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

transformers/models/__init__.py

@@ -23,6 +23,7 @@ if TYPE_CHECKING:
     from .albert import *
     from .align import *
     from .altclip import *
+    from .apertus import *
     from .arcee import *
     from .aria import *
     from .audio_spectrogram_transformer import *
@@ -107,6 +108,7 @@ if TYPE_CHECKING:
     from .dinov3_vit import *
     from .distilbert import *
     from .dit import *
+    from .doge import *
     from .donut import *
     from .dots1 import *
     from .dpr import *
@@ -119,7 +121,11 @@ if TYPE_CHECKING:
     from .emu3 import *
     from .encodec import *
     from .encoder_decoder import *
+    from .eomt import *
     from .ernie import *
+    from .ernie4_5 import *
+    from .ernie4_5_moe import *
+    from .ernie4_5_vl_moe import *
     from .esm import *
     from .evolla import *
     from .exaone4 import *
@@ -144,9 +150,11 @@ if TYPE_CHECKING:
     from .git import *
     from .glm import *
     from .glm4 import *
+    from .glm4_moe import *
     from .glm4v import *
     from .glm4v_moe import *
     from .glm46v import *
+    from .glmasr import *
     from .glpn import *
     from .got_ocr2 import *
     from .gpt2 import *
@@ -181,10 +189,12 @@ if TYPE_CHECKING:
     from .instructblip import *
     from .instructblipvideo import *
     from .internvl import *
+    from .jais2 import *
     from .jamba import *
     from .janus import *
     from .jetmoe import *
     from .kosmos2 import *
+    from .kosmos2_5 import *
     from .kyutai_speech_to_text import *
     from .lasr import *
     from .layoutlm import *
@@ -220,6 +230,7 @@ if TYPE_CHECKING:
     from .mbart50 import *
     from .megatron_bert import *
     from .megatron_gpt2 import *
+    from .metaclip_2 import *
     from .mgp_str import *
     from .mimi import *
     from .minimax import *
@@ -231,6 +242,7 @@ if TYPE_CHECKING:
     from .mlcd import *
     from .mllama import *
     from .mluke import *
+    from .mm_grounding_dino import *
     from .mobilebert import *
     from .mobilenet_v1 import *
     from .mobilenet_v2 import *
@@ -270,6 +282,9 @@ if TYPE_CHECKING:
     from .parakeet import *
     from .patchtsmixer import *
     from .patchtst import *
+    from .pe_audio import *
+    from .pe_audio_video import *
+    from .pe_video import *
     from .pegasus import *
     from .pegasus_x import *
     from .perceiver import *
@@ -281,6 +296,7 @@ if TYPE_CHECKING:
     from .phimoe import *
     from .phobert import *
     from .pix2struct import *
+    from .pixio import *
     from .pixtral import *
     from .plbart import *
     from .poolformer import *
@@ -317,8 +333,10 @@ if TYPE_CHECKING:
     from .sam import *
     from .sam2 import *
     from .sam2_video import *
+    from .sam3 import *
     from .sam3_tracker import *
     from .sam3_tracker_video import *
+    from .sam3_video import *
     from .sam_hq import *
     from .seamless_m4t import *
     from .seamless_m4t_v2 import *
@@ -330,6 +348,7 @@ if TYPE_CHECKING:
     from .shieldgemma2 import *
     from .siglip import *
     from .siglip2 import *
+    from .smollm3 import *
     from .smolvlm import *
     from .speech_encoder_decoder import *
     from .speech_to_text import *

transformers/models/afmoe/modeling_afmoe.py

@@ -25,11 +25,11 @@ from typing import Optional, Union
 import torch
 from torch import nn
 
+from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_kernel_func_from_hub, use_kernelized_func
-from ...integrations.hub_kernels import use_kernel_forward_from_hub
+from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub, use_kernelized_func
 from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
 from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, MoeModelOutputWithPast
@@ -58,7 +58,7 @@ class AfmoeRotaryEmbedding(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(
@@ -531,20 +531,11 @@ class AfmoePreTrainedModel(PreTrainedModel):
 
     def _init_weights(self, module):
         """Initialize the weights"""
-        if isinstance(module, nn.Linear):
-            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
-            if module.bias is not None:
-                nn.init.zeros_(module.bias)
-        elif isinstance(module, nn.Embedding):
-            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
-            if module.padding_idx is not None:
-                nn.init.zeros_(module.weight[module.padding_idx])
-        elif isinstance(module, AfmoeRMSNorm):
-            nn.init.ones_(module.weight)
-        elif isinstance(module, AfmoeTokenChoiceRouter):
-            nn.init.zeros_(module.gate.weight)
+        super()._init_weights(module)
+        if isinstance(module, AfmoeTokenChoiceRouter):
+            init.zeros_(module.gate.weight)
         elif isinstance(module, AfmoeMoE):
-            nn.init.zeros_(module.expert_bias)
+            init.zeros_(module.expert_bias)
 
 
 @auto_docstring

transformers/models/afmoe/modular_afmoe.py

@@ -20,6 +20,7 @@ from typing import Optional
 import torch
 from torch import nn
 
+from ... import initialization as init
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
 from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
@@ -350,20 +351,11 @@ class AfmoePreTrainedModel(PreTrainedModel):
 
     def _init_weights(self, module):
         """Initialize the weights"""
-        if isinstance(module, nn.Linear):
-            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
-            if module.bias is not None:
-                nn.init.zeros_(module.bias)
-        elif isinstance(module, nn.Embedding):
-            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
-            if module.padding_idx is not None:
-                nn.init.zeros_(module.weight[module.padding_idx])
-        elif isinstance(module, AfmoeRMSNorm):
-            nn.init.ones_(module.weight)
-        elif isinstance(module, AfmoeTokenChoiceRouter):
-            nn.init.zeros_(module.gate.weight)
+        super()._init_weights(module)
+        if isinstance(module, AfmoeTokenChoiceRouter):
+            init.zeros_(module.gate.weight)
         elif isinstance(module, AfmoeMoE):
-            nn.init.zeros_(module.expert_bias)
+            init.zeros_(module.expert_bias)
 
 
 @auto_docstring

transformers/models/aimv2/modeling_aimv2.py

@@ -414,6 +414,10 @@ class Aimv2PreTrainedModel(PreTrainedModel):
                 init.constant_(module.logit_scale, math.log(1 / 0.07))
         elif isinstance(module, Aimv2AttentionPoolingHead):
             init.normal_(module.cls_token, mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, Aimv2VisionEmbeddings):
+            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
+        elif isinstance(module, Aimv2TextEmbeddings):
+            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
 
 
 @auto_docstring(

transformers/models/aimv2/modular_aimv2.py

@@ -457,6 +457,10 @@ class Aimv2PreTrainedModel(PreTrainedModel):
                 init.constant_(module.logit_scale, math.log(1 / 0.07))
         elif isinstance(module, Aimv2AttentionPoolingHead):
             init.normal_(module.cls_token, mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, Aimv2VisionEmbeddings):
+            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
+        elif isinstance(module, Aimv2TextEmbeddings):
+            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
 
 
 @auto_docstring(

transformers/models/albert/modeling_albert.py

@@ -320,6 +320,9 @@ class AlbertPreTrainedModel(PreTrainedModel):
             init.ones_(module.weight)
         elif isinstance(module, AlbertMLMHead):
             init.zeros_(module.bias)
+        elif isinstance(module, AlbertEmbeddings):
+            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
+            init.zeros_(module.token_type_ids)
 
 
 @dataclass

transformers/models/align/modeling_align.py

@@ -781,9 +781,9 @@ class AlignTextEncoder(nn.Module):
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             layer_outputs = layer_module(
-                hidden_states=hidden_states,
-                attention_mask=attention_mask,
-                output_attentions=output_attentions,
+                hidden_states,
+                attention_mask,
+                output_attentions,
                 **kwargs,
             )
 
@@ -844,6 +844,13 @@ class AlignPreTrainedModel(PreTrainedModel):
         if isinstance(module, (nn.LayerNorm, nn.BatchNorm2d)):
             init.zeros_(module.bias)
             init.ones_(module.weight)
+            if getattr(module, "running_mean", None) is not None:
+                init.zeros_(module.running_mean)
+                init.ones_(module.running_var)
+                init.zeros_(module.num_batches_tracked)
+        elif isinstance(module, AlignTextEmbeddings):
+            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
+            init.zeros_(module.token_type_ids)
 
 
 @auto_docstring(
@@ -976,6 +983,8 @@ class AlignVisionModel(AlignPreTrainedModel):
     main_input_name = "pixel_values"
     input_modalities = ("image",)
     supports_gradient_checkpointing = False
+    _input_embed_layer = "convolution"
+    _no_split_modules = ["AlignVisionBlock"]
 
     def __init__(self, config: AlignVisionConfig):
         super().__init__(config)
@@ -994,9 +1003,6 @@ class AlignVisionModel(AlignPreTrainedModel):
         # Initialize weights and apply final processing
         self.post_init()
 
-    def get_input_embeddings(self) -> nn.Module:
-        return self.vision_model.embeddings.convolution
-
     @can_return_tuple
     @auto_docstring
     def forward(

transformers/models/altclip/modeling_altclip.py

@@ -393,9 +393,9 @@ class AltRobertaEncoder(nn.Module):
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             layer_outputs = layer_module(
-                hidden_states=hidden_states,
-                attention_mask=attention_mask,
-                output_attentions=output_attentions,
+                hidden_states,
+                attention_mask,
+                output_attentions,
                 **kwargs,
             )
 
@@ -780,6 +780,7 @@ class AltCLIPPreTrainedModel(PreTrainedModel):
             init.normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor)
             init.normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor)
             init.normal_(module.position_embedding.weight, std=module.config.initializer_range * factor)
+            init.copy_(module.position_ids, torch.arange(module.num_positions).expand((1, -1)))
         elif isinstance(module, AltCLIPAttention):
             factor = self.config.initializer_factor
             in_proj_std = (module.embed_dim**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
@@ -815,6 +816,9 @@ class AltCLIPPreTrainedModel(PreTrainedModel):
             # Here we need the check explicitly, as we slice the weight in the `zeros_` call, so it looses the flag
             if module.padding_idx is not None and not getattr(module.weight, "_is_hf_initialized", False):
                 init.zeros_(module.weight[module.padding_idx])
+        elif isinstance(module, AltRobertaEmbeddings):
+            init.copy_(module.position_ids, torch.arange(module.position_ids.shape[-1]).expand((1, -1)))
+            init.zeros_(module.token_type_ids)
 
 
 class AltCLIPVisionTransformer(nn.Module):

transformers/models/apertus/modeling_apertus.py

@@ -25,7 +25,7 @@ from typing import Optional, Union
 import torch
 from torch import nn
 
-from ...activations import ACT2FN
+from ...activations import ACT2CLS, ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
 from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub, use_kernelized_func
@@ -49,6 +49,8 @@ class ApertusMLP(nn.Module):
         self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
         self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
         self.act_fn = ACT2FN[config.hidden_act]
+        if config.hidden_act == "xielu":
+            self.act_fn = ACT2CLS["xielu"](dtype=config.dtype)
 
     def forward(self, x):
         return self.down_proj(self.act_fn(self.up_proj(x)))
@@ -92,7 +94,7 @@ class ApertusRotaryEmbedding(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/apertus/modular_apertus.py

@@ -19,6 +19,7 @@ from typing import Optional
 import torch
 from torch import nn
 
+from ...activations import ACT2CLS
 from ...cache_utils import Cache
 from ...configuration_utils import PreTrainedConfig
 from ...modeling_rope_utils import RopeParameters
@@ -192,9 +193,11 @@ class ApertusConfig(PreTrainedConfig):
 
 class ApertusMLP(NemotronMLP):
     def __init__(self, config):
-        super().__init__()
+        super().__init__(config)
         self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
         self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        if config.hidden_act == "xielu":
+            self.act_fn = ACT2CLS["xielu"](dtype=config.dtype)
 
 
 class ApertusRMSNorm(LlamaRMSNorm):

transformers/models/arcee/modeling_arcee.py

@@ -99,7 +99,7 @@ class ArceeRotaryEmbedding(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/aria/modeling_aria.py

@@ -636,7 +636,7 @@ class AriaTextRotaryEmbedding(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(
@@ -1203,6 +1203,7 @@ class AriaForConditionalGeneration(AriaPreTrainedModel, GenerationMixin):
         attention_mask=None,
         cache_position=None,
         logits_to_keep=None,
+        is_first_iteration=False,
         **kwargs,
     ):
         model_inputs = super().prepare_inputs_for_generation(
@@ -1212,12 +1213,15 @@ class AriaForConditionalGeneration(AriaPreTrainedModel, GenerationMixin):
             attention_mask=attention_mask,
             cache_position=cache_position,
             logits_to_keep=logits_to_keep,
+            is_first_iteration=is_first_iteration,
             **kwargs,
         )
 
-        if cache_position[0] == 0:
-            # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
-            # Otherwise we need pixel values to be passed to model
+        if is_first_iteration or not kwargs.get("use_cache", True):
+            # Pixel values are used only in the first iteration if available
+            # In subsquent iterations, they are already merged with text and cached
+            # NOTE: first iteration doesn't have to be prefill, it can be the first
+            # iteration with a question and cached system prompt (continue generate from cache)
             model_inputs["pixel_values"] = pixel_values
             model_inputs["pixel_mask"] = pixel_mask
 

transformers/models/aria/modular_aria.py

@@ -1500,6 +1500,7 @@ class AriaForConditionalGeneration(LlavaForConditionalGeneration):
         attention_mask=None,
         cache_position=None,
         logits_to_keep=None,
+        is_first_iteration=False,
         **kwargs,
     ):
         model_inputs = super().prepare_inputs_for_generation(
@@ -1509,12 +1510,15 @@ class AriaForConditionalGeneration(LlavaForConditionalGeneration):
             attention_mask=attention_mask,
             cache_position=cache_position,
             logits_to_keep=logits_to_keep,
+            is_first_iteration=is_first_iteration,
             **kwargs,
         )
 
-        if cache_position[0] == 0:
-            # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
-            # Otherwise we need pixel values to be passed to model
+        if is_first_iteration or not kwargs.get("use_cache", True):
+            # Pixel values are used only in the first iteration if available
+            # In subsquent iterations, they are already merged with text and cached
+            # NOTE: first iteration doesn't have to be prefill, it can be the first
+            # iteration with a question and cached system prompt (continue generate from cache)
             model_inputs["pixel_values"] = pixel_values
             model_inputs["pixel_mask"] = pixel_mask
 

transformers/models/audioflamingo3/processing_audioflamingo3.py

@@ -32,9 +32,6 @@ if is_torch_available():
 
 logger = logging.get_logger(__name__)
 
-MAX_AUDIO_LEN = 10 * 60  # 10 minutes
-DEFAULT_TRANSCRIPTION_PROMPT = "Transcribe the input speech."
-
 
 class AudioFlamingo3ProcessorKwargs(ProcessingKwargs, total=False):
     _defaults = {
@@ -63,32 +60,41 @@ class AudioFlamingo3Processor(ProcessorMixin):
     [`Qwen2TokenizerFast`]. See the [`~AudioFlamingo3Processor.__call__`] for more information.
 
     Args:
-        feature_extractor ([`WhisperFeatureExtractor`]):
-            The feature extractor is a required input.
-        tokenizer ([`Qwen2TokenizerFast`]):
-            The tokenizer is a required input.
-        chat_template (`Optional[str]`, *optional*):
-            The Jinja template to use for formatting the conversation. If not provided, the tokenizer's default chat
-            template will be used.
-        audio_token (`Optional[str]`, *optional*, defaults to `"<sound>"`):
-            Special token used to represent audio inputs in the chat template.
+            feature_extractor ([`WhisperFeatureExtractor`]):
+                The feature extractor is a required input.
+            tokenizer ([`Qwen2TokenizerFast`]):
+                The tokenizer is a required input.
+            chat_template (`Optional[str]`, *optional*):
+                The Jinja template to use for formatting the conversation. If not provided, the tokenizer's default chat
+                template will be used.
+            audio_token (`Optional[str]`, *optional*, defaults to `"<sound>"`):
+                Special token used to represent audio inputs in the chat template.
+            default_transcription_prompt (`str`, *optional*, defaults to `"Transcribe the input speech."`):
+                Default prompt to use for transcription tasks when applying transcription requests.
+            max_audio_len (`int`, *optional*, defaults to 600):
+                Maximum length of audio sequences in seconds. Audio longer than this will be truncated.
     """
 
-    attributes = ["feature_extractor", "tokenizer"]
-    feature_extractor_class = "WhisperFeatureExtractor"
-    tokenizer_class = "Qwen2TokenizerFast"
-
     def __init__(
         self,
         feature_extractor,
         tokenizer,
         chat_template=None,
         audio_token="<sound>",
+        default_transcription_prompt="Transcribe the input speech.",
+        max_audio_len=600,
     ):
         self.audio_token = audio_token
         self.audio_token_id = tokenizer.convert_tokens_to_ids(audio_token)
+        self.default_transcription_prompt = default_transcription_prompt
+        self.max_audio_len = max_audio_len
         super().__init__(feature_extractor, tokenizer, chat_template=chat_template)
 
+    def _get_audio_token_length(self, audio_lengths: "torch.Tensor") -> "torch.Tensor":
+        conv_output_lengths = (audio_lengths - 1) // 2 + 1  # After conv2 downsampling
+        audio_tokens_lengths = (conv_output_lengths - 2) // 2 + 1  # After avg pooling
+        return audio_tokens_lengths
+
     def __call__(
         self,
         text: Union[TextInput, list[TextInput]],
@@ -143,7 +149,7 @@ class AudioFlamingo3Processor(ProcessorMixin):
 
             # Determine number of chunks per sample, and flatten
             window_size = int(audio_kwargs["sampling_rate"] * audio_kwargs["chunk_length"])
-            max_windows = int(MAX_AUDIO_LEN // audio_kwargs["chunk_length"])
+            max_windows = int(self.max_audio_len // audio_kwargs["chunk_length"])
 
             per_sample_windows: list[int] = []
             flat_chunks: list[np.ndarray] = []
@@ -153,7 +159,7 @@ class AudioFlamingo3Processor(ProcessorMixin):
                 n_win = max(1, (n_samples + window_size - 1) // window_size)
                 if n_win > max_windows:
                     logger.warning(
-                        f"Audio duration ({n_samples / audio_kwargs['sampling_rate']:.1f}s) exceeds {MAX_AUDIO_LEN}s; truncating to first {MAX_AUDIO_LEN}s."
+                        f"Audio duration ({n_samples / audio_kwargs['sampling_rate']:.1f}s) exceeds {self.max_audio_len}s; truncating to first {self.max_audio_len}s."
                     )
                     n_win = max_windows
                 per_sample_windows.append(n_win)
@@ -171,8 +177,7 @@ class AudioFlamingo3Processor(ProcessorMixin):
 
             # Compute sequence lengths token counting
             audio_lengths = torch.stack([s.sum() for s in torch.split(padding_mask.sum(-1), per_sample_windows)])
-            conv_output_lengths = (audio_lengths - 1) // 2 + 1  # After conv2 downsampling
-            audio_tokens_lengths = (conv_output_lengths - 2) // 2 + 1  # After avg pooling
+            audio_tokens_lengths = self._get_audio_token_length(audio_lengths)
 
             # expand audio tokens in text
             for i, audio_length in enumerate(audio_tokens_lengths):
@@ -236,7 +241,7 @@ class AudioFlamingo3Processor(ProcessorMixin):
             raise ValueError("`audio` must contain at least one sample.")
 
         if prompt is None:
-            prompts = [DEFAULT_TRANSCRIPTION_PROMPT] * batch_size
+            prompts = [self.default_transcription_prompt] * batch_size
         elif isinstance(prompt, str):
             prompts = [prompt] * batch_size
         elif isinstance(prompt, (list, tuple)):
@@ -247,7 +252,7 @@ class AudioFlamingo3Processor(ProcessorMixin):
             prompts = []
             for item in prompt:
                 if item is None:
-                    prompts.append(DEFAULT_TRANSCRIPTION_PROMPT)
+                    prompts.append(self.default_transcription_prompt)
                 elif isinstance(item, str):
                     prompts.append(item)
                 else:

transformers/models/auto/auto_factory.py

@@ -543,7 +543,7 @@ def add_generation_mixin_to_remote_model(model_class):
 
 class _LazyAutoMapping(OrderedDict[type[PreTrainedConfig], _LazyAutoMappingValue]):
     """
-    " A mapping config to object (model or tokenizer for instance) that will load keys and values when it is accessed.
+    A mapping config to object (model or tokenizer for instance) that will load keys and values when it is accessed.
 
     Args:
         - config_mapping: The map model type to config class

transformers/models/auto/configuration_auto.py

@@ -142,6 +142,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("ernie", "ErnieConfig"),
         ("ernie4_5", "Ernie4_5Config"),
         ("ernie4_5_moe", "Ernie4_5_MoeConfig"),
+        ("ernie4_5_vl_moe", "Ernie4_5_VL_MoeConfig"),
         ("esm", "EsmConfig"),
         ("evolla", "EvollaConfig"),
         ("exaone4", "Exaone4Config"),
@@ -179,6 +180,8 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("glm4v_moe_vision", "Glm4vMoeVisionConfig"),
         ("glm4v_text", "Glm4vTextConfig"),
         ("glm4v_vision", "Glm4vVisionConfig"),
+        ("glmasr", "GlmAsrConfig"),
+        ("glmasr_encoder", "GlmAsrEncoderConfig"),
         ("glpn", "GLPNConfig"),
         ("got_ocr2", "GotOcr2Config"),
         ("gpt-sw3", "GPT2Config"),
@@ -215,6 +218,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("instructblipvideo", "InstructBlipVideoConfig"),
         ("internvl", "InternVLConfig"),
         ("internvl_vision", "InternVLVisionConfig"),
+        ("jais2", "Jais2Config"),
         ("jamba", "JambaConfig"),
         ("janus", "JanusConfig"),
         ("jetmoe", "JetMoeConfig"),
@@ -306,6 +310,12 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("parakeet_encoder", "ParakeetEncoderConfig"),
         ("patchtsmixer", "PatchTSMixerConfig"),
         ("patchtst", "PatchTSTConfig"),
+        ("pe_audio", "PeAudioConfig"),
+        ("pe_audio_encoder", "PeAudioEncoderConfig"),
+        ("pe_audio_video", "PeAudioVideoConfig"),
+        ("pe_audio_video_encoder", "PeAudioVideoEncoderConfig"),
+        ("pe_video", "PeVideoConfig"),
+        ("pe_video_encoder", "PeVideoEncoderConfig"),
         ("pegasus", "PegasusConfig"),
         ("pegasus_x", "PegasusXConfig"),
         ("perceiver", "PerceiverConfig"),
@@ -316,6 +326,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("phi4_multimodal", "Phi4MultimodalConfig"),
         ("phimoe", "PhimoeConfig"),
         ("pix2struct", "Pix2StructConfig"),
+        ("pixio", "PixioConfig"),
         ("pixtral", "PixtralVisionConfig"),
         ("plbart", "PLBartConfig"),
         ("poolformer", "PoolFormerConfig"),
@@ -582,6 +593,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("ernie", "ERNIE"),
         ("ernie4_5", "Ernie4_5"),
         ("ernie4_5_moe", "Ernie4_5_MoE"),
+        ("ernie4_5_vl_moe", "Ernie4_5_VL_MoE"),
         ("esm", "ESM"),
         ("evolla", "Evolla"),
         ("exaone4", "EXAONE-4.0"),
@@ -622,6 +634,8 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("glm4v_moe_vision", "Glm4vMoeVisionModel"),
         ("glm4v_text", "GLM4V"),
         ("glm4v_vision", "Glm4vVisionModel"),
+        ("glmasr", "GLM-ASR"),
+        ("glmasr_encoder", "GLM-ASR Encoder"),
         ("glpn", "GLPN"),
         ("got_ocr2", "GOT-OCR2"),
         ("gpt-sw3", "GPT-Sw3"),
@@ -659,6 +673,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("instructblipvideo", "InstructBlipVideo"),
         ("internvl", "InternVL"),
         ("internvl_vision", "InternVLVision"),
+        ("jais2", "Jais2"),
         ("jamba", "Jamba"),
         ("janus", "Janus"),
         ("jetmoe", "JetMoe"),
@@ -762,6 +777,12 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("parakeet_encoder", "ParakeetEncoder"),
         ("patchtsmixer", "PatchTSMixer"),
         ("patchtst", "PatchTST"),
+        ("pe_audio", "PeAudio"),
+        ("pe_audio_encoder", "PeAudioEncoder"),
+        ("pe_audio_video", "PeAudioVideo"),
+        ("pe_audio_video_encoder", "PeAudioVideoEncoder"),
+        ("pe_video", "PeVideo"),
+        ("pe_video_encoder", "PeVideoEncoder"),
         ("pegasus", "Pegasus"),
         ("pegasus_x", "PEGASUS-X"),
         ("perceiver", "Perceiver"),
@@ -773,6 +794,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("phimoe", "Phimoe"),
         ("phobert", "PhoBERT"),
         ("pix2struct", "Pix2Struct"),
+        ("pixio", "Pixio"),
         ("pixtral", "Pixtral"),
         ("plbart", "PLBart"),
         ("poolformer", "PoolFormer"),
@@ -955,6 +977,7 @@ SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict[str, str](
         ("glm4v_moe_vision", "glm4v_moe"),
         ("glm4v_text", "glm4v"),
         ("glm4v_moe_text", "glm4v_moe"),
+        ("glmasr_encoder", "glmasr"),
         ("grounding-dino", "grounding_dino"),
         ("mm-grounding-dino", "mm_grounding_dino"),
         ("idefics3_vision", "idefics3"),
@@ -981,6 +1004,10 @@ SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict[str, str](
         ("llama4_text", "llama4"),
         ("blip_2_qformer", "blip_2"),
         ("fastspeech2_conformer_with_hifigan", "fastspeech2_conformer"),
+        ("perception_encoder", "perception_lm"),
+        ("pe_audio_encoder", "pe_audio"),
+        ("pe_video_encoder", "pe_video"),
+        ("pe_audio_video_encoder", "pe_audio_video"),
         ("video_llama_3_vision", "video_llama_3"),
         ("parakeet_encoder", "parakeet"),
         ("parakeet_ctc", "parakeet"),