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

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
@@ -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(
@@ -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
@@ -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
@@ -32,7 +33,7 @@ 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
@@ -409,7 +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:
         """
@@ -725,6 +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
@@ -1414,6 +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
@@ -1430,13 +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/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, use_kernelized_func
+from ...integrations import use_experts_implementation, use_kernel_forward_from_hub, use_kernelized_func
 from ...masking_utils import create_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
@@ -40,7 +40,13 @@ from ...modeling_outputs import ModelOutput, MoeModelOutputWithPast
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_torchdynamo_compiling
+from ...utils import (
+    TransformersKwargs,
+    auto_docstring,
+    can_return_tuple,
+    is_grouped_mm_available,
+    is_torchdynamo_compiling,
+)
 from ...utils.generic import OutputRecorder, check_model_inputs, maybe_autocast
 from .configuration_glm4v_moe import Glm4vMoeConfig, Glm4vMoeTextConfig, Glm4vMoeVisionConfig
 
@@ -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(
@@ -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 = {
@@ -602,9 +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
@@ -637,6 +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__()
@@ -671,20 +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__()
@@ -713,7 +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:
         """
@@ -840,8 +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,
@@ -1763,6 +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
@@ -1779,13 +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 (
@@ -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__)

transformers/models/glmasr/configuration_glmasr.py

@@ -0,0 +1,197 @@
+# 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 ...configuration_utils import PreTrainedConfig
+from ..auto import CONFIG_MAPPING, AutoConfig
+
+
+class GlmAsrEncoderConfig(PreTrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`GlmAsrEncoder`]. It is used to instantiate a
+    glmasr audio encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the audio encoder of the glmasr
+    architecture.
+
+    e.g. [zai-org/GLM-ASR-Nano-2512](https://huggingface.co/zai-org/GLM-ASR-Nano-2512)
+
+    Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PreTrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 1280):
+            Dimensionality of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 5120):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 20):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*):
+            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, check out [this
+            paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler.
+        max_position_embeddings (`int`, *optional*, defaults to 1500):
+            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.
+        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`.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        num_mel_bins (`int`, *optional*, defaults to 128):
+            Number of mel features used per input features. Should correspond to the value used in the
+            `GlmAsrProcessor` class.
+
+    ```python
+    >>> from transformers import GlmAsrEncoderConfig, GlmAsrEncoder
+
+    >>> # Initializing a GlmAsrEncoderConfig
+    >>> configuration = GlmAsrEncoderConfig()
+
+    >>> # Initializing a GlmAsrEncoder (with random weights)
+    >>> model = GlmAsrEncoder(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "glmasr_encoder"
+
+    def __init__(
+        self,
+        hidden_size=1280,
+        intermediate_size=5120,
+        num_hidden_layers=32,
+        num_attention_heads=20,
+        num_key_value_heads=None,
+        hidden_act="gelu",
+        max_position_embeddings=1500,
+        initializer_range=0.02,
+        rope_parameters=None,
+        attention_dropout=0.0,
+        num_mel_bins=128,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.head_dim = hidden_size // num_attention_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.rope_parameters = rope_parameters
+        self.attention_dropout = attention_dropout
+        self.num_mel_bins = num_mel_bins
+
+        kwargs.setdefault("partial_rotary_factor", 0.5)
+        super().__init__(**kwargs)
+
+
+class GlmAsrConfig(PreTrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`GlmAsrForConditionalGeneration`]. It is used to instantiate an
+    glmasr model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the glmasr-Mini-3B.
+
+    e.g. [zai-org/GLM-ASR-Nano-2512](https://huggingface.co/zai-org/GLM-ASR-Nano-2512)
+
+    Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PreTrainedConfig`] for more information.
+
+    Args:
+        audio_config (`Union[AutoConfig, dict]`, *optional*):
+            The config object or dictionary of the audio encoder.
+        text_config (`Union[AutoConfig, dict]`, *optional*):
+            The config object or dictionary of the text model.
+        audio_token_id (`int`, *optional*, defaults to 59260):
+            The audio token index to encode the audio prompt.
+        projector_hidden_act (`str`, *optional*, defaults to `"gelu"`):
+            The activation function (function or string) in the multi-modal projector.
+
+    ```python
+    >>> from transformers import GlmAsrForConditionalGeneration, GlmAsrConfig
+
+    >>> # Initializing a glmasr configuration
+    >>> configuration = GlmAsrConfig()
+
+    >>> # Initializing a GLM-ASR-Nano-2512 model with random weights
+    >>> model = GlmAsrForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "glmasr"
+    sub_configs = {"text_config": AutoConfig, "audio_config": AutoConfig}
+
+    _default_text_config_kwargs = {
+        "vocab_size": 59264,
+        "hidden_size": 2048,
+        "intermediate_size": 6144,
+        "num_hidden_layers": 28,
+        "num_attention_heads": 16,
+        "num_key_value_heads": 4,
+        "max_position_embeddings": 8192,
+        "rms_norm_eps": 1e-05,
+        "use_cache": True,
+        "eos_token_id": [59246, 59253, 59255],
+        "rope_parameters": {"rope_theta": 10000.0, "rope_type": "default"},
+    }
+
+    def __init__(
+        self,
+        audio_config=None,
+        text_config=None,
+        audio_token_id=59260,
+        projector_hidden_act="gelu",
+        **kwargs,
+    ):
+        if isinstance(audio_config, dict):
+            audio_config["model_type"] = audio_config.get("model_type", "glmasr_encoder")
+            audio_config = CONFIG_MAPPING[audio_config["model_type"]](**audio_config)
+        elif audio_config is None:
+            audio_config = CONFIG_MAPPING["glmasr_encoder"]()
+        self.audio_config = audio_config
+
+        if isinstance(text_config, dict):
+            text_config["model_type"] = text_config.get("model_type", "llama")
+            text_config = CONFIG_MAPPING[text_config["model_type"]](
+                **{**self._default_text_config_kwargs, **text_config}
+            )
+        elif text_config is None:
+            text_config = CONFIG_MAPPING["llama"](**self._default_text_config_kwargs)
+        self.text_config = text_config
+
+        self.vocab_size = text_config.vocab_size
+        self.hidden_size = text_config.hidden_size
+        self.audio_token_id = audio_token_id
+        self.projector_hidden_act = projector_hidden_act
+
+        super().__init__(**kwargs)
+
+
+__all__ = ["GlmAsrEncoderConfig", "GlmAsrConfig"]