] = None
-    aux_loss: Optional[torch.FloatTensor] = None
+    loss: torch.FloatTensor | None = None
+    logits: torch.FloatTensor | None = None
+    past_key_values: Cache | None = None
+    hidden_states: tuple[torch.FloatTensor] | None = None
+    attentions: tuple[torch.FloatTensor] | None = None
+    rope_deltas: torch.LongTensor | None = None
+    aux_loss: torch.FloatTensor | None = None
+
+
+class Glm4vMoeVisionRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    def forward(self, seqlen: int) -> torch.Tensor:
+        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.outer(seq, self.inv_freq)
+        return freqs
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class Glm4vMoeRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Glm4vMoeRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
 
 
 class Glm4vMoeisionMlp(nn.Module):
@@ -671,20 +545,6 @@ class Glm4vMoeVisionPatchEmbed(nn.Module):
         return hidden_states
 
 
-class Glm4vMoeVisionRotaryEmbedding(nn.Module):
-    inv_freq: torch.Tensor  # fix linting for `register_buffer`
-
-    def __init__(self, dim: int, theta: float = 10000.0) -> None:
-        super().__init__()
-        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-
-    def forward(self, seqlen: int) -> torch.Tensor:
-        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
-        freqs = torch.outer(seq, self.inv_freq)
-        return freqs
-
-
 class Glm4vMoeVisionPatchMerger(nn.Module):
     def __init__(self, dim: int, context_dim: int, hidden_act: str, bias: bool = False) -> None:
         super().__init__()
@@ -713,7 +573,7 @@ class Glm4vMoeVisionEmbeddings(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.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
+        self.interpolated_method = "bicubic"
 
     def forward(self, embeddings, lengths, image_shapes, h_coords, w_coords) -> torch.Tensor:
         """
@@ -732,57 +592,45 @@ class Glm4vMoeVisionEmbeddings(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
 
-        # Move coordinates to correct device
-        h_coords, w_coords = h_coords.to(device), w_coords.to(device)
-
-        # 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)
-            )
+        # 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
-            )
+        # 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
-            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
+        # 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)
+        # 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="bicubic", align_corners=False, padding_mode="border"
-            )
+        # 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)
+        # 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
@@ -821,8 +669,8 @@ class Glm4vMoeVisionAttention(nn.Module):
         self,
         hidden_states: torch.Tensor,
         cu_seqlens: torch.Tensor,
-        rotary_pos_emb: Optional[torch.Tensor] = None,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        rotary_pos_emb: torch.Tensor | None = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
         **kwargs,
     ) -> torch.Tensor:
         seq_length = hidden_states.shape[0]
@@ -840,8 +688,8 @@ class Glm4vMoeVisionAttention(nn.Module):
         if self.config._attn_implementation != "eager":
             attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
-        if self.config._attn_implementation == "flash_attention_2":
-            # Flash Attention 2: Use cu_seqlens for variable length attention
+        if "flash" in self.config._attn_implementation:
+            # Flash Attention: Use cu_seqlens for variable length attention
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
             attn_output, _ = attention_interface(
                 self,
@@ -898,8 +746,8 @@ class Glm4vMoeVisionBlock(GradientCheckpointingLayer):
         self,
         hidden_states: torch.Tensor,
         cu_seqlens: torch.Tensor,
-        rotary_pos_emb: Optional[torch.Tensor] = None,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        rotary_pos_emb: torch.Tensor | None = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
         **kwargs,
     ) -> torch.Tensor:
         hidden_states = hidden_states + self.attn(
@@ -989,7 +837,6 @@ class Glm4vMoeVisionModel(Glm4vMoePreTrainedModel):
         """
         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())
@@ -1004,7 +851,13 @@ class Glm4vMoeVisionModel(Glm4vMoePreTrainedModel):
         )
         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], image_type_ids[:, 1])
+        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(
@@ -1025,6 +878,83 @@ class Glm4vMoeVisionModel(Glm4vMoePreTrainedModel):
         return hidden_states
 
 
+class Glm4vMoeTextRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: Glm4vMoeTextConfig, device=None):
+        super().__init__()
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+
+        self.rope_type = self.config.rope_parameters["rope_type"]
+        rope_init_fn: Callable = self.compute_default_rope_parameters
+        if self.rope_type != "default":
+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+        self.mrope_section = config.rope_parameters.get("mrope_section", [8, 12, 12])
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: Glm4vMoeTextConfig | None = None,
+        device: Optional["torch.device"] = None,
+        seq_len: int | None = None,
+    ) -> tuple["torch.Tensor", float]:
+        """
+        Computes the inverse frequencies according to the original RoPE implementation
+        Args:
+            config ([`~transformers.PreTrainedConfig`]):
+                The model configuration.
+            device (`torch.device`):
+                The device to use for initialization of the inverse frequencies.
+            seq_len (`int`, *optional*):
+                The current sequence length. Unused for this type of RoPE.
+        Returns:
+            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
+            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
+        """
+        base = config.rope_parameters["rope_theta"]
+        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 1.0)
+        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+        dim = int(head_dim * partial_rotary_factor)
+
+        attention_factor = 1.0  # Unused in this type of RoPE
+
+        # Compute the inverse frequencies
+        inv_freq = 1.0 / (
+            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
+        )
+        return inv_freq, attention_factor
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    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
+
+
 @auto_docstring
 class Glm4vMoeTextModel(Glm4vMoePreTrainedModel):
     config: Glm4vMoeTextConfig
@@ -1050,13 +980,13 @@ class Glm4vMoeTextModel(Glm4vMoePreTrainedModel):
     @check_model_inputs
     def forward(
         self,
-        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,
+        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],
     ) -> MoeModelOutputWithPast:
         if (input_ids is None) ^ (inputs_embeds is not None):
@@ -1134,6 +1064,30 @@ class Glm4vMoeTextModel(Glm4vMoePreTrainedModel):
         )
 
 
+@dataclass
+@auto_docstring(
+    custom_intro="""
+    Base class for Llava outputs, with hidden states and attentions.
+    """
+)
+class Glm4vMoeModelOutputWithPast(ModelOutput):
+    r"""
+    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
+
+        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+        `past_key_values` input) to speed up sequential decoding.
+    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+        The rope index difference between sequence length and multimodal rope.
+    """
+
+    last_hidden_state: torch.FloatTensor | None = None
+    past_key_values: Cache | None = None
+    hidden_states: tuple[torch.FloatTensor] | None = None
+    attentions: tuple[torch.FloatTensor] | None = None
+    rope_deltas: torch.LongTensor | None = None
+
+
 @auto_docstring
 class Glm4vMoeModel(Glm4vMoePreTrainedModel):
     base_model_prefix = "model"
@@ -1160,10 +1114,10 @@ class Glm4vMoeModel(Glm4vMoePreTrainedModel):
 
     def get_rope_index(
         self,
-        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,
+        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,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """
         Calculate the 3D rope index based on image and video's temporal, height and width in LLM.
@@ -1349,7 +1303,7 @@ class Glm4vMoeModel(Glm4vMoePreTrainedModel):
             return position_ids, mrope_position_deltas
 
     def get_video_features(
-        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
+        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: torch.LongTensor | None = None
     ):
         """
         Encodes videos into continuous embeddings that can be forwarded to the language model.
@@ -1372,7 +1326,7 @@ class Glm4vMoeModel(Glm4vMoePreTrainedModel):
         video_embeds = torch.split(video_embeds, split_sizes)
         return video_embeds
 
-    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
+    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: torch.LongTensor | None = None):
         """
         Encodes images into continuous embeddings that can be forwarded to the language model.
 
@@ -1392,8 +1346,8 @@ class Glm4vMoeModel(Glm4vMoePreTrainedModel):
         self,
         input_ids: torch.LongTensor,
         inputs_embeds: torch.FloatTensor,
-        image_features: Optional[torch.FloatTensor] = None,
-        video_features: Optional[torch.FloatTensor] = None,
+        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
@@ -1433,19 +1387,19 @@ class Glm4vMoeModel(Glm4vMoePreTrainedModel):
     @can_return_tuple
     def forward(
         self,
-        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,
+        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,
         **kwargs: Unpack[TransformersKwargs],
-    ) -> Union[tuple, Glm4vMoeModelOutputWithPast]:
+    ) -> tuple | Glm4vMoeModelOutputWithPast:
         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.
@@ -1538,11 +1492,11 @@ class Glm4vMoeModel(Glm4vMoePreTrainedModel):
 
 
 def load_balancing_loss_func(
-    gate_logits: Union[torch.Tensor, tuple[torch.Tensor], None],
-    num_experts: Optional[int] = None,
+    gate_logits: torch.Tensor | tuple[torch.Tensor] | None,
+    num_experts: int | None = None,
     top_k=2,
-    attention_mask: Optional[torch.Tensor] = None,
-) -> Union[torch.Tensor, int]:
+    attention_mask: torch.Tensor | None = None,
+) -> torch.Tensor | int:
     r"""
     Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
 
@@ -1639,31 +1593,31 @@ class Glm4vMoeForConditionalGeneration(Glm4vMoePreTrainedModel, GenerationMixin)
         self.model.set_input_embeddings(value)
 
     def get_video_features(
-        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
+        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: torch.LongTensor | None = None
     ):
         return self.model.get_video_features(pixel_values_videos, video_grid_thw)
 
-    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
+    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: torch.LongTensor | None = None):
         return self.model.get_image_features(pixel_values, image_grid_thw)
 
     @auto_docstring
     @check_model_inputs
     def forward(
         self,
-        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,
+        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,
         **kwargs: Unpack[TransformersKwargs],
-    ) -> Union[tuple, Glm4vMoeCausalLMOutputWithPast]:
+    ) -> tuple | Glm4vMoeCausalLMOutputWithPast:
         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, ...,
@@ -1763,6 +1717,7 @@ class Glm4vMoeForConditionalGeneration(Glm4vMoePreTrainedModel, GenerationMixin)
         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
@@ -1779,13 +1734,14 @@ class Glm4vMoeForConditionalGeneration(Glm4vMoePreTrainedModel, GenerationMixin)
             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 cache_position[0] != 0:
+        if not is_first_iteration and use_cache:
             model_inputs["pixel_values"] = None
             model_inputs["pixel_values_videos"] = None
 
@@ -1793,8 +1749,8 @@ class Glm4vMoeForConditionalGeneration(Glm4vMoePreTrainedModel, GenerationMixin)
 
     def _get_image_nums_and_video_nums(
         self,
-        input_ids: Optional[torch.LongTensor],
-        inputs_embeds: Optional[torch.Tensor] = None,
+        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.
@@ -1851,7 +1807,7 @@ class Glm4vMoeForConditionalGeneration(Glm4vMoePreTrainedModel, GenerationMixin)
         self,
         expand_size: int = 1,
         is_encoder_decoder: bool = False,
-        input_ids: Optional[torch.LongTensor] = None,
+        input_ids: torch.LongTensor | None = None,
         **model_kwargs,
     ) -> tuple[torch.LongTensor, dict[str, Any]]:
         # Overwritten -- Support for expanding tensors without a batch size dimension