transformers 5.0.0rc0__py3-none-any.whl → 5.0.0rc2__py3-none-any.whl

transformers/models/glm4_moe/modeling_glm4_moe.py

@@ -30,7 +30,7 @@ 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
+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
@@ -38,8 +38,8 @@ from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 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
-from ...utils.generic import check_model_inputs
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_grouped_mm_available
+from ...utils.generic import check_model_inputs, maybe_autocast
 from .configuration_glm4_moe import Glm4MoeConfig
 
 
@@ -60,7 +60,7 @@ class Glm4MoeRotaryEmbedding(nn.Module):
         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
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
@@ -101,7 +101,7 @@ class Glm4MoeRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
@@ -193,6 +193,7 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     return q_embed, k_embed
 
 
+@use_kernelized_func(apply_rotary_pos_emb)
 class Glm4MoeAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -331,6 +332,7 @@ class Glm4MoeRMSNorm(nn.Module):
         return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
 
 
+@use_experts_implementation
 class Glm4MoeNaiveMoe(nn.Module):
     """Collection of expert weights stored as 3D tensors."""
 
@@ -338,7 +340,7 @@ class Glm4MoeNaiveMoe(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]
@@ -485,18 +487,22 @@ class Glm4MoePreTrainedModel(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 = {
         "hidden_states": Glm4MoeDecoderLayer,
         "attentions": Glm4MoeAttention,
     }
+    _keep_in_fp32_modules_strict = ["e_score_correction_bias"]
 
     @torch.no_grad()
     def _init_weights(self, module):
         super()._init_weights(module)
         if isinstance(module, Glm4MoeTopkRouter):
             init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
+            init.zeros_(module.e_score_correction_bias)
         elif isinstance(module, Glm4MoeNaiveMoe):
             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)

transformers/models/glm4_moe/modular_glm4_moe.py

@@ -273,7 +273,7 @@ class Glm4MoeDecoderLayer(DeepseekV3DecoderLayer):
 
 
 class Glm4MoePreTrainedModel(DeepseekV3PreTrainedModel):
-    _can_compile_fullgraph = False
+    pass
 
 
 class Glm4MoeModel(DeepseekV3Model):

transformers/models/glm4v/configuration_glm4v.py

@@ -234,7 +234,9 @@ class Glm4vTextConfig(PreTrainedConfig):
         self.attention_dropout = attention_dropout
         self.rope_parameters = rope_parameters
 
-        super().__init__(tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope"}, **kwargs)
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope_section"}, **kwargs
+        )
 
 
 class Glm4vConfig(PreTrainedConfig):

transformers/models/glm4v/image_processing_glm4v.py

@@ -353,7 +353,6 @@ class Glm4vImageProcessor(BaseImageProcessor):
             image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
                 Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
                 `True`.
-                The max pixels of the image to resize the image.
             patch_size (`int`, *optional*, defaults to `self.patch_size`):
                 The spatial patch size of the vision encoder.
             temporal_patch_size (`int`, *optional*, defaults to `self.temporal_patch_size`):
@@ -380,12 +379,9 @@ class Glm4vImageProcessor(BaseImageProcessor):
                 - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
 
         """
-        # Try to use config values if set, otherwise fallback to global defaults
         size = size if size is not None else self.size
         if size is not None and ("shortest_edge" not in size or "longest_edge" not in size):
             raise ValueError("size must contain 'shortest_edge' and 'longest_edge' keys.")
-        elif size is None:
-            size = {"shortest_edge": 112 * 112, "longest_edge": 28 * 28 * 15000}
 
         do_resize = do_resize if do_resize is not None else self.do_resize
         resample = resample if resample is not None else self.resample

transformers/models/glm4v/modeling_glm4v.py

@@ -28,6 +28,7 @@ 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 ...generation import GenerationMixin
@@ -40,7 +41,7 @@ from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_torchdynamo_compiling
-from ...utils.generic import check_model_inputs
+from ...utils.generic import check_model_inputs, maybe_autocast
 from .configuration_glm4v import Glm4vConfig, Glm4vTextConfig, Glm4vVisionConfig
 
 
@@ -104,6 +105,8 @@ class Glm4vVisionRotaryEmbedding(nn.Module):
 
     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)
 
@@ -141,7 +144,6 @@ class Glm4vVisionEmbeddings(nn.Module):
         self.num_patches = (self.image_size // self.patch_size) ** 2
         self.num_positions = self.num_patches
         self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
-        self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
 
     def forward(self, embeddings, lengths, image_shapes, h_coords, w_coords) -> torch.Tensor:
         """
@@ -313,8 +315,8 @@ class Glm4vVisionAttention(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,
@@ -403,7 +405,7 @@ class Glm4vTextRotaryEmbedding(nn.Module):
         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
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
@@ -446,7 +448,7 @@ class Glm4vTextRotaryEmbedding(nn.Module):
         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 torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        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
@@ -705,6 +707,12 @@ class Glm4vPreTrainedModel(PreTrainedModel):
         "attentions": Glm4vTextAttention,
     }
 
+    def _init_weights(self, module):
+        super()._init_weights(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
@@ -768,7 +776,7 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
         rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
         return rotary_pos_emb, pos_ids
 
-    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor) -> torch.Tensor:
+    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
         """
         Args:
             hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
@@ -1487,6 +1495,7 @@ class Glm4vForConditionalGeneration(Glm4vPreTrainedModel, 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
@@ -1503,13 +1512,14 @@ class Glm4vForConditionalGeneration(Glm4vPreTrainedModel, 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
 

transformers/models/glm4v/modular_glm4v.py

@@ -22,6 +22,7 @@ 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
@@ -32,11 +33,11 @@ from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_outputs import BaseModelOutputWithPast
 from ...modeling_rope_utils import RopeParameters
-from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...tokenization_utils_base import PreTokenizedInput, TextInput
 from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_torchdynamo_compiling, logging
-from ...utils.generic import check_model_inputs
+from ...utils.generic import check_model_inputs, maybe_autocast
 from ...video_utils import VideoInput
 from ..glm4.modeling_glm4 import Glm4MLP, Glm4RMSNorm, Glm4RotaryEmbedding, eager_attention_forward
 from ..qwen2_5_vl.modeling_qwen2_5_vl import (
@@ -271,7 +272,9 @@ class Glm4vTextConfig(PreTrainedConfig):
         self.attention_dropout = attention_dropout
         self.rope_parameters = rope_parameters
 
-        super().__init__(tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope"}, **kwargs)
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope_section"}, **kwargs
+        )
 
 
 class Glm4vConfig(PreTrainedConfig):
@@ -407,7 +410,6 @@ class Glm4vVisionEmbeddings(nn.Module):
         self.num_patches = (self.image_size // self.patch_size) ** 2
         self.num_positions = self.num_patches
         self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
-        self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
 
     def forward(self, embeddings, lengths, image_shapes, h_coords, w_coords) -> torch.Tensor:
         """
@@ -509,7 +511,7 @@ class Glm4vTextRotaryEmbedding(Glm4RotaryEmbedding):
         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 torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        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
@@ -723,6 +725,12 @@ class Glm4vPreTrainedModel(Qwen2_5_VLPreTrainedModel):
         "attentions": Glm4vTextAttention,
     }
 
+    def _init_weights(self, module):
+        PreTrainedModel._init_weights(self, module)
+        if isinstance(module, Glm4vVisionRotaryEmbedding):
+            inv_freq = 1.0 / (module.theta ** (torch.arange(0, module.dim, 2, dtype=torch.float) / module.dim))
+            init.copy_(module.inv_freq, inv_freq)
+
 
 class Glm4vVisionModel(Glm4vPreTrainedModel):
     config: Glm4vVisionConfig
@@ -786,7 +794,7 @@ class Glm4vVisionModel(Glm4vPreTrainedModel):
         rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
         return rotary_pos_emb, pos_ids
 
-    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor) -> torch.Tensor:
+    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
         """
         Args:
             hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
@@ -1412,6 +1420,7 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
         pixel_values_videos=None,
         image_grid_thw=None,
         video_grid_thw=None,
+        is_first_iteration=False,
         **kwargs,
     ):
         # Overwritten -- in specific circumstances we don't want to forward image inputs to the model
@@ -1428,13 +1437,14 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
             image_grid_thw=image_grid_thw,
             video_grid_thw=video_grid_thw,
             use_cache=use_cache,
+            is_first_iteration=is_first_iteration,
             **kwargs,
         )
 
         # GLM-4.1V position_ids are prepareed with rope_deltas in forward
         model_inputs["position_ids"] = None
 
-        if cache_position[0] != 0:
+        if not is_first_iteration and use_cache:
             model_inputs["pixel_values"] = None
             model_inputs["pixel_values_videos"] = None
 

transformers/models/glm4v_moe/configuration_glm4v_moe.py

@@ -280,7 +280,9 @@ class Glm4vMoeTextConfig(PreTrainedConfig):
         self.first_k_dense_replace = first_k_dense_replace
         self.norm_topk_prob = norm_topk_prob
         self.router_aux_loss_coef = router_aux_loss_coef
-        super().__init__(tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope"}, **kwargs)
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope_section"}, **kwargs
+        )
 
 
 class Glm4vMoeConfig(PreTrainedConfig):

transformers/models/glm4v_moe/modeling_glm4v_moe.py

@@ -32,7 +32,7 @@ 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
+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
@@ -40,8 +40,14 @@ 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.generic import OutputRecorder, check_model_inputs
+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
 
 
@@ -107,7 +113,7 @@ class Glm4vMoeTextRotaryEmbedding(nn.Module):
         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
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
 
     @staticmethod
     def compute_default_rope_parameters(
@@ -150,7 +156,7 @@ class Glm4vMoeTextRotaryEmbedding(nn.Module):
         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 torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        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
@@ -299,6 +305,7 @@ def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim
     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"""
 
@@ -322,7 +329,6 @@ class Glm4vMoeTextAttention(nn.Module):
             config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
         )
         self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
-        self.rotary_fn = apply_rotary_pos_emb
         self.rope_parameters = config.rope_parameters
 
     def forward(
@@ -395,6 +401,7 @@ class Glm4vMoeTextTopkRouter(nn.Module):
         return router_logits
 
 
+@use_experts_implementation
 class Glm4vMoeTextNaiveMoe(nn.Module):
     """Collection of expert weights stored as 3D tensors."""
 
@@ -402,7 +409,7 @@ 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]
@@ -586,7 +593,9 @@ 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 = {
@@ -594,6 +603,7 @@ class Glm4vMoePreTrainedModel(PreTrainedModel):
         "attentions": Glm4vMoeTextAttention,
         "router_logits": OutputRecorder(nn.Linear, layer_name="mlp.gate", index=0),
     }
+    _keep_in_fp32_modules_strict = ["e_score_correction_bias"]
     input_modalities = ("text", "image", "video")
 
     @torch.no_grad()
@@ -601,9 +611,13 @@ 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
@@ -636,6 +650,22 @@ class Glm4vMoeCausalLMOutputWithPast(ModelOutput):
     aux_loss: Optional[torch.FloatTensor] = 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
+
+
 class Glm4vMoeisionMlp(nn.Module):
     def __init__(self, config, bias: bool = False):
         super().__init__()
@@ -670,20 +700,6 @@ 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__()
@@ -712,7 +728,6 @@ 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)
 
     def forward(self, embeddings, lengths, image_shapes, h_coords, w_coords) -> torch.Tensor:
         """
@@ -839,8 +854,8 @@ 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,
@@ -975,7 +990,7 @@ class Glm4vMoeVisionModel(Glm4vMoePreTrainedModel):
         rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
         return rotary_pos_emb, pos_ids
 
-    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor) -> torch.Tensor:
+    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
         """
         Args:
             hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
@@ -1762,6 +1777,7 @@ 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
@@ -1778,13 +1794,14 @@ 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
 

transformers/models/glm4v_moe/modular_glm4v_moe.py

@@ -18,6 +18,7 @@ from typing import Optional, Union
 import torch
 import torch.nn as nn
 
+from ... import initialization as init
 from ...cache_utils import Cache, DynamicCache
 from ...configuration_utils import PreTrainedConfig
 from ...masking_utils import create_causal_mask
@@ -46,6 +47,7 @@ from ..glm4v.modeling_glm4v import (
     Glm4vTextModel,
     Glm4vTextRotaryEmbedding,
     Glm4vVisionModel,
+    Glm4vVisionRotaryEmbedding,
     rotate_half,
 )
 from ..qwen3_vl_moe.modeling_qwen3_vl_moe import (
@@ -227,7 +229,7 @@ class Glm4vMoeTextConfig(Glm4MoeConfig, RotaryEmbeddingConfigMixin):
         self.norm_topk_prob = norm_topk_prob
         self.router_aux_loss_coef = router_aux_loss_coef
         PreTrainedConfig.__init__(
-            self, tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope"}, **kwargs
+            self, tie_word_embeddings=tie_word_embeddings, ignore_keys_at_rope_validation={"mrope_section"}, **kwargs
         )
 
 
@@ -479,11 +481,21 @@ class Glm4vMoePreTrainedModel(Glm4MoePreTrainedModel):
         "router_logits": OutputRecorder(nn.Linear, layer_name="mlp.gate", index=0),
     }
 
+    def _init_weights(self, module):
+        super()._init_weights(module)
+        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)
+
 
 class Glm4vMoeCausalLMOutputWithPast(Qwen3VLMoeCausalLMOutputWithPast):
     pass
 
 
+class Glm4vMoeVisionRotaryEmbedding(Glm4vVisionRotaryEmbedding):
+    pass
+
+
 @auto_docstring
 class Glm4vMoeVisionModel(Glm4vVisionModel):
     pass

transformers/models/glmasr/__init__.py

@@ -0,0 +1,30 @@
+# coding=utf-8
+# Copyright 2025 the HuggingFace 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.
+
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_glmasr import *
+    from .modeling_glmasr import *
+    from .processing_glmasr import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)