transformers 5.0.0rc3__py3-none-any.whl → 5.1.0__py3-none-any.whl

transformers/models/glm_ocr/modular_glm_ocr.py

@@ -0,0 +1,428 @@
+# Copyright 2026 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 collections.abc import Callable
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...modeling_outputs import BaseModelOutputWithPooling
+from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
+from ..glm4v.configuration_glm4v import Glm4vConfig, Glm4vTextConfig, Glm4vVisionConfig
+from ..glm4v.modeling_glm4v import (
+    Glm4vForConditionalGeneration,
+    Glm4VisionMlp,
+    Glm4vModel,
+    Glm4vModelOutputWithPast,
+    Glm4vPreTrainedModel,
+    Glm4vRMSNorm,
+    Glm4vTextAttention,
+    Glm4vVisionAttention,
+    Glm4vVisionBlock,
+    Glm4vVisionModel,
+    Glm4vVisionPatchMerger,
+    apply_rotary_pos_emb_vision,
+    eager_attention_forward,
+    is_flash_attention_requested,
+)
+
+
+class GlmOcrRMSNorm(Glm4vRMSNorm):
+    pass
+
+
+class GlmOcrVisionMlp(Glm4VisionMlp):
+    def __init__(self, config, bias: bool = True):
+        super().__init__(config)
+        self.intermediate_size = config.intermediate_size
+
+
+class GlmOcrVisionConfig(Glm4vVisionConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`GlmOcrVisionConfig`]. It is used to instantiate a
+    GLM-OCR model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of
+    GLM-OCR [zai-org/GLM-OCR](https://huggingface.co/zai-org/GLM-OCR).
+
+    Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PreTrainedConfig`] for more information.
+
+    Args:
+        depth (`int`, *optional*, defaults to 24):
+            Number of layers (depth) in the model.
+        hidden_size (`int`, *optional*, defaults to 1024):
+            Dimensionality of the encoder layers and the pooler layer.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"silu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        attention_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries, keys and values.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            Dropout probability for attention weights.
+        num_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer architecture.
+        in_channels (`int`, *optional*, defaults to 3):
+            Number of input channels.
+        image_size (`int` or `list[int]`, *optional*, defaults to 336):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 14):
+            The size (resolution) of each patch.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the rms normalization layers.
+        spatial_merge_size (`int`, *optional*, defaults to 2):
+            The size used for merging spatial dimensions.
+        temporal_patch_size (`int`, *optional*, defaults to 2):
+            The size used for patches along the temporal dimension.
+        out_hidden_size (`int`, *optional*, defaults to 1536):
+            The output hidden size of the vision model.
+        intermediate_size (`int`, *optional*, defaults to 4096):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+    """
+
+    def __init__(
+        self,
+        depth=24,
+        hidden_size=1024,
+        hidden_act="silu",
+        attention_bias=True,
+        num_heads=16,
+        image_size=336,
+        out_hidden_size=1536,
+        intermediate_size=4096,
+        **super_kwargs,
+    ):
+        super().__init__(**super_kwargs)
+
+
+class GlmOcrTextConfig(Glm4vTextConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`GlmOcrTextConfig`]. It is used to instantiate a
+    GLM-OCR model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of
+    GLM-OCR [zai-org/GLM-OCR](https://huggingface.co/zai-org/GLM-OCR).
+
+    Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PreTrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 59392):
+            Vocabulary size of the GlmOcr model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`GlmOcrModel`]
+        hidden_size (`int`, *optional*, defaults to 1024):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 4096):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 16):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 131072):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        rope_parameters (`RopeParameters`, *optional*):
+            Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain
+            a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE
+            with longer `max_position_embeddings`.
+        pad_token_id (`int`, *optional*):
+            The id of the padding token.
+
+    ```python
+    >>> from transformers import GlmOcrTextModel, GlmOcrConfig
+
+    >>> # Initializing a GLM-OCR style configuration
+    >>> configuration = GlmOcrConfig()
+
+    >>> # Initializing a model from the GLM-OCR style configuration
+    >>> model = GlmOcrTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    def __init__(
+        self,
+        vocab_size: int | None = 59392,
+        hidden_size: int | None = 1024,
+        intermediate_size: int | None = 4096,
+        num_hidden_layers: int | None = 16,
+        num_attention_heads: int | None = 16,
+        num_key_value_heads: int | None = 8,
+        max_position_embeddings: int | None = 131072,
+        **super_kwargs,
+    ):
+        super().__init__(**super_kwargs)
+
+
+class GlmOcrConfig(Glm4vConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`GlmOcrModel`]. It is used to instantiate a
+    GLM-OCR model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of
+    GLM-OCR [zai-org/GLM-OCR](https://huggingface.co/zai-org/GLM-OCR).
+
+    Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PreTrainedConfig`] for more information.
+
+    Args:
+        text_config (`Union[PreTrainedConfig, dict]`, *optional*, defaults to `GlmOcrTextConfig`):
+            The config object or dictionary of the text backbone.
+        vision_config (`Union[PreTrainedConfig, dict]`,  *optional*, defaults to `GlmOcrVisionConfig`):
+            The config object or dictionary of the vision backbone.
+        image_token_id (`int`, *optional*, defaults to 59280):
+            The image token index to encode the image prompt.
+        video_token_id (`int`, *optional*, defaults to 59281):
+            The video token index to encode the image prompt.
+        image_start_token_id (`int`, *optional*, defaults to 59256):
+            The image start token index to encode the start of image.
+        image_end_token_id (`int`, *optional*, defaults to 59257):
+            The image end token index to encode the end of image.
+        video_start_token_id (`int`, *optional*, defaults to 59258):
+            The video start token index to encode the start of video.
+        video_end_token_id (`int`, *optional*, defaults to 59259):
+            The video end token index to encode the end of video.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+
+    ```python
+    >>> from transformers import GlmOcrForConditionalGeneration, GlmOcrConfig
+
+    >>> # Initializing a GLM-OCR style configuration
+    >>> configuration = GlmOcrConfig()
+
+    >>> # Initializing a model from the GLM-OCR style configuration
+    >>> model = GlmOcrForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    def __init__(
+        self,
+        text_config=None,
+        vision_config=None,
+        image_token_id=59280,
+        video_token_id=59281,
+        image_start_token_id=59256,
+        image_end_token_id=59257,
+        video_start_token_id=59258,
+        video_end_token_id=59259,
+        tie_word_embeddings=False,
+        **super_kwargs,
+    ):
+        super().__init__(**super_kwargs)
+
+
+class GlmOcrTextAttention(Glm4vTextAttention, nn.Module):
+    def __init__(self, config: GlmOcrTextConfig, layer_idx: int | None = None):
+        super().__init__()
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+
+
+class GlmOcrPreTrainedModel(Glm4vPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"model\.language_model\.layers\.16.*"]
+
+
+class GlmOcrModelOutputWithPast(Glm4vModelOutputWithPast):
+    pass
+
+
+class GlmOcrVisionAttention(Glm4vVisionAttention):
+    def __init__(self, config: GlmOcrVisionConfig) -> None:
+        super().__init__()
+        self.qkv = nn.Linear(config.hidden_size, config.hidden_size * 3, bias=config.attention_bias)
+        self.proj = nn.Linear(config.hidden_size, config.hidden_size, bias=config.attention_bias)
+        self.q_norm = GlmOcrRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        self.k_norm = GlmOcrRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        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]
+        query_states, key_states, value_states = (
+            self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
+        )
+
+        query_states = self.q_norm(query_states)
+        key_states = self.k_norm(key_states)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb_vision(query_states, key_states, cos, sin)
+        query_states = query_states.transpose(0, 1).unsqueeze(0)
+        key_states = key_states.transpose(0, 1).unsqueeze(0)
+        value_states = value_states.transpose(0, 1).unsqueeze(0)
+
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
+
+        if is_flash_attention_requested(self.config):
+            # Flash Attention: Use cu_seqlens for variable length attention
+            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
+            attn_output, _ = attention_interface(
+                self,
+                query_states,
+                key_states,
+                value_states,
+                attention_mask=None,
+                scaling=self.scaling,
+                dropout=0.0 if not self.training else self.attention_dropout,
+                cu_seq_lens_q=cu_seqlens,
+                cu_seq_lens_k=cu_seqlens,
+                max_length_q=max_seqlen,
+                max_length_k=max_seqlen,
+                is_causal=False,
+                **kwargs,
+            )
+        else:
+            # Other implementations: Process each chunk separately
+            lengths = cu_seqlens[1:] - cu_seqlens[:-1]
+            splits = [
+                torch.split(tensor, lengths.tolist(), dim=2) for tensor in (query_states, key_states, value_states)
+            ]
+
+            attn_outputs = [
+                attention_interface(
+                    self,
+                    q,
+                    k,
+                    v,
+                    attention_mask=None,
+                    scaling=self.scaling,
+                    dropout=0.0 if not self.training else self.attention_dropout,
+                    is_causal=False,
+                    **kwargs,
+                )[0]
+                for q, k, v in zip(*splits)
+            ]
+            attn_output = torch.cat(attn_outputs, dim=1)
+
+        attn_output = attn_output.reshape(seq_length, -1).contiguous()
+        attn_output = self.proj(attn_output)
+        return attn_output
+
+
+class GlmOcrVisionBlock(Glm4vVisionBlock):
+    def __init__(self, config) -> None:
+        super().__init__()
+        self.mlp = GlmOcrVisionMlp(config, bias=config.attention_bias)
+
+
+class GlmOcrVisionPatchMerger(Glm4vVisionPatchMerger):
+    pass
+
+
+class GlmOcrVisionModel(Glm4vVisionModel):
+    def __init__(self, config) -> None:
+        super().__init__(config)
+        del self.embeddings
+        del self.post_conv_layernorm
+        self.merger = GlmOcrVisionPatchMerger(
+            dim=config.out_hidden_size,
+            context_dim=config.out_hidden_size * config.in_channels,
+            hidden_act=config.hidden_act,
+        )
+
+    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
+        r"""
+        hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
+            The final hidden states of the model.
+        grid_thw (`torch.Tensor` of shape `(num_images_or_videos, 3)`):
+            The temporal, height and width of feature shape of each image in LLM.
+
+        Returns:
+            `torch.Tensor`: hidden_states.
+        """
+        hidden_states = self.patch_embed(hidden_states)
+        rotary_pos_emb, image_type_ids = self.rot_pos_emb(grid_thw)
+        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
+        position_embeddings = (emb.cos(), emb.sin())
+
+        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
+            dim=0,
+            # Select dtype based on the following factors:
+            #  - FA2 requires that cu_seqlens_q must have dtype int32
+            #  - torch.onnx.export requires that cu_seqlens_q must have same dtype as grid_thw
+            # See https://github.com/huggingface/transformers/pull/34852 for more information
+            dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
+        )
+        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+
+        for blk in self.blocks:
+            hidden_states = blk(
+                hidden_states,
+                cu_seqlens=cu_seqlens,
+                position_embeddings=position_embeddings,
+            )
+
+        hidden_states = self.post_layernorm(hidden_states)
+
+        hidden_states = hidden_states.view(
+            -1, self.spatial_merge_size, self.spatial_merge_size, hidden_states.shape[-1]
+        )
+        hidden_states = hidden_states.permute(0, 3, 1, 2)
+        hidden_states = self.downsample(hidden_states).view(-1, self.config.out_hidden_size)
+
+        merged_hidden_states = self.merger(hidden_states)
+        return BaseModelOutputWithPooling(
+            last_hidden_state=hidden_states,
+            pooler_output=merged_hidden_states,
+        )
+
+
+class GlmOcrModel(Glm4vModel):
+    pass
+
+
+class GlmOcrForConditionalGeneration(Glm4vForConditionalGeneration):
+    pass
+
+
+__all__ = [
+    "GlmOcrConfig",
+    "GlmOcrTextConfig",
+    "GlmOcrVisionConfig",
+    "GlmOcrTextModel",  # noqa: F822
+    "GlmOcrVisionModel",
+    "GlmOcrModel",
+    "GlmOcrPreTrainedModel",
+    "GlmOcrForConditionalGeneration",
+]

transformers/models/glmasr/modeling_glmasr.py

@@ -26,12 +26,12 @@ from ...cache_utils import Cache
 from ...generation import GenerationMixin
 from ...integrations import use_kernelized_func
 from ...modeling_layers import GradientCheckpointingLayer
-from ...modeling_outputs import BaseModelOutput, CausalLMOutputWithPast
+from ...modeling_outputs import BaseModelOutputWithPooling, 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, is_torch_available
-from ...utils.generic import check_model_inputs, maybe_autocast
+from ...utils import TransformersKwargs, auto_docstring, is_torch_available
+from ...utils.generic import can_return_tuple, check_model_inputs, maybe_autocast
 from ..auto import AutoModel, AutoModelForCausalLM
 from .configuration_glmasr import GlmAsrConfig, GlmAsrEncoderConfig
 
@@ -205,9 +205,9 @@ class GlmAsrAttention(nn.Module):
         cos, sin = position_embeddings
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -292,6 +292,10 @@ class GlmAsrEncoder(GlmAsrPreTrainedModel):
     main_input_name = "input_features"
     input_modalities = "audio"
     _no_split_modules = ["GlmAsrEncoderLayer"]
+    _can_record_outputs = {
+        "hidden_states": GlmAsrEncoderLayer,
+        "attentions": GlmAsrAttention,
+    }
 
     def __init__(self, config: GlmAsrEncoderConfig):
         super().__init__(config)
@@ -322,7 +326,7 @@ class GlmAsrEncoder(GlmAsrPreTrainedModel):
             hidden_states = encoder_layer(hidden_states, position_embeddings=position_embeddings, **kwargs)
 
         hidden_states = self.norm(hidden_states)
-        return BaseModelOutput(last_hidden_state=hidden_states)
+        return BaseModelOutputWithPooling(last_hidden_state=hidden_states)
 
 
 class GlmAsrMultiModalProjector(nn.Module):
@@ -382,26 +386,27 @@ class GlmAsrForConditionalGeneration(GlmAsrPreTrainedModel, GenerationMixin):
     def get_decoder(self):
         return self.language_model.get_decoder()
 
+    @can_return_tuple
+    @auto_docstring(
+        custom_intro="Compute audio embeddings from log-mel input features using the audio encoder and multi-modal projector."
+    )
     def get_audio_features(
-        self, input_features: torch.FloatTensor, input_features_mask: torch.Tensor
-    ) -> torch.FloatTensor:
-        """
-        This method is used to get the audio embeddings from input features (a log mel spectrogram), meaning inferring the audio encoder and the multi-modal projector.
-        Args:
-            input_features (`torch.FloatTensor`):
-                Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
-                obtained by loading a `.flac` or `.wav` audio file into an array of type `list[float]` or a
-                `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
-                `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
-                and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
-            input_features_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
-                Mask to avoid performing attention on padded feature indices.
-
-        Returns:
-            `torch.FloatTensor`:
-                The audio embeddings.
+        self,
+        input_features: torch.FloatTensor,
+        input_features_mask: torch.Tensor,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        input_features (`torch.FloatTensor`):
+            Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
+            obtained by loading a `.flac` or `.wav` audio file into an array of type `list[float]` or a
+            `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
+            `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
+            and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
+        input_features_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
+            Mask to avoid performing attention on padded feature indices.
         """
-        audio_outputs = self.audio_tower(input_features)
+        audio_outputs = self.audio_tower(input_features, return_dict=True, **kwargs)
         audio_hidden_states = audio_outputs.last_hidden_state
         audio_hidden_states = audio_hidden_states.reshape(
             input_features.shape[0], -1, self.config.audio_config.intermediate_size
@@ -415,8 +420,9 @@ class GlmAsrForConditionalGeneration(GlmAsrPreTrainedModel, GenerationMixin):
         post_lengths = (audio_lengths - merge_factor) // merge_factor + 1
 
         valid_mask = torch.arange(audio_embeds.shape[1], device=post_lengths.device)[None, :] < post_lengths[:, None]
-        audio_embeds = audio_embeds[valid_mask.to(audio_embeds.device)]
-        return audio_embeds
+        audio_outputs.pooler_output = audio_embeds[valid_mask.to(audio_embeds.device)]
+
+        return audio_outputs
 
     @can_return_tuple
     @auto_docstring
@@ -468,7 +474,7 @@ class GlmAsrForConditionalGeneration(GlmAsrPreTrainedModel, GenerationMixin):
             inputs_embeds = self.get_input_embeddings()(input_ids)
 
         if input_features is not None and input_ids is not None:
-            audio_embeds = self.get_audio_features(input_features, input_features_mask)
+            audio_embeds = self.get_audio_features(input_features, input_features_mask, return_dict=True).pooler_output
 
             # replace text-audio token placeholders with audio embeddings
             audio_token_mask = (input_ids == self.config.audio_token_id).unsqueeze(-1)

transformers/models/glmasr/modular_glmasr.py

@@ -21,11 +21,11 @@ from ...audio_utils import AudioInput, make_list_of_audio
 from ...cache_utils import Cache
 from ...feature_extraction_utils import BatchFeature
 from ...modeling_layers import GradientCheckpointingLayer
-from ...modeling_outputs import BaseModelOutput, CausalLMOutputWithPast
+from ...modeling_outputs import BaseModelOutputWithPooling, CausalLMOutputWithPast
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, is_torch_available, logging
-from ...utils.generic import check_model_inputs
+from ...utils.generic import can_return_tuple, check_model_inputs
 from ..audioflamingo3.modeling_audioflamingo3 import (
     AudioFlamingo3ForConditionalGeneration,
     AudioFlamingo3MultiModalProjector,
@@ -227,9 +227,9 @@ class GlmAsrAttention(LlamaAttention):
         cos, sin = position_embeddings
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -305,6 +305,10 @@ class GlmAsrEncoder(GlmAsrPreTrainedModel):
     main_input_name = "input_features"
     input_modalities = "audio"
     _no_split_modules = ["GlmAsrEncoderLayer"]
+    _can_record_outputs = {
+        "hidden_states": GlmAsrEncoderLayer,
+        "attentions": GlmAsrAttention,
+    }
 
     def __init__(self, config: GlmAsrEncoderConfig):
         super().__init__(config)
@@ -335,7 +339,7 @@ class GlmAsrEncoder(GlmAsrPreTrainedModel):
             hidden_states = encoder_layer(hidden_states, position_embeddings=position_embeddings, **kwargs)
 
         hidden_states = self.norm(hidden_states)
-        return BaseModelOutput(last_hidden_state=hidden_states)
+        return BaseModelOutputWithPooling(last_hidden_state=hidden_states)
 
 
 class GlmAsrMultiModalProjector(AudioFlamingo3MultiModalProjector):
@@ -351,10 +355,17 @@ class GlmAsrMultiModalProjector(AudioFlamingo3MultiModalProjector):
     """
 )
 class GlmAsrForConditionalGeneration(AudioFlamingo3ForConditionalGeneration):
+    @can_return_tuple
+    @auto_docstring(
+        custom_intro="Compute audio embeddings from log-mel input features using the audio encoder and multi-modal projector."
+    )
     def get_audio_features(
-        self, input_features: torch.FloatTensor, input_features_mask: torch.Tensor
-    ) -> torch.FloatTensor:
-        audio_outputs = self.audio_tower(input_features)
+        self,
+        input_features: torch.FloatTensor,
+        input_features_mask: torch.Tensor,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        audio_outputs = self.audio_tower(input_features, return_dict=True, **kwargs)
         audio_hidden_states = audio_outputs.last_hidden_state
         audio_hidden_states = audio_hidden_states.reshape(
             input_features.shape[0], -1, self.config.audio_config.intermediate_size
@@ -368,8 +379,9 @@ class GlmAsrForConditionalGeneration(AudioFlamingo3ForConditionalGeneration):
         post_lengths = (audio_lengths - merge_factor) // merge_factor + 1
 
         valid_mask = torch.arange(audio_embeds.shape[1], device=post_lengths.device)[None, :] < post_lengths[:, None]
-        audio_embeds = audio_embeds[valid_mask.to(audio_embeds.device)]
-        return audio_embeds
+        audio_outputs.pooler_output = audio_embeds[valid_mask.to(audio_embeds.device)]
+
+        return audio_outputs
 
     def forward(
         self,

transformers/models/glpn/image_processing_glpn_fast.py

@@ -16,7 +16,7 @@
 from typing import Optional
 
 import torch
-from torchvision.transforms.v2 import functional as F
+import torchvision.transforms.v2.functional as tvF
 
 from ...image_processing_utils import BatchFeature
 from ...image_processing_utils_fast import BaseImageProcessorFast, group_images_by_shape, reorder_images
@@ -50,7 +50,7 @@ class GLPNImageProcessorFast(BaseImageProcessorFast):
         self,
         image: "torch.Tensor",
         size_divisor: int,
-        interpolation: Optional["F.InterpolationMode"] = None,
+        interpolation: Optional["tvF.InterpolationMode"] = None,
         antialias: bool = True,
         **kwargs,
     ) -> "torch.Tensor":
@@ -83,7 +83,7 @@ class GLPNImageProcessorFast(BaseImageProcessorFast):
         images: list["torch.Tensor"],
         do_resize: bool,
         size_divisor: int | None = None,
-        interpolation: Optional["F.InterpolationMode"] = None,
+        interpolation: Optional["tvF.InterpolationMode"] = None,
         do_rescale: bool = True,
         rescale_factor: float | None = 1 / 255,
         do_normalize: bool = False,

transformers/models/glpn/modeling_glpn.py

@@ -609,10 +609,12 @@ class GLPNForDepthEstimation(GLPNPreTrainedModel):
         >>> import torch
         >>> import numpy as np
         >>> from PIL import Image
-        >>> import requests
+        >>> import httpx
+        >>> from io import BytesIO
 
         >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
-        >>> image = Image.open(requests.get(url, stream=True).raw)
+        >>> with httpx.stream("GET", url) as response:
+        ...     image = Image.open(BytesIO(response.read()))
 
         >>> image_processor = AutoImageProcessor.from_pretrained("vinvino02/glpn-kitti")
         >>> model = GLPNForDepthEstimation.from_pretrained("vinvino02/glpn-kitti")