Upload folder using huggingface_hub

config.json

@@ -0,0 +1,402 @@
+{
+  "architectures": [
+    "NemotronParseForConditionalGeneration"
+  ],
+  "auto_map": {
+    "AutoConfig": "hf_nemotron_parse_config.NemotronParseConfig",
+    "AutoModel": "hf_nemotron_parse_modeling.NemotronParseForConditionalGeneration",
+    "AutoImageProcessor": "hf_nemotron_parse_processor.NemotronParseImageProcessor",
+    "AutoProcessor": "hf_nemotron_parse_processor.NemotronParseProcessor"
+  },
+  "bos_token_id": 0,
+  "decoder": {
+    "_attn_implementation": "sdpa",
+    "_name_or_path": "",
+    "activation_dropout": 0.0,
+    "activation_function": "gelu",
+    "add_cross_attention": true,
+    "add_final_layer_norm": true,
+    "architectures": null,
+    "attention_dropout": 0.0,
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": 0,
+    "chunk_size_feed_forward": 0,
+    "classifier_dropout": 0.0,
+    "cross_attention_hidden_size": null,
+    "d_model": 1024,
+    "decoder_attention_heads": 16,
+    "decoder_ffn_dim": 4096,
+    "decoder_layerdrop": 0.0,
+    "decoder_layers": 10,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "dropout": 0.1,
+    "early_stopping": false,
+    "encoder_attention_heads": 16,
+    "encoder_ffn_dim": 4096,
+    "encoder_layerdrop": 0.0,
+    "encoder_layers": 12,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 2,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": 2,
+    "hidden_size": 1024,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1",
+      "2": "LABEL_2"
+    },
+    "init_std": 0.02,
+    "is_decoder": true,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1,
+      "LABEL_2": 2
+    },
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "nemotron_parse_text",
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_hidden_layers": 12,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": 1,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "scale_embedding": true,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": false,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "bfloat16",
+    "torchscript": false,
+    "transformers_version": "4.51.3",
+    "typical_p": 1.0,
+    "use_bfloat16": true,
+    "use_cache": true,
+    "vocab_size": 52352
+  },
+  "decoder_start_token_id": 2,
+  "encoder": {
+    "_attn_implementation": "eager",
+    "_name_or_path": "nvidia/C-RADIOv2-H",
+    "adaptor_configs": {},
+    "adaptor_names": null,
+    "add_cross_attention": false,
+    "architectures": [
+      "RADIOModel"
+    ],
+    "args": {
+      "aa": null,
+      "amp": true,
+      "amp_dtype": "bfloat16",
+      "amp_impl": "native",
+      "aug_repeats": 0,
+      "aug_splits": 0,
+      "bn_eps": null,
+      "bn_momentum": null,
+      "cache_dir": null,
+      "channels_last": false,
+      "checkpoint_hist": 10,
+      "chk_keep_forever": 100,
+      "class_map": "",
+      "clip_grad": null,
+      "clip_mode": "norm",
+      "cls_token_per_teacher": true,
+      "coco_annotations_file": "/datasets/coco2017-adlsa/annotations/captions_val2017.json",
+      "coco_image_dir": "/datasets/coco2017-adlsa/val2017",
+      "color_jitter": 0.4,
+      "cooldown_epochs": 0,
+      "cpe_max_size": 2048,
+      "crd_loss": false,
+      "crd_loss_weight": 0.8,
+      "crop_pct": null,
+      "cutmix": 0.0,
+      "cutmix_minmax": null,
+      "dataset_download": false,
+      "debug_full_knn": false,
+      "decay_epochs": 90,
+      "decay_milestones": [
+        90,
+        180,
+        270
+      ],
+      "decay_rate": 0.1,
+      "depchain": true,
+      "dist_bn": "reduce",
+      "dist_norm_weight": 0.0,
+      "distributed": true,
+      "drop": 0.0,
+      "drop_block": null,
+      "drop_connect": null,
+      "drop_path": null,
+      "dtype": "bfloat16",
+      "epoch_repeats": 0.0,
+      "eval": false,
+      "eval_metric": "knn_top1",
+      "eval_teacher": false,
+      "eval_teacher_only": false,
+      "eval_throughput": false,
+      "fast_norm": false,
+      "fd_loss_fn": "MSE",
+      "feature_normalization": "SHIP_NORM",
+      "feature_summarizer": "cls_token",
+      "feature_upscale_factor": null,
+      "force_new_wandb_id": false,
+      "force_spectral_reparam": true,
+      "freeze_bn": false,
+      "fsdp": false,
+      "fuser": "",
+      "gp": null,
+      "grad_accum_steps": 1,
+      "grad_checkpointing": false,
+      "head_init_bias": null,
+      "head_init_scale": null,
+      "head_warmup": 5,
+      "head_weight_decay": 0.001,
+      "hflip": 0.5,
+      "img_size": null,
+      "in_chans": null,
+      "initial_checkpoint": null,
+      "input_size": null,
+      "interpolation": "",
+      "layer_decay": null,
+      "local_rank": 0,
+      "log_interval": 50,
+      "log_mlflow": false,
+      "log_wandb": true,
+      "loss_auto_balance": false,
+      "lr_base": 0.1,
+      "lr_base_scale": "",
+      "lr_base_size": 256,
+      "lr_cycle_decay": 0.5,
+      "lr_cycle_limit": 1,
+      "lr_cycle_mul": 1.0,
+      "lr_k_decay": 1.0,
+      "lr_noise": null,
+      "lr_noise_pct": 0.67,
+      "lr_noise_std": 1.0,
+      "mean": null,
+      "mesa": false,
+      "min_lr": 0,
+      "mixup": 0.0,
+      "mixup_mode": "batch",
+      "mixup_off_epoch": 0,
+      "mixup_prob": 1.0,
+      "mixup_switch_prob": 0.5,
+      "mlp_hidden_size": 1520,
+      "mlp_num_inner": 3,
+      "mlp_version": "v2",
+      "model": "vit_huge_patch16_224",
+      "model_kwargs": {},
+      "model_norm": false,
+      "momentum": 0.9,
+      "no_aug": false,
+      "no_ddp_bb": true,
+      "no_prefetcher": false,
+      "no_resume_opt": false,
+      "num_classes": null,
+      "opt_betas": null,
+      "opt_eps": null,
+      "patience_epochs": 10,
+      "pin_mem": false,
+      "prefetcher": true,
+      "pretrained": false,
+      "rank": 0,
+      "ratio": [
+        0.75,
+        1.3333333333333333
+      ],
+      "recount": 1,
+      "recovery_interval": 0,
+      "register_multiple": 8,
+      "remode": "pixel",
+      "reprob": 0.0,
+      "reset_loss_state": false,
+      "resplit": false,
+      "save_images": false,
+      "scale": [
+        0.5,
+        1.0
+      ],
+      "sched": "cosine",
+      "seed": 42,
+      "smoothing": 0.1,
+      "spectral_heads": false,
+      "spectral_reparam": false,
+      "split_bn": false,
+      "start_epoch": null,
+      "std": null,
+      "stream_teachers": true,
+      "sync_bn": false,
+      "synchronize_step": false,
+      "teachers": [
+        {
+          "fd_normalize": false,
+          "feature_distillation": true,
+          "input_size": 378,
+          "model": "ViT-H-14-378-quickgelu",
+          "name": "clip",
+          "pretrained": "dfn5b",
+          "type": "open_clip",
+          "use_summary": true
+        },
+        {
+          "fd_normalize": false,
+          "feature_distillation": true,
+          "input_size": 378,
+          "model": "ViT-SO400M-14-SigLIP-384",
+          "name": "siglip",
+          "pretrained": "webli",
+          "type": "open_clip",
+          "use_summary": true
+        },
+        {
+          "fd_normalize": false,
+          "feature_distillation": true,
+          "input_size": 378,
+          "model": "dinov2_vitg14_reg",
+          "name": "dino_v2",
+          "type": "dino_v2",
+          "use_summary": true
+        },
+        {
+          "fd_normalize": false,
+          "feature_distillation": true,
+          "input_size": 1024,
+          "model": "vit-h",
+          "name": "sam",
+          "type": "sam",
+          "use_summary": false
+        }
+      ],
+      "torchcompile": null,
+      "torchscript": false,
+      "train_interpolation": "random",
+      "train_split": "train",
+      "tta": 0,
+      "use_coco": false,
+      "use_multi_epochs_loader": false,
+      "val_ema_only": false,
+      "val_split": "val",
+      "vflip": 0.0,
+      "vitdet_version": 1,
+      "wandb_entity": "",
+      "wandb_job_type": "",
+      "wandb_name": "",
+      "wandb_project": "",
+      "warmup_lr": 1e-05,
+      "warmup_prefix": false,
+      "worker_seeding": "all",
+      "workers": 8,
+      "world_size": 256
+    },
+    "auto_map": {
+      "AutoConfig": "nvidia/C-RADIOv2-H--hf_model.RADIOConfig",
+      "AutoModel": "nvidia/C-RADIOv2-H--hf_model.RADIOModel"
+    },
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": null,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": null,
+    "exponential_decay_length_penalty": null,
+    "feature_normalizer_config": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "inter_feature_normalizer_config": null,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "max_resolution": 2048,
+    "min_length": 0,
+    "model_type": "",
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "patch_size": 16,
+    "preferred_resolution": [
+      768,
+      768
+    ],
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "bfloat16",
+    "torchscript": false,
+    "transformers_version": "4.51.3",
+    "typical_p": 1.0,
+    "use_bfloat16": true,
+    "version": "radio_v2.5-h",
+    "vitdet_window_size": null
+  },
+  "eos_token_id": 2,
+  "image_size": [
+    2048,
+    1648
+  ],
+  "is_encoder_decoder": true,
+  "max_sequence_length": 9000,
+  "model_type": "nemotron_parse",
+  "pad_token_id": 1,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.3",
+  "vocab_size": 52327
+}

generation_config.json

@@ -0,0 +1,13 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "decoder_start_token_id": 2,
+  "eos_token_id": 2,
+  "forced_eos_token_id": 2,
+  "pad_token_id": 1,
+  "max_new_tokens": 9000,
+  "do_sample": false,
+  "num_beams": 1,
+  "repetition_penalty": 1.1,
+  "transformers_version": "4.51.3"
+}

hf_nemotron_parse_config.py

@@ -0,0 +1,136 @@
+from os import truncate
+from quopri import decodestring
+from transformers import PretrainedConfig
+from typing import List, Optional
+
+from transformers.dynamic_module_utils import get_class_from_dynamic_module
+
+class NemotronParseTextConfig(PretrainedConfig):
+    """
+    Configuration class for NemotronParse text decoder (mBART-based).
+    """
+    model_type = "nemotron_parse_text"
+
+    def __init__(
+        self,
+        vocab_size: int = 250027,
+        d_model: int = 1024,
+        encoder_layers: int = 12,
+        decoder_layers: int = 12,
+        encoder_attention_heads: int = 16,
+        decoder_attention_heads: int = 16,
+        decoder_ffn_dim: int = 4096,
+        encoder_ffn_dim: int = 4096,
+        activation_function: str = "gelu",
+        dropout: float = 0.1,
+        attention_dropout: float = 0.0,
+        activation_dropout: float = 0.0,
+        classifier_dropout: float = 0.0,
+        init_std: float = 0.02,
+        encoder_layerdrop: float = 0.0,
+        decoder_layerdrop: float = 0.0,
+        scale_embedding: bool = False,
+        use_cache: bool = True,
+        num_labels: int = 3,
+        forced_eos_token_id: int = 2,
+        add_cross_attention: bool = True,  # Enable cross-attention for vision-encoder-decoder
+        is_decoder: bool = True,  # This is a decoder
+        max_sequence_length: int = 9000,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.encoder_layers = encoder_layers
+        self.decoder_layers = decoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_attention_heads = decoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.activation_function = activation_function
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.classifier_dropout = classifier_dropout
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.scale_embedding = scale_embedding
+        self.use_cache = use_cache
+        self.num_labels = num_labels
+        self.add_cross_attention = add_cross_attention
+        self.is_decoder = is_decoder
+        
+        # Add hidden_size as alias for d_model (for compatibility)
+        self.hidden_size = self.d_model
+        self.forced_eos_token_id = forced_eos_token_id
+        self.num_attention_heads = self.encoder_attention_heads
+
+        self.max_sequence_length = max_sequence_length
+
+
+class NemotronParseConfig(PretrainedConfig):
+    """
+    Configuration class for NemotronParse model.
+    
+    This configuration class is used to store the configuration of a [`NemotronParseForConditionalGeneration`] model.
+    It is used to instantiate an NemotronParse model according to the specified arguments, defining the vision and text model configs.
+    """
+    model_type = "nemotron_parse"
+    is_composition = True
+    max_sequence_length = 9000
+    
+    def __init__(
+        self,
+        encoder: Optional[dict] = None,
+        decoder: Optional[dict] = None,
+        tie_word_embeddings: bool = False,
+        decoder_start_token_id: int = 2,
+        pad_token_id: int = 1,
+        eos_token_id: int = 2,
+        bos_token_id: int = 0,
+        image_size: List[int] = [2048, 1648],
+        is_encoder_decoder: bool = True,
+        max_sequence_length: int = 9000,
+        **kwargs
+    ):
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            decoder_start_token_id=decoder_start_token_id,
+            pad_token_id=pad_token_id,
+            eos_token_id=eos_token_id,
+            bos_token_id=bos_token_id,
+            max_sequence_length=max_sequence_length,
+            **kwargs
+        )
+        
+        
+        if decoder is None:
+            decoder = {}
+            
+        if encoder is not None:
+            assert "auto_map" in encoder and "AutoConfig" in encoder["auto_map"]
+            vision_auto_config = get_class_from_dynamic_module(*encoder["auto_map"]["AutoConfig"].split("--")[::-1])
+            self.encoder = vision_auto_config(**encoder)
+        else:
+            self.encoder = PretrainedConfig()
+
+        decoder["max_sequence_length"] = max_sequence_length
+        self.decoder = NemotronParseTextConfig(**decoder)
+        self.image_size = image_size
+        
+        # Initialize vocab size from text config
+        self.vocab_size = self.decoder.vocab_size
+        self.is_encoder_decoder = is_encoder_decoder
+        self.max_sequence_length = max_sequence_length
+        
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+        """
+        output = super().to_dict()
+        output["encoder"] = self.encoder.to_dict()
+        output["decoder"] = self.decoder.to_dict()
+        output["model_type"] = self.model_type
+        output["is_encoder_decoder"] = self.is_encoder_decoder
+        return output 

hf_nemotron_parse_modeling.py

@@ -0,0 +1,585 @@
+import math
+import torch
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss
+from transformers import PreTrainedModel, GenerationMixin
+from transformers.models.vision_encoder_decoder.modeling_vision_encoder_decoder import VisionEncoderDecoderModel
+from transformers.models.vision_encoder_decoder.configuration_vision_encoder_decoder import VisionEncoderDecoderConfig
+from transformers.modeling_outputs import Seq2SeqLMOutput
+from transformers.models.mbart.modeling_mbart import MBartPreTrainedModel, MBartConfig, MBartScaledWordEmbedding, MBartDecoderLayer, BaseModelOutputWithPastAndCrossAttentions
+from transformers.models.donut.modeling_donut_swin import DonutSwinModelOutput
+from einops import rearrange
+from typing import Optional, List, Union, Tuple
+import warnings
+from transformers.modeling_outputs import BaseModelOutput
+from transformers.models.encoder_decoder.modeling_encoder_decoder import shift_tokens_right
+from hf_nemotron_parse_config import NemotronParseConfig
+from transformers import AutoModel
+import time
+from transformers.modeling_attn_mask_utils import (
+    _prepare_4d_attention_mask,
+    _prepare_4d_attention_mask_for_sdpa,
+    _prepare_4d_causal_attention_mask,
+    _prepare_4d_causal_attention_mask_for_sdpa,
+)
+
+
+class NemotronParseDecoder(MBartPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`MBartDecoderLayer`]
+
+    Args:
+        config: MBartConfig
+        embed_tokens (nn.Embedding): output embedding
+    """
+
+    def __init__(self, config: MBartConfig, embed_tokens: Optional[nn.Embedding] = None):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        self.embed_tokens = MBartScaledWordEmbedding(
+            config.vocab_size, config.d_model, self.padding_idx, embed_scale=embed_scale
+        )
+
+        if embed_tokens is not None:
+            self.embed_tokens.weight = embed_tokens.weight
+
+        self.layers = nn.ModuleList([MBartDecoderLayer(config) for _ in range(config.decoder_layers)])
+        self.config = config
+
+        self.layernorm_embedding = nn.LayerNorm(config.d_model)
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        cross_attn_head_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
+        r"""
+        Args:
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+                [`PreTrainedTokenizer.__call__`] for details.
+
+                [What are input IDs?](../glossary#input-ids)
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
+                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
+                selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the cross-attention modules in the decoder to avoid performing
+                cross-attention on hidden heads. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
+                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input = input_ids
+            input_shape = input.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            input = inputs_embeds[:, :, -1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if self.config._attn_implementation == "flash_attention_2":
+            # 2d mask is passed through the layers
+            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+        elif self.config._attn_implementation == "sdpa" and not output_attentions and cross_attn_head_mask is None:
+            # output_attentions=True & cross_attn_head_mask can not be supported when using SDPA, and we fall back on
+            # the manual implementation that requires a 4D causal mask in all cases.
+            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+                attention_mask,
+                input_shape,
+                inputs_embeds,
+                past_key_values_length,
+            )
+        else:
+            # 4d mask is passed through the layers
+            attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask, input_shape, inputs_embeds, past_key_values_length
+            )
+
+        # expand encoder attention mask
+        if encoder_hidden_states is not None and encoder_attention_mask is not None:
+            if self.config._attn_implementation == "flash_attention_2":
+                encoder_attention_mask = encoder_attention_mask if 0 in encoder_attention_mask else None
+            elif self.config._attn_implementation == "sdpa" and cross_attn_head_mask is None and not output_attentions:
+                # output_attentions=True & cross_attn_head_mask can not be supported when using SDPA, and we fall back on
+                # the manual implementation that requires a 4D causal mask in all cases.
+                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+                encoder_attention_mask = _prepare_4d_attention_mask_for_sdpa(
+                    encoder_attention_mask,
+                    inputs_embeds.dtype,
+                    tgt_len=input_shape[-1],
+                )
+            else:
+                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+                encoder_attention_mask = _prepare_4d_attention_mask(
+                    encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
+                )
+        hidden_states = inputs_embeds
+        hidden_states = self.layernorm_embedding(hidden_states)
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing`. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+        next_decoder_cache = () if use_cache else None
+
+        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
+        for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
+            if attn_mask is not None:
+                if attn_mask.size()[0] != len(self.layers):
+                    raise ValueError(
+                        f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
+                        f" {attn_mask.size()[0]}."
+                    )
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            if self.training:
+                dropout_probability = torch.rand([])
+                if dropout_probability < self.layerdrop:
+                    continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    head_mask[idx] if head_mask is not None else None,
+                    cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
+                    None,
+                    output_attentions,
+                    use_cache,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                    cross_attn_layer_head_mask=(
+                        cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None
+                    ),
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        hidden_states = self.layer_norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class RadioWithNeck(nn.Module):
+    """Vision encoder using RADIO model with custom neck."""
+    
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        
+        self.model_encoder = AutoModel.from_config(config, trust_remote_code=True)
+        
+        # Neck components
+        last_hidden_state = 1024
+        self.conv1 = nn.Conv1d(1280, last_hidden_state, 1)
+        self.layer_norm1 = nn.LayerNorm(last_hidden_state, eps=1e-06, elementwise_affine=True)
+        self.conv2 = nn.Conv2d(last_hidden_state, last_hidden_state, kernel_size=(1,4), stride=(1,4), padding=0, bias=False)
+        self.layer_norm2 = nn.LayerNorm(last_hidden_state, eps=1e-06, elementwise_affine=True)
+        self.sum_proj = nn.Linear(3840, last_hidden_state)
+        self.layer_norm3 = nn.LayerNorm(last_hidden_state, eps=1e-06, elementwise_affine=True)
+        
+    def forward(self, pixel_values, output_attentions=False, output_hidden_states=False, return_dict=False, **kwargs):
+        radio_output = self.model_encoder(pixel_values)
+        summary, feature = radio_output
+        
+        
+        output = self.conv1(feature.permute(0,2,1)).permute(0,2,1)
+        output = self.layer_norm1(output)
+        
+        patch_size = self.config.patch_size
+        output = rearrange(output, 'b (h w) d -> b d h w',
+                    h=pixel_values.shape[-2] // patch_size,
+                    w=pixel_values.shape[-1] // patch_size)
+        
+        output = self.conv2(output)
+        output = rearrange(output, 'b d h w -> b (h w) d')
+        output = self.layer_norm2(output)
+        summary = self.layer_norm3(self.sum_proj(summary))
+        output = torch.cat((output, summary.unsqueeze(1)), dim=1)
+        
+        return DonutSwinModelOutput(last_hidden_state=output)
+
+
+class NemotronParsePreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models.
+    """
+    config_class = NemotronParseConfig
+    base_model_prefix = "vision_encoder_decoder"  # Use VisionEncoderDecoder prefix
+    main_input_name = "pixel_values"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["RadioWithNeck", "MBartDecoder"]
+    _skip_keys_device_placement = "past_key_values"
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=self.config.decoder.init_std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.decoder.init_std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+# Based on transformers.models.encoder_decoder.modeling_encoder_decoder
+class NemotronParseForConditionalGeneration(NemotronParsePreTrainedModel, GenerationMixin):
+    """
+    NemotronParse model for conditional generation tasks.
+    
+    This model combines a RADIO-based vision encoder with an mBART-based text decoder.
+    """
+    
+    def __init__(self, config: NemotronParseConfig):
+        super().__init__(config)
+        
+        self.encoder = RadioWithNeck(config.encoder)
+        self.encoder.main_input_name = 'pixel_values'
+        self.encoder = self.encoder.to(config.encoder.torch_dtype)
+        
+        self.decoder = NemotronParseDecoder(config.decoder)
+        self.decoder = self.decoder.to(config.decoder.torch_dtype)
+
+        self.lm_head = nn.Linear(config.decoder.d_model, config.decoder.vocab_size, bias=False, dtype=config.decoder.torch_dtype)
+ 
+        # Extra heads
+        num_extra_heads = getattr(config, 'num_extra_heads', 0)
+        self.decoder.extra_heads = nn.ModuleList([
+            nn.Linear(config.decoder.d_model, config.decoder.d_model) 
+            for _ in range(num_extra_heads)
+        ])
+        self.decoder.extra_proj = nn.ModuleList([
+            nn.Linear(config.decoder.d_model, config.decoder.d_model)
+            for _ in range(num_extra_heads)
+        ])
+        
+        # Class token index for loss weighting
+        self.class_token_indx_start = getattr(config, 'class_token_start_idx', 50000)
+        
+        self.post_init()
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+    
+    def get_input_embeddings(self):
+        return self.decoder.get_input_embeddings()
+
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.BoolTensor] = None,
+        encoder_outputs: Optional[Tuple[torch.FloatTensor]] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        __subflavors__: Optional[str] = None,
+        __keys__: Optional[List[str]] = None,
+        return_sample_losses: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]:
+        
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        kwargs_encoder = {argument: value for argument, value in kwargs.items() if not argument.startswith("decoder_")}
+
+        kwargs_decoder = {
+            argument[len("decoder_") :]: value for argument, value in kwargs.items() if argument.startswith("decoder_")
+        }
+
+        if encoder_outputs is None:
+            if pixel_values is None:
+                raise ValueError("You have to specify pixel_values")
+
+            encoder_outputs = self.encoder(
+                pixel_values,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                **kwargs_encoder,
+            )
+
+        elif isinstance(encoder_outputs, tuple):
+            encoder_outputs = BaseModelOutput(*encoder_outputs)
+
+        encoder_hidden_states = encoder_outputs[0]
+
+        encoder_attention_mask = None
+
+        if (labels is not None) and (decoder_input_ids is None and decoder_inputs_embeds is None):
+            decoder_input_ids = shift_tokens_right(
+                labels, self.config.pad_token_id, self.config.decoder_start_token_id
+            )
+
+        output_hidden_states = True
+
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            inputs_embeds=decoder_inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            use_cache=use_cache,
+            past_key_values=past_key_values,
+            return_dict=return_dict,
+            **kwargs_decoder,
+        )
+        loss = None
+
+        if labels is not None:
+            main_logits = self.lm_head(decoder_outputs.last_hidden_state)
+            logits = [main_logits]
+            decoder_inputs_embeds = decoder_outputs.inputs_embeds
+            for iii, head in enumerate(self.decoder.extra_heads):
+
+                decoder_input_embeds_shift = self.decoder.extra_proj[iii](torch.cat((decoder_inputs_embeds[:,1:,:], torch.zeros_like(decoder_inputs_embeds[:,0,:].unsqueeze(1))), axis=1))
+                hidden = head(decoder_outputs['hidden_states'][-1] + decoder_input_embeds_shift)
+                logits.append(self.lm_head(hidden))  # Use main lm_head, NOT decoder.lm_head
+
+            logits = torch.stack(logits, dim=-2)
+            loss_fct = CrossEntropyLoss(reduction="none")
+
+            losses_per_head = []
+            tokens_per_head = []
+            for head_num in range(len(self.decoder.extra_heads)+1):
+                logits_head = logits[:,:,head_num,:]
+                labels_head = torch.cat(
+                    (labels[:, head_num:], torch.full_like(labels[:, :head_num], -100)),
+                    1
+                )
+                loss_full = loss_fct(logits_head.permute(0, 2, 1), labels_head)
+                loss_full[labels_head >= self.class_token_indx_start] *= 10
+                losses_per_head.append(loss_full.sum(1))
+                tokens_per_head.append((labels_head != -100).sum(1))
+
+            losses_per_sample = torch.stack(losses_per_head, dim=1).sum(1)
+            tokens_per_sample = torch.stack(tokens_per_head, dim=1).sum(1)
+            loss = losses_per_sample.sum() / (tokens_per_sample.sum() + 1e-6)
+            if return_sample_losses is not None:
+                return_sample_losses.copy_(losses_per_sample.detach() / (tokens_per_sample + 1e-6))
+
+        if not return_dict:
+            if loss is not None:
+                return (loss,) + decoder_outputs + encoder_outputs
+            else:
+                return decoder_outputs + encoder_outputs
+        output_logits = self.lm_head(decoder_outputs.last_hidden_state)
+        return Seq2SeqLMOutput(
+            loss=loss,
+            logits=output_logits, 
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+    def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
+        return shift_tokens_right(labels, self.config.pad_token_id, self.config.decoder_start_token_id)
+
+
+    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of: Optional[int] = None):
+        """Resize token embeddings and update lm_head accordingly."""
+        # Resize decoder embeddings
+        new_embeddings = self.decoder.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
+        
+        # Update lm_head to match new vocab size
+        if new_embeddings is not None:
+            old_vocab_size, hidden_size = self.lm_head.weight.shape
+            new_vocab_size = new_embeddings.num_embeddings
+            
+            if old_vocab_size != new_vocab_size:
+                print(f"Resizing lm_head from {old_vocab_size} to {new_vocab_size} tokens")
+                new_lm_head = nn.Linear(hidden_size, new_vocab_size, bias=False, device=self.lm_head.weight.device, dtype=self.lm_head.weight.dtype)
+                
+                # Copy old weights to new lm_head
+                num_tokens_to_copy = min(old_vocab_size, new_vocab_size)
+                new_lm_head.weight.data[:num_tokens_to_copy] = self.lm_head.weight.data[:num_tokens_to_copy]
+                
+                # Update reference
+                self.lm_head = new_lm_head
+                # DO NOT update decoder.lm_head - keep them separate
+        
+        return new_embeddings
+
+    def _reorder_cache(self, past_key_values, beam_idx):
+        # apply decoder cache reordering here
+        return self.decoder._reorder_cache(past_key_values, beam_idx)
+
+
+# Copied from transformers.models.encoder_decoder.modeling_encoder_decoder.shift_tokens_right
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
+    if decoder_start_token_id is None:
+        raise ValueError("Make sure to set the decoder_start_token_id attribute of the model's configuration.")
+    shifted_input_ids[:, 0] = decoder_start_token_id
+
+    if pad_token_id is None:
+        raise ValueError("Make sure to set the pad_token_id attribute of the model's configuration.")
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+    return shifted_input_ids

hf_nemotron_parse_processor.py

@@ -0,0 +1,376 @@
+import numpy as np
+from PIL import Image
+from typing import List, Optional, Union, Dict, Any
+import torch
+from torchvision import transforms as T
+import albumentations as A
+import cv2
+import json
+
+from transformers import ProcessorMixin, BaseImageProcessor, ImageProcessingMixin
+from transformers.tokenization_utils_base import BatchEncoding
+from transformers.image_utils import ChannelDimension, ImageInput, PILImageResampling, infer_channel_dimension_format
+from transformers.utils import TensorType
+
+
+class NemotronParseImageProcessor(BaseImageProcessor, ImageProcessingMixin):
+    """
+    Image processor for NemotronParse model.
+    
+    This processor inherits from BaseImageProcessor to be compatible with transformers AutoImageProcessor.
+    """
+    
+    model_input_names = ["pixel_values"]
+    
+    def __init__(
+        self,
+        final_size: tuple = (2048, 1648),
+        **kwargs,
+    ):
+        clean_kwargs = {}
+        for k, v in kwargs.items():
+            if not k.startswith('_') and k not in ['transform', 'torch_transform']:
+                clean_kwargs[k] = v
+        
+        if 'size' in clean_kwargs:
+            size_config = clean_kwargs.pop('size')
+            if isinstance(size_config, dict):
+                if 'longest_edge' in size_config:
+                    longest_edge = size_config['longest_edge']
+                    if isinstance(longest_edge, (list, tuple)):
+                        final_size = tuple(int(x) for x in longest_edge)
+                    else:
+                        final_size = (int(longest_edge), int(longest_edge))
+                elif 'height' in size_config and 'width' in size_config:
+                    final_size = (int(size_config['height']), int(size_config['width']))
+        
+        super().__init__(**clean_kwargs)
+        
+        if isinstance(final_size, (list, tuple)) and len(final_size) >= 2:
+            self.final_size = (int(final_size[0]), int(final_size[1]))
+        elif isinstance(final_size, (int, float)):
+            self.final_size = (int(final_size), int(final_size))
+        else:
+            self.final_size = (2048, 1648)  # Default fallback
+        
+        self._create_transforms()
+    
+    def _create_transforms(self):
+        """Create transform objects (not serialized to JSON)."""
+        if isinstance(self.final_size, (list, tuple)):
+            self.target_height, self.target_width = int(self.final_size[0]), int(self.final_size[1])
+        else:
+            self.target_height = self.target_width = int(self.final_size)
+        
+        self.transform = A.Compose([
+            A.PadIfNeeded(
+                min_height=self.target_height, 
+                min_width=self.target_width, 
+                border_mode=cv2.BORDER_CONSTANT, 
+                value=[255, 255, 255],
+                p=1.0
+            ),
+        ])
+        
+        self.torch_transform = T.Compose([
+            T.ToTensor(),
+            # Note: Normalization is done within RADIO model
+        ])
+
+    def to_dict(self):
+        """Override to exclude non-serializable transforms."""
+        output = super().to_dict()
+        output.pop('transform', None)
+        output.pop('torch_transform', None)
+        return output
+    
+    @classmethod
+    def from_dict(cls, config_dict: dict, **kwargs):
+        """Override to recreate transforms after loading."""
+        config_dict = config_dict.copy()
+        config_dict.pop('transform', None)
+        config_dict.pop('torch_transform', None)
+        
+        # Clean any problematic entries
+        for key in list(config_dict.keys()):
+            if key.startswith('_') or config_dict[key] is None:
+                config_dict.pop(key, None)
+        
+        # Ensure numeric types are correct
+        if 'final_size' in config_dict:
+            final_size = config_dict['final_size']
+            if isinstance(final_size, (list, tuple)):
+                config_dict['final_size'] = tuple(int(x) for x in final_size)
+        
+        try:
+            return cls(**config_dict, **kwargs)
+        except Exception as e:
+            print(f"Warning: Error in from_dict: {e}")
+            print("Using default parameters...")
+            return cls(**kwargs)
+    
+    def save_pretrained(self, save_directory, **kwargs):
+        """Save image processor configuration."""
+        import os
+        import json
+        
+        os.makedirs(save_directory, exist_ok=True)
+        
+        # Save preprocessor config in standard HuggingFace format
+        config = {
+            "feature_extractor_type": "NemotronParseImageProcessor",
+            "image_processor_type": "NemotronParseImageProcessor", 
+            "processor_class": "NemotronParseImageProcessor",
+            "size": {
+                "height": self.final_size[0],
+                "width": self.final_size[1],
+                "longest_edge": self.final_size
+            },
+            "final_size": self.final_size,
+        }
+        
+        config_path = os.path.join(save_directory, "preprocessor_config.json")
+        with open(config_path, 'w') as f:
+            json.dump(config, f, indent=2)
+
+    def _resize_with_aspect_ratio(self, image: np.ndarray) -> np.ndarray:
+        """Resize image maintaining aspect ratio (exact replica of original LongestMaxSizeHW)."""
+        height, width = image.shape[:2]
+        max_size_height = self.target_height
+        max_size_width = self.target_width
+        
+        # Original LongestMaxSizeHW algorithm from custom_augmentations.py
+        aspect_ratio = width / height
+        new_height = height
+        new_width = width
+
+        if height > max_size_height:
+            new_height = max_size_height
+            new_width = int(new_height * aspect_ratio)
+
+        if new_width > max_size_width:
+            new_width = max_size_width
+            new_height = int(new_width / aspect_ratio)
+        
+        return cv2.resize(image, (new_width, new_height), interpolation=cv2.INTER_LINEAR)
+    
+    def _pad_to_size(self, image: np.ndarray) -> np.ndarray:
+        """Pad image to target size with white padding (matches A.PadIfNeeded behavior)."""
+        h, w = image.shape[:2]
+        min_height, min_width = self.target_height, self.target_width
+        
+        pad_h = max(0, min_height - h)
+        pad_w = max(0, min_width - w)
+        
+        if pad_h == 0 and pad_w == 0:
+            return image
+        
+        if len(image.shape) == 3:
+            padded = np.pad(
+                image, 
+                ((0, pad_h), (0, pad_w), (0, 0)), 
+                mode='constant', 
+                constant_values=255
+            )
+        else:
+            padded = np.pad(
+                image, 
+                ((0, pad_h), (0, pad_w)), 
+                mode='constant', 
+                constant_values=255
+            )
+        
+        return padded
+    
+    def preprocess(
+        self,
+        images: ImageInput,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs,
+    ) -> Dict[str, torch.Tensor]:
+        """
+        Preprocess an image or batch of images for the NemotronParse model.
+        
+        Args:
+            images: Input image(s)
+            return_tensors: Type of tensors to return
+        """
+        
+        # Ensure images is a list
+        if not isinstance(images, list):
+            images = [images]
+        
+        # Convert PIL images to numpy arrays if needed
+        processed_images = []
+        for image in images:
+            if isinstance(image, Image.Image):
+                image = np.asarray(image)
+            processed_images.append(image)
+        
+        # Apply NemotronParse-specific transforms
+        pixel_values = []
+        for image in processed_images:
+            processed_image = self._resize_with_aspect_ratio(image)
+            
+            if self.transform is not None:
+                transformed = self.transform(image=processed_image)
+                processed_image = transformed["image"]
+            else:
+                # Fallback: just pad to target size
+                processed_image = self._pad_to_size(processed_image)
+            
+            pixel_values_tensor = self.torch_transform(processed_image)
+            
+            if pixel_values_tensor.shape[0] == 1:
+                pixel_values_tensor = pixel_values_tensor.expand(3, -1, -1)
+                
+            pixel_values.append(pixel_values_tensor)
+                        
+        pixel_values = torch.stack(pixel_values)
+        
+        data = {"pixel_values": pixel_values}
+        
+        if return_tensors is not None:
+            data = self._convert_output_format(data, return_tensors)
+            
+        return data
+    
+    def _convert_output_format(self, data: Dict[str, torch.Tensor], return_tensors: Union[str, TensorType]) -> Dict:
+        """Convert output format based on return_tensors parameter."""
+        if return_tensors == "pt" or return_tensors == TensorType.PYTORCH:
+            return data
+        elif return_tensors == "np" or return_tensors == TensorType.NUMPY:
+            return {k: v.numpy() for k, v in data.items()}
+        else:
+            return data
+    
+    def __call__(self, images: Union[Image.Image, List[Image.Image]], **kwargs) -> Dict[str, torch.Tensor]:
+        """Process images for the model (backward compatibility)."""
+        return self.preprocess(images, **kwargs)
+
+
+class NemotronParseProcessor(ProcessorMixin):
+    
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "NemotronParseImageProcessor"
+    tokenizer_class = ("PreTrainedTokenizer", "PreTrainedTokenizerFast")
+    
+    def __init__(self, image_processor=None, tokenizer=None, **kwargs):
+        if image_processor is None:
+            image_processor = NemotronParseImageProcessor(**kwargs)
+               
+        super().__init__(image_processor, tokenizer)
+            
+    
+    def __call__(
+        self,
+        images: Union[Image.Image, List[Image.Image]] = None,
+        text: Union[str, List[str]] = None,
+        add_special_tokens: bool = True,
+        padding: Union[bool, str] = False,
+        truncation: Union[bool, str] = False,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        pad_to_multiple_of: Optional[int] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_offsets_mapping: bool = False,
+        return_token_type_ids: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+        return_tensors: Optional[Union[str, "TensorType"]] = None,
+        **kwargs
+    ) -> BatchEncoding:
+        """
+        Main method to prepare for the model one or several text(s) and image(s).
+        """
+        
+        # Process images
+        if images is not None:
+            image_inputs = self.image_processor(images, **kwargs)
+        else:
+            image_inputs = {}
+        
+        # Process text
+        if text is not None:
+            text_inputs = self.tokenizer(
+                text,
+                add_special_tokens=add_special_tokens,
+                padding=padding,
+                truncation=truncation,
+                max_length=max_length,
+                stride=stride,
+                pad_to_multiple_of=pad_to_multiple_of,
+                return_attention_mask=return_attention_mask,
+                return_overflowing_tokens=return_overflowing_tokens,
+                return_special_tokens_mask=return_special_tokens_mask,
+                return_offsets_mapping=return_offsets_mapping,
+                return_token_type_ids=return_token_type_ids,
+                return_length=return_length,
+                verbose=verbose,
+                return_tensors=return_tensors,
+                **kwargs,
+            )
+        else:
+            text_inputs = {}
+        
+        # Combine inputs
+        return BatchEncoding({**image_inputs, **text_inputs})
+    
+    def decode(self, *args, **kwargs):
+        """Decode token ids to strings."""
+        return self.tokenizer.decode(*args, **kwargs)
+    
+    def batch_decode(self, *args, **kwargs):
+        """Batch decode token ids to strings."""
+        return self.tokenizer.batch_decode(*args, **kwargs)
+    
+    def post_process_generation(self, sequences, fix_markdown=False):
+        """Post-process generated sequences."""
+        if hasattr(self.tokenizer, 'post_process_generation'):
+            return self.tokenizer.post_process_generation(sequences, fix_markdown=fix_markdown)
+        else:
+            # Fallback processing
+            if isinstance(sequences, str):
+                sequences = [sequences]
+            
+            processed = []
+            for seq in sequences:
+                # Basic cleaning
+                seq = seq.replace('<s>', '').replace('</s>', '').strip()
+                processed.append(seq)
+            
+            return processed[0] if len(processed) == 1 else processed
+    
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        """
+        Load processor from pretrained model.
+        
+        This method is compatible with AutoProcessor.from_pretrained().
+        """
+        # Use the parent class's from_pretrained method which handles auto-loading
+        return super().from_pretrained(pretrained_model_name_or_path, **kwargs)
+    
+    def save_pretrained(self, save_directory, **kwargs):
+        """
+        Save processor to directory.
+        
+        This method is compatible with AutoProcessor/AutoImageProcessor loading.
+        """
+        import os
+        os.makedirs(save_directory, exist_ok=True)
+        
+        # Save tokenizer with proper configuration for AutoTokenizer
+        print("Saving tokenizer for AutoTokenizer compatibility...")
+        self.tokenizer.save_pretrained(save_directory, **kwargs)
+        
+        # Save image processor  
+        print("Saving image processor...")
+        self.image_processor.save_pretrained(save_directory, **kwargs)
+        
+        # Use the parent class's save_pretrained method for processor config
+        super().save_pretrained(save_directory, **kwargs)
+        print(f"NemotronParseProcessor saved to {save_directory}")
+        print(f"AutoTokenizer.from_pretrained('{save_directory}') should now work!")

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81cae06dbfa407fce43e8624cc25a167340eff9e710492e890039605e2ac2570
+size 3827116504

preprocessor_config.json

@@ -0,0 +1,20 @@
+{
+  "feature_extractor_type": "NemotronParseImageProcessor",
+  "image_processor_type": "NemotronParseImageProcessor",
+  "processor_class": "NemotronParseProcessor",
+  "do_normalize": false,
+  "do_rescale": true,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "height": 2048,
+    "width": 1648,
+    "longest_edge": [
+      2048,
+      1648
+    ]
+  },
+  "final_size": [
+    2048,
+    1648
+  ]
+}

special_tokens_map.json

@@ -0,0 +1,39 @@
+{
+  "additional_special_tokens": [
+    {
+      "content": "<predict_classes>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false
+    }
+  ],
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}