transformers 5.0.0rc1__py3-none-any.whl → 5.0.0rc2__py3-none-any.whl

transformers/models/git/modeling_git.py

@@ -26,8 +26,9 @@ from torch import nn
 from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
+from ...configuration_utils import PreTrainedConfig
 from ...generation import GenerationMixin
-from ...modeling_attn_mask_utils import _prepare_4d_attention_mask
+from ...masking_utils import create_masks_for_generate
 from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -69,6 +70,104 @@ class GitVisionModelOutput(ModelOutput):
     attentions: Optional[tuple[torch.FloatTensor, ...]] = None
 
 
+# Copied from transformers.models.gemma3.modeling_gemma3.token_type_ids_mask_function
+def token_type_ids_mask_function(
+    token_type_ids: Optional[torch.Tensor],
+    image_group_ids: Optional[torch.Tensor],
+) -> Optional[Callable]:
+    """
+    This function adds the correct offsets to the `q_idx` and `kv_idx` as the torch API can only accept lengths,
+    not start and end indices.
+    """
+    # Do not return an additional mask in this case
+    if token_type_ids is None:
+        return None
+
+    def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
+        # If it's 1 for both query and key/value, we are in an image block
+        # NOTE: static cache shape goes beyond input seq length, while token_type_ids.shape[1] == input seq length
+        # Since vmap doesn't support `if statement` we workaround it with `torch.where`
+        safe_q_idx = torch.where(q_idx < token_type_ids.shape[1], q_idx, 0)
+        safe_kv_idx = torch.where(kv_idx < token_type_ids.shape[1], kv_idx, 0)
+
+        token_type_ids_at_q_idx = token_type_ids[batch_idx, safe_q_idx]
+        token_type_ids_at_q_idx = torch.where(q_idx < token_type_ids.shape[1], token_type_ids_at_q_idx, 0)
+
+        token_type_ids_at_kv_idx = token_type_ids[batch_idx, safe_kv_idx]
+        token_type_ids_at_kv_idx = torch.where(kv_idx < token_type_ids.shape[1], token_type_ids_at_kv_idx, 0)
+
+        image_group_ids_at_q_idx = image_group_ids[batch_idx, safe_q_idx]
+        image_group_ids_at_q_idx = torch.where(q_idx < image_group_ids.shape[1], image_group_ids_at_q_idx, -1)
+
+        image_group_ids_at_kv_idx = image_group_ids[batch_idx, safe_kv_idx]
+        image_group_ids_at_kv_idx = torch.where(kv_idx < image_group_ids.shape[1], image_group_ids_at_kv_idx, -1)
+
+        is_image_block = (token_type_ids_at_q_idx == 1) & (token_type_ids_at_kv_idx == 1)
+        same_image_block = image_group_ids_at_q_idx == image_group_ids_at_kv_idx
+
+        # This is bidirectional attention whenever we are dealing with image tokens
+        return is_image_block & same_image_block
+
+    return inner_mask
+
+
+# Copied from transformers.models.gemma3.modeling_gemma3.create_causal_mask_mapping
+def create_causal_mask_mapping(
+    config: PreTrainedConfig,
+    input_embeds: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    cache_position: torch.Tensor,
+    past_key_values: Optional[Cache],
+    position_ids: Optional[torch.Tensor],
+    token_type_ids: Optional[torch.Tensor] = None,
+    pixel_values: Optional[torch.FloatTensor] = None,
+    is_training: bool = False,
+    is_first_iteration: Optional[bool] = None,
+    **kwargs,
+) -> dict:
+    """
+    Overwrites the base `create_masks_for_generate` with `token_type_ids` masking to create the causal mask mapping
+    for all kinds of forward passes. Gemma3 uses a bidirectional mask for images.
+
+    Uses `pixel_values` as an optional input to disambiguate edge cases.
+    """
+    if is_training and token_type_ids is None:
+        raise ValueError("`token_type_ids` is required as a model input when training")
+
+    mask_kwargs = {
+        "config": config.get_text_config(),
+        "input_embeds": input_embeds,
+        "attention_mask": attention_mask,
+        "cache_position": cache_position,
+        "past_key_values": past_key_values,
+        "position_ids": position_ids,
+    }
+    # NOTE: this `may_have_image_input` logic is not flawless, it fails when we're using a cache eagerly initialized
+    # (e.g. compiled prefill) AND `pixel_values` are not provided (i.e. the image data is provided through other
+    # means). Determining prefill in that case requires checking data values, which is not compile-compatible.
+    is_first_iteration = (
+        is_first_iteration
+        if is_first_iteration is not None
+        else (past_key_values is None or not past_key_values.is_initialized or pixel_values is not None)
+    )
+    if token_type_ids is not None and is_first_iteration:
+        # We need to pass an additional mask function to account for token type ids, and it needs to be an `or` (to
+        # undo the causal masking)
+
+        # First find where a new image block starts: 1 if image and previous not image
+        # The images cannot attend to future images, but can attend to all prev images and to itself bidirectionally
+        is_image = (token_type_ids == 1).to(cache_position.device)
+        is_previous_image = nn.functional.pad(is_image, (1, 0), value=0)[:, :-1]
+        new_image_start = is_image & ~is_previous_image
+        image_group_ids = torch.cumsum(new_image_start.int(), dim=1) - 1
+        image_group_ids = torch.where(is_image, image_group_ids, -1)
+        mask_kwargs["or_mask_function"] = token_type_ids_mask_function(
+            token_type_ids.to(cache_position.device), image_group_ids
+        )
+
+    return create_masks_for_generate(**mask_kwargs)
+
+
 class GitEmbeddings(nn.Module):
     """Construct the embeddings from word and position embeddings."""
 
@@ -148,17 +247,15 @@ class GitSelfAttention(nn.Module):
         hidden_states: torch.Tensor,
         attention_mask: Optional[torch.FloatTensor] = None,
         past_key_values: Optional[Cache] = None,
-        output_attentions: Optional[bool] = False,
-        pixel_values_present: Optional[bool] = False,
+        cache_position: Optional[torch.Tensor] = None,
     ) -> tuple[torch.Tensor]:
-        batch_size, seq_length, _ = hidden_states.shape
+        batch_size = hidden_states.shape[0]
         query_layer = (
             self.query(hidden_states)
             .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)
             .transpose(1, 2)
         )
 
-        cutoff = self.image_patch_tokens if pixel_values_present else 0
         key_layer = (
             self.key(hidden_states)
             .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)
@@ -170,12 +267,9 @@ class GitSelfAttention(nn.Module):
             .transpose(1, 2)
         )
         if past_key_values is not None:
-            # NOTE: like in other caches, we store the text component. In GIT it means we discard the image component.
-            key_layer_past, value_layer_past = past_key_values.update(
-                key_layer[:, :, cutoff:, :], value_layer[:, :, cutoff:, :], self.layer_idx
+            key_layer, value_layer = past_key_values.update(
+                key_layer, value_layer, self.layer_idx, cache_kwargs={"cache_position": cache_position}
             )
-            key_layer = torch.cat([key_layer[:, :, :cutoff, :], key_layer_past], dim=2)
-            value_layer = torch.cat([value_layer[:, :, :cutoff, :], value_layer_past], dim=2)
 
         # Take the dot product between "query" and "key" to get the raw attention scores.
         attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
@@ -232,15 +326,14 @@ class GitAttention(nn.Module):
         hidden_states: torch.Tensor,
         attention_mask: Optional[torch.FloatTensor] = None,
         past_key_values: Optional[Cache] = None,
+        cache_position: Optional[torch.Tensor] = None,
         output_attentions: Optional[bool] = False,
-        pixel_values_present: Optional[bool] = False,
     ) -> tuple[torch.Tensor]:
         attn_output, self_attn_weights = self.self(
             hidden_states,
             attention_mask,
             past_key_values,
-            output_attentions,
-            pixel_values_present,
+            cache_position=cache_position,
         )
         attention_output = self.output(attn_output, hidden_states)
         return attention_output, self_attn_weights
@@ -291,8 +384,8 @@ class GitLayer(GradientCheckpointingLayer):
         hidden_states: torch.Tensor,
         attention_mask: Optional[torch.FloatTensor] = None,
         past_key_values: Optional[Cache] = None,
+        cache_position: Optional[torch.Tensor] = None,
         output_attentions: Optional[bool] = False,
-        pixel_values_present: Optional[bool] = False,
     ) -> tuple[torch.Tensor]:
         # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
         attention_output, self_attention_weights = self.attention(
@@ -300,7 +393,7 @@ class GitLayer(GradientCheckpointingLayer):
             attention_mask,
             output_attentions=output_attentions,
             past_key_values=past_key_values,
-            pixel_values_present=pixel_values_present,
+            cache_position=cache_position,
         )
 
         layer_output = apply_chunking_to_forward(
@@ -329,8 +422,8 @@ class GitEncoder(nn.Module):
         use_cache: Optional[bool] = None,
         output_attentions: Optional[bool] = False,
         output_hidden_states: Optional[bool] = False,
-        pixel_values_present: Optional[bool] = False,
         return_dict: Optional[bool] = True,
+        cache_position: Optional[torch.Tensor] = None,
     ) -> Union[tuple[torch.Tensor], BaseModelOutputWithPast]:
         if self.gradient_checkpointing and self.training:
             if use_cache:
@@ -353,7 +446,7 @@ class GitEncoder(nn.Module):
                 attention_mask,
                 past_key_values,
                 output_attentions,
-                pixel_values_present,
+                cache_position,
             )
 
             hidden_states = layer_outputs[0]
@@ -396,6 +489,7 @@ class GitPreTrainedModel(PreTrainedModel):
             init.normal_(module.class_embedding, mean=0.0, std=self.config.initializer_range)
             init.normal_(module.patch_embedding.weight, std=self.config.initializer_range)
             init.normal_(module.position_embedding.weight, std=self.config.initializer_range)
+            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
         if isinstance(module, nn.Linear):
             init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
             if module.bias is not None:
@@ -408,6 +502,8 @@ class GitPreTrainedModel(PreTrainedModel):
         elif isinstance(module, nn.LayerNorm):
             init.zeros_(module.bias)
             init.ones_(module.weight)
+        elif isinstance(module, GitEmbeddings):
+            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
 
 
 # Copied from transformers.models.clip.modeling_clip.CLIPVisionEmbeddings with CLIP->Git
@@ -903,62 +999,6 @@ class GitModel(GitPreTrainedModel):
     def set_input_embeddings(self, value):
         self.embeddings.word_embeddings = value
 
-    def _generate_future_mask(self, size: int, dtype: torch.dtype, device: torch.device) -> torch.Tensor:
-        # Default mask is for forward direction. Flip for backward direction.
-        mask = torch.triu(torch.ones(size, size, device=device, dtype=dtype), diagonal=1)
-        mask = mask.masked_fill(mask == 1, float("-inf"))
-        return mask
-
-    def create_attention_mask(self, tgt, memory, tgt_mask, past_key_values_length, memory_key_padding_mask=None):
-        num_tgt = tgt.shape[1]
-        num_memory = memory.shape[1]
-        device = tgt.device
-        dtype = tgt.dtype
-        top_left = torch.zeros((num_memory, num_memory), device=device, dtype=dtype)
-        top_right = torch.full(
-            (num_memory, num_tgt + past_key_values_length),
-            float("-inf"),
-            device=tgt.device,
-            dtype=dtype,
-        )
-        bottom_left = torch.zeros(
-            (num_tgt, num_memory),
-            dtype=dtype,
-            device=tgt_mask.device,
-        )
-
-        if past_key_values_length > 0:
-            tgt_mask = torch.zeros(
-                (tgt_mask.shape[0], tgt_mask.shape[0] + past_key_values_length),
-                dtype=dtype,
-                device=tgt_mask.device,
-            )
-
-        left = torch.cat((top_left, bottom_left), dim=0)
-        right = torch.cat((top_right, tgt_mask.to(dtype)), dim=0)
-
-        full_attention_mask = torch.cat((left, right), dim=1)[None, :]
-
-        if memory_key_padding_mask is None:
-            memory_key_padding_mask = torch.full((memory.shape[0], memory.shape[1]), fill_value=False, device=device)
-        # if it is False, it means valid. That is, it is not a padding
-        if memory_key_padding_mask.dtype != torch.bool:
-            raise ValueError("Memory key padding mask must be a boolean tensor.")
-        zero_negative_infinity = torch.zeros_like(memory_key_padding_mask, dtype=tgt.dtype)
-        zero_negative_infinity[memory_key_padding_mask] = float("-inf")
-        full_attention_mask = full_attention_mask.expand(
-            (memory_key_padding_mask.shape[0], num_memory + num_tgt, num_memory + past_key_values_length + num_tgt)
-        )
-        full_attention_mask = full_attention_mask.clone()
-        origin_left = full_attention_mask[:, :, :num_memory]
-        update = zero_negative_infinity[:, None, :]
-        full_attention_mask[:, :, :num_memory] = origin_left + update
-
-        # add axis for multi-head
-        full_attention_mask = full_attention_mask[:, None, :, :]
-
-        return full_attention_mask
-
     @auto_docstring
     def forward(
         self,
@@ -973,6 +1013,7 @@ class GitModel(GitPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         interpolate_pos_encoding: bool = False,
         return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.Tensor] = None,
         **kwargs,
     ) -> Union[tuple[torch.Tensor], BaseModelOutputWithPooling]:
         r"""
@@ -1005,15 +1046,6 @@ class GitModel(GitPreTrainedModel):
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
-        elif input_ids is not None:
-            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
-            input_shape = input_ids.size()
-        elif inputs_embeds is not None:
-            input_shape = inputs_embeds.size()[:-1]
-        else:
-            raise ValueError("You have to specify either input_ids or inputs_embeds")
-
-        seq_length = input_shape[1]
 
         # past_key_values_length
         past_key_values_length = 0
@@ -1024,7 +1056,23 @@ class GitModel(GitPreTrainedModel):
                 else past_key_values.get_seq_length()
             )
 
-        projected_visual_features = None
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+
+        if cache_position is None:
+            cache_position = torch.arange(
+                past_key_values_length,
+                past_key_values_length + embedding_output.shape[1],
+                device=embedding_output.device,
+            )
+
+        # Always create `token_type_ids` so we can re-use Gemma3 style mask preparation fn
+        token_type_ids = torch.zeros_like(embedding_output, dtype=torch.int)[..., 0]
+
         if pixel_values is not None:
             if pixel_values.ndim == 4:
                 # here we assume pixel_values is of shape (batch_size, num_channels, height, width)
@@ -1050,60 +1098,54 @@ class GitModel(GitPreTrainedModel):
 
             projected_visual_features = self.visual_projection(visual_features)
 
-        embedding_output = self.embeddings(
-            input_ids=input_ids,
-            position_ids=position_ids,
-            inputs_embeds=inputs_embeds,
-            past_key_values_length=past_key_values_length,
-        )
-
-        if projected_visual_features is None:
-            projected_visual_features = torch.zeros(
-                (embedding_output.shape[0], 0, embedding_output.shape[2]),
-                dtype=embedding_output.dtype,
-                device=embedding_output.device,
+            # Repeat visual features to match embedding batch size.
+            projected_visual_features = projected_visual_features.repeat(
+                embedding_output.size(0) // projected_visual_features.size(0), 1, 1
             )
 
-        # Repeat visual features to match embedding batch size.
-        projected_visual_features = projected_visual_features.repeat(
-            embedding_output.size(0) // projected_visual_features.size(0), 1, 1
-        )
-
-        # concatenate patch token and text token embeddings
-        hidden_states = torch.cat((projected_visual_features, embedding_output), dim=1)
-
-        # By default, an additive causal mask is created
-        # for masking the future (one direction).
-        tgt_mask = self._generate_future_mask(seq_length, embedding_output.dtype, embedding_output.device)
+            # concatenate patch token and text token embeddings
+            embedding_output = torch.cat((projected_visual_features, embedding_output), dim=1)
+            image_token_type_ids = torch.ones_like(projected_visual_features, dtype=torch.int)[..., 0]
+            token_type_ids = torch.cat([image_token_type_ids, token_type_ids], dim=-1)
+            cache_position = torch.arange(embedding_output.shape[1], device=embedding_output.device, dtype=torch.int)
+            if attention_mask is not None:
+                attention_mask = torch.cat([torch.ones_like(image_token_type_ids), attention_mask], dim=-1)
+        elif past_key_values is not None and input_ids.shape[1] == 1:
+            # Expand attention mask and cache position with image tokens because GIT doesn't add image
+            # placeholder tokens when processing. Doesn't worth the refactor, low usage!
+            cache_position = torch.tensor(
+                [past_key_values_length], dtype=cache_position.dtype, device=cache_position.device
+            )
+            extended_attention_mask = torch.ones(
+                (attention_mask.shape[0], past_key_values_length - attention_mask.shape[1] + 1),
+                dtype=attention_mask.dtype,
+                device=attention_mask.device,
+            )
+            attention_mask = torch.cat([extended_attention_mask, attention_mask], dim=-1)
 
-        # Create an attention mask of shape (batch_size, 1, tgt_seq_len, src_seq_len)
-        combined_attention_mask = self.create_attention_mask(
-            tgt=embedding_output,
-            memory=projected_visual_features,
-            tgt_mask=tgt_mask,
-            past_key_values_length=past_key_values_length,
+        # Images attend each other bidirectionally while text remains causal
+        causal_mask = create_causal_mask_mapping(
+            self.config,
+            embedding_output,
+            attention_mask,
+            cache_position,
+            past_key_values,
+            None,
+            token_type_ids,
+            pixel_values,
         )
 
-        if attention_mask is not None:
-            # if the user provides an attention mask, we add it to the default one
-            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
-            expanded_attn_mask = _prepare_4d_attention_mask(
-                attention_mask, embedding_output.dtype, tgt_len=input_shape[-1]
-            ).to(embedding_output.device)
-            if past_key_values_length > 0:
-                expanded_attn_mask = expanded_attn_mask[:, :, -past_key_values_length:, :]
-            else:
-                combined_attention_mask[:, :, -input_shape[1] :, -input_shape[1] :] += expanded_attn_mask
+        hidden_states = embedding_output
 
         encoder_outputs = self.encoder(
             hidden_states,
-            attention_mask=combined_attention_mask,
+            attention_mask=causal_mask,
             past_key_values=past_key_values,
             use_cache=use_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
-            pixel_values_present=pixel_values is not None,
+            cache_position=cache_position,
         )
         sequence_output = encoder_outputs[0]
 
@@ -1157,6 +1199,7 @@ class GitForCausalLM(GitPreTrainedModel, GenerationMixin):
         interpolate_pos_encoding: bool = False,
         return_dict: Optional[bool] = None,
         logits_to_keep: Union[int, torch.Tensor] = 0,
+        cache_position: Optional[torch.Tensor] = None,
         **kwargs,
     ) -> Union[tuple[torch.Tensor], CausalLMOutputWithPast]:
         r"""
@@ -1306,6 +1349,7 @@ class GitForCausalLM(GitPreTrainedModel, GenerationMixin):
             output_hidden_states=output_hidden_states,
             interpolate_pos_encoding=interpolate_pos_encoding,
             return_dict=return_dict,
+            cache_position=cache_position,
         )
 
         hidden_states = outputs[0]
@@ -1339,7 +1383,15 @@ class GitForCausalLM(GitPreTrainedModel, GenerationMixin):
         )
 
     def prepare_inputs_for_generation(
-        self, input_ids, past_key_values=None, attention_mask=None, use_cache=None, **kwargs
+        self,
+        input_ids,
+        past_key_values=None,
+        pixel_values=None,
+        attention_mask=None,
+        use_cache=None,
+        cache_position=None,
+        is_first_iteration=False,
+        **kwargs,
     ):
         # Overwritten -- `git` has special cache handling and doesn't support generating from `inputs_embeds` atm
 
@@ -1364,11 +1416,14 @@ class GitForCausalLM(GitPreTrainedModel, GenerationMixin):
         model_inputs = {
             "input_ids": input_ids,
             "attention_mask": attention_mask,
-            "pixel_values": kwargs.get("pixel_values"),
             "past_key_values": past_key_values,
             "use_cache": use_cache,
+            "cache_position": cache_position,
         }
 
+        if is_first_iteration or not use_cache:
+            model_inputs["pixel_values"] = pixel_values
+
         # Forward ALL kwargs that are uninitialized (e.g. `use_cache`).
         for key, value in kwargs.items():
             if key not in model_inputs:

transformers/models/glm/modeling_glm.py

@@ -79,7 +79,7 @@ class GlmRotaryEmbedding(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(

transformers/models/glm4/modeling_glm4.py

@@ -284,7 +284,7 @@ class Glm4RotaryEmbedding(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(

transformers/models/glm46v/image_processing_glm46v.py

@@ -354,7 +354,6 @@ class Glm46VImageProcessor(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`):
@@ -381,12 +380,9 @@ class Glm46VImageProcessor(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/glm46v/modeling_glm46v.py

@@ -639,6 +639,7 @@ class Glm46VForConditionalGeneration(Glm46VPreTrainedModel, 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
@@ -655,13 +656,14 @@ class Glm46VForConditionalGeneration(Glm46VPreTrainedModel, 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/glm46v/modular_glm46v.py

@@ -110,6 +110,9 @@ class Glm46VPreTrainedModel(Glm4vPreTrainedModel):
     _can_record_outputs = None
     _no_split_modules = None
 
+    def _init_weights(self, module):
+        raise AttributeError("Not needed")
+
 
 class Glm46VModel(Glm4vModel):
     _no_split_modules = None

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, 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
@@ -38,7 +38,7 @@ 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 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(
@@ -332,6 +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."""
 
@@ -339,7 +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]
@@ -486,7 +487,9 @@ 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,
@@ -499,6 +502,7 @@ class Glm4MoePreTrainedModel(PreTrainedModel):
         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/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