transformers 5.0.0rc0__py3-none-any.whl → 5.0.0rc1__py3-none-any.whl

transformers/models/apertus/modeling_apertus.py

@@ -28,7 +28,7 @@ from torch import nn
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_kernel_forward_from_hub, use_kernel_func_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
 from ...modeling_layers import GenericForTokenClassification, GradientCheckpointingLayer
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
@@ -36,7 +36,7 @@ from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
-from ...utils.generic import check_model_inputs
+from ...utils.generic import check_model_inputs, maybe_autocast
 from .configuration_apertus import ApertusConfig
 
 
@@ -131,7 +131,7 @@ class ApertusRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
@@ -213,6 +213,7 @@ def eager_attention_forward(
     return attn_output, attn_weights
 
 
+@use_kernelized_func(apply_rotary_pos_emb)
 class ApertusAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -238,7 +239,6 @@ class ApertusAttention(nn.Module):
         self.o_proj = nn.Linear(
             config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
         )
-        self.rotary_fn = apply_rotary_pos_emb
         self.q_norm = ApertusRMSNorm(self.head_dim, config.rms_norm_eps)
         self.k_norm = ApertusRMSNorm(self.head_dim, config.rms_norm_eps)
 

transformers/models/arcee/modeling_arcee.py

@@ -30,7 +30,7 @@ from transformers.utils import auto_docstring
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_kernel_forward_from_hub, use_kernel_func_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
 from ...modeling_layers import (
     GenericForQuestionAnswering,
@@ -43,7 +43,7 @@ from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, can_return_tuple
-from ...utils.generic import check_model_inputs
+from ...utils.generic import check_model_inputs, maybe_autocast
 from .configuration_arcee import ArceeConfig
 
 
@@ -138,7 +138,7 @@ class ArceeRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
@@ -220,6 +220,7 @@ def eager_attention_forward(
     return attn_output, attn_weights
 
 
+@use_kernelized_func(apply_rotary_pos_emb)
 class ArceeAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -245,7 +246,6 @@ class ArceeAttention(nn.Module):
         self.o_proj = nn.Linear(
             config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
         )
-        self.rotary_fn = apply_rotary_pos_emb
 
     def forward(
         self,

transformers/models/aria/modeling_aria.py

@@ -29,7 +29,7 @@ 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_kernel_func_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
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
@@ -38,7 +38,7 @@ from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
-from ...utils.generic import check_model_inputs
+from ...utils.generic import check_model_inputs, maybe_autocast
 from ..auto import AutoModel
 from .configuration_aria import AriaConfig, AriaTextConfig
 
@@ -444,6 +444,7 @@ def eager_attention_forward(
     return attn_output, attn_weights
 
 
+@use_kernelized_func(apply_rotary_pos_emb)
 class AriaTextAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -469,7 +470,6 @@ class AriaTextAttention(nn.Module):
         self.o_proj = nn.Linear(
             config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
         )
-        self.rotary_fn = apply_rotary_pos_emb
 
     def forward(
         self,
@@ -675,7 +675,7 @@ class AriaTextRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling

transformers/models/audioflamingo3/modeling_audioflamingo3.py

@@ -323,6 +323,7 @@ class AudioFlamingo3Encoder(AudioFlamingo3PreTrainedModel):
         self,
         input_features: torch.Tensor,
         input_features_mask: Optional[torch.Tensor] = None,
+        **kwargs,
     ):
         r"""
         Args:

transformers/models/audioflamingo3/modular_audioflamingo3.py

@@ -60,6 +60,7 @@ class AudioFlamingo3Encoder(Qwen2AudioEncoder):
         self,
         input_features: torch.Tensor,
         input_features_mask: Optional[torch.Tensor] = None,
+        **kwargs,
     ):
         r"""
         Args:

transformers/models/auto/configuration_auto.py

@@ -148,6 +148,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("falcon", "FalconConfig"),
         ("falcon_h1", "FalconH1Config"),
         ("falcon_mamba", "FalconMambaConfig"),
+        ("fast_vlm", "FastVlmConfig"),
         ("fastspeech2_conformer", "FastSpeech2ConformerConfig"),
         ("fastspeech2_conformer_with_hifigan", "FastSpeech2ConformerWithHifiGanConfig"),
         ("flaubert", "FlaubertConfig"),
@@ -220,6 +221,8 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("kosmos-2", "Kosmos2Config"),
         ("kosmos-2.5", "Kosmos2_5Config"),
         ("kyutai_speech_to_text", "KyutaiSpeechToTextConfig"),
+        ("lasr_ctc", "LasrCTCConfig"),
+        ("lasr_encoder", "LasrEncoderConfig"),
         ("layoutlm", "LayoutLMConfig"),
         ("layoutlmv2", "LayoutLMv2Config"),
         ("layoutlmv3", "LayoutLMv3Config"),
@@ -297,6 +300,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("ovis2", "Ovis2Config"),
         ("owlv2", "Owlv2Config"),
         ("owlvit", "OwlViTConfig"),
+        ("paddleocr_vl", "PaddleOCRVLConfig"),
         ("paligemma", "PaliGemmaConfig"),
         ("parakeet_ctc", "ParakeetCTCConfig"),
         ("parakeet_encoder", "ParakeetEncoderConfig"),
@@ -585,6 +589,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("falcon3", "Falcon3"),
         ("falcon_h1", "FalconH1"),
         ("falcon_mamba", "FalconMamba"),
+        ("fast_vlm", "FastVlm"),
         ("fastspeech2_conformer", "FastSpeech2Conformer"),
         ("fastspeech2_conformer_with_hifigan", "FastSpeech2ConformerWithHifiGan"),
         ("flan-t5", "FLAN-T5"),
@@ -660,6 +665,9 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("kosmos-2", "KOSMOS-2"),
         ("kosmos-2.5", "KOSMOS-2.5"),
         ("kyutai_speech_to_text", "KyutaiSpeechToText"),
+        ("lasr", "Lasr"),
+        ("lasr_ctc", "Lasr"),
+        ("lasr_encoder", "LasrEncoder"),
         ("layoutlm", "LayoutLM"),
         ("layoutlmv2", "LayoutLMv2"),
         ("layoutlmv3", "LayoutLMv3"),
@@ -747,6 +755,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("ovis2", "Ovis2"),
         ("owlv2", "OWLv2"),
         ("owlvit", "OWL-ViT"),
+        ("paddleocr_vl", "PaddleOCRVL"),
         ("paligemma", "PaliGemma"),
         ("parakeet", "Parakeet"),
         ("parakeet_ctc", "Parakeet"),
@@ -975,6 +984,8 @@ SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict[str, str](
         ("video_llama_3_vision", "video_llama_3"),
         ("parakeet_encoder", "parakeet"),
         ("parakeet_ctc", "parakeet"),
+        ("lasr_encoder", "lasr"),
+        ("lasr_ctc", "lasr"),
         ("wav2vec2-bert", "wav2vec2_bert"),
     ]
 )

transformers/models/auto/feature_extraction_auto.py

@@ -49,6 +49,8 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
         ("granite_speech", "GraniteSpeechFeatureExtractor"),
         ("hubert", "Wav2Vec2FeatureExtractor"),
         ("kyutai_speech_to_text", "KyutaiSpeechToTextFeatureExtractor"),
+        ("lasr_ctc", "LasrFeatureExtractor"),
+        ("lasr_encoder", "LasrFeatureExtractor"),
         ("markuplm", "MarkupLMFeatureExtractor"),
         ("mimi", "EncodecFeatureExtractor"),
         ("moonshine", "Wav2Vec2FeatureExtractor"),

transformers/models/auto/image_processing_auto.py

@@ -153,6 +153,7 @@ else:
             ("ovis2", ("Ovis2ImageProcessor", "Ovis2ImageProcessorFast")),
             ("owlv2", ("Owlv2ImageProcessor", "Owlv2ImageProcessorFast")),
             ("owlvit", ("OwlViTImageProcessor", "OwlViTImageProcessorFast")),
+            ("paddleocr_vl", ("PaddleOCRVLImageProcessor", "PaddleOCRVLImageProcessorFast")),
             ("paligemma", ("SiglipImageProcessor", "SiglipImageProcessorFast")),
             ("perceiver", ("PerceiverImageProcessor", "PerceiverImageProcessorFast")),
             ("perception_lm", (None, "PerceptionLMImageProcessorFast")),

transformers/models/auto/modeling_auto.py

@@ -151,6 +151,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("falcon", "FalconModel"),
         ("falcon_h1", "FalconH1Model"),
         ("falcon_mamba", "FalconMambaModel"),
+        ("fast_vlm", "FastVlmModel"),
         ("fastspeech2_conformer", "FastSpeech2ConformerModel"),
         ("fastspeech2_conformer_with_hifigan", "FastSpeech2ConformerWithHifiGan"),
         ("flaubert", "FlaubertModel"),
@@ -221,6 +222,8 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("kosmos-2", "Kosmos2Model"),
         ("kosmos-2.5", "Kosmos2_5Model"),
         ("kyutai_speech_to_text", "KyutaiSpeechToTextModel"),
+        ("lasr_ctc", "LasrForCTC"),
+        ("lasr_encoder", "LasrEncoder"),
         ("layoutlm", "LayoutLMModel"),
         ("layoutlmv2", "LayoutLMv2Model"),
         ("layoutlmv3", "LayoutLMv3Model"),
@@ -996,6 +999,7 @@ MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES = OrderedDict(
         ("deepseek_vl_hybrid", "DeepseekVLHybridForConditionalGeneration"),
         ("emu3", "Emu3ForConditionalGeneration"),
         ("evolla", "EvollaForProteinText2Text"),
+        ("fast_vlm", "FastVlmForConditionalGeneration"),
         ("florence2", "Florence2ForConditionalGeneration"),
         ("fuyu", "FuyuForCausalLM"),
         ("gemma3", "Gemma3ForConditionalGeneration"),
@@ -1022,6 +1026,7 @@ MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES = OrderedDict(
         ("mistral3", "Mistral3ForConditionalGeneration"),
         ("mllama", "MllamaForConditionalGeneration"),
         ("ovis2", "Ovis2ForConditionalGeneration"),
+        ("paddleocr_vl", "PaddleOCRVLForConditionalGeneration"),
         ("paligemma", "PaliGemmaForConditionalGeneration"),
         ("perception_lm", "PerceptionLMForConditionalGeneration"),
         ("pix2struct", "Pix2StructForConditionalGeneration"),
@@ -1581,6 +1586,7 @@ MODEL_FOR_CTC_MAPPING_NAMES = OrderedDict(
         # Model for Connectionist temporal classification (CTC) mapping
         ("data2vec-audio", "Data2VecAudioForCTC"),
         ("hubert", "HubertForCTC"),
+        ("lasr_ctc", "LasrForCTC"),
         ("parakeet_ctc", "ParakeetForCTC"),
         ("sew", "SEWForCTC"),
         ("sew-d", "SEWDForCTC"),

transformers/models/auto/processing_auto.py

@@ -114,6 +114,7 @@ PROCESSOR_MAPPING_NAMES = OrderedDict(
         ("ovis2", "Ovis2Processor"),
         ("owlv2", "Owlv2Processor"),
         ("owlvit", "OwlViTProcessor"),
+        ("paddleocr_vl", "PaddleOCRVLProcessor"),
         ("paligemma", "PaliGemmaProcessor"),
         ("perception_lm", "PerceptionLMProcessor"),
         ("phi4_multimodal", "Phi4MultimodalProcessor"),
@@ -317,7 +318,6 @@ class AutoProcessor:
                     processor_class = config_dict.get("processor_class", None)
                     if "AutoProcessor" in config_dict.get("auto_map", {}):
                         processor_auto_map = config_dict["auto_map"]["AutoProcessor"]
-
             # Saved as feature extractor
             if preprocessor_config_file is None:
                 preprocessor_config_file = cached_file(
@@ -345,16 +345,24 @@ class AutoProcessor:
                     processor_auto_map = config_dict["auto_map"]["AutoProcessor"]
 
         if processor_class is None:
-            # Otherwise, load config, if it can be loaded.
-            if not isinstance(config, PreTrainedConfig):
-                config = AutoConfig.from_pretrained(
-                    pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
-                )
+            # Last resort: try loading the model config to get processor_class.
+            # This handles cases where processor info is only in config.json (not in any
+            # preprocessor/tokenizer config files). AutoConfig.from_pretrained may raise
+            # ValueError if the model_type is unrecognized or the config is invalid -
+            # we catch and ignore this to allow fallback to AutoTokenizer/AutoImageProcessor.
+            try:
+                if not isinstance(config, PreTrainedConfig):
+                    config = AutoConfig.from_pretrained(
+                        pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
+                    )
 
-            # And check if the config contains the processor class.
-            processor_class = getattr(config, "processor_class", None)
-            if hasattr(config, "auto_map") and "AutoProcessor" in config.auto_map:
-                processor_auto_map = config.auto_map["AutoProcessor"]
+                processor_class = getattr(config, "processor_class", None)
+                if hasattr(config, "auto_map") and "AutoProcessor" in config.auto_map:
+                    processor_auto_map = config.auto_map["AutoProcessor"]
+            except ValueError:
+                # Config loading failed (unrecognized model_type, invalid config, etc.)
+                # Continue to fallback logic below (AutoTokenizer, AutoImageProcessor, etc.)
+                pass
 
         if processor_class is not None:
             processor_class = processor_class_from_name(processor_class)