transformers 5.0.0rc0__py3-none-any.whl → 5.0.0rc2__py3-none-any.whl

transformers/models/gemma3n/modeling_gemma3n.py

@@ -32,21 +32,19 @@ from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
+from ...integrations import use_kernelized_func
 from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
-from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import ModelOutput, TransformersKwargs, auto_docstring, can_return_tuple, logging
+from ...utils import ModelOutput, TransformersKwargs, auto_docstring, can_return_tuple
+from ...utils.generic import check_model_inputs, maybe_autocast
 from ..auto import AutoModel
 from .configuration_gemma3n import Gemma3nAudioConfig, Gemma3nConfig, Gemma3nTextConfig, Gemma3nVisionConfig
 
 
-logger = logging.get_logger(__name__)
-
-
 @dataclass
 @auto_docstring(
     custom_intro="""
@@ -331,6 +329,16 @@ class Gemma3nAudioAttention(nn.Module):
         r_softplus_0 = 1.0 / torch.nn.functional.softplus(torch.tensor(0.0))
         self.register_buffer("q_scale", (q_scale * r_softplus_0).clone().detach(), persistent=False)
 
+        local_causal_valid_mask = self.create_local_causal_valid_mask()
+        self.register_buffer("local_causal_valid_mask", local_causal_valid_mask, persistent=False)
+
+        self.register_buffer(
+            "softcap",
+            torch.tensor(self.attention_logits_soft_cap).float(),
+            persistent=False,
+        )
+
+    def create_local_causal_valid_mask(self):
         lower_causal_mask = torch.tril(
             torch.ones((self.context_size, self.chunk_size), dtype=torch.bool),
             diagonal=0,
@@ -341,13 +349,7 @@ class Gemma3nAudioAttention(nn.Module):
         )
         local_causal_valid_mask = torch.ones((self.chunk_size, self.context_size), dtype=torch.bool)
         local_causal_valid_mask = local_causal_valid_mask * lower_causal_mask * upper_causal_mask
-        self.register_buffer("local_causal_valid_mask", local_causal_valid_mask, persistent=False)
-
-        self.register_buffer(
-            "softcap",
-            torch.tensor(self.attention_logits_soft_cap).float(),
-            persistent=False,
-        )
+        return local_causal_valid_mask
 
     def _pad_dim1(self, x: torch.Tensor, pad_left: int, pad_right: int) -> torch.Tensor:
         batch, _, *tail_shape = x.shape
@@ -921,9 +923,10 @@ class Gemma3nAudioEncoder(PreTrainedModel):
         self.conformer = nn.ModuleList(
             [Gemma3nAudioConformerBlock(config) for _ in range(config.conf_num_hidden_layers)]
         )
+        self.post_init()
 
     def forward(
-        self, audio_mel: torch.Tensor, audio_mel_mask: torch.BoolTensor
+        self, audio_mel: torch.Tensor, audio_mel_mask: torch.BoolTensor, **kwargs
     ) -> tuple[torch.Tensor, torch.BoolTensor]:
         """Encodes a batch of MELs.
 
@@ -985,6 +988,7 @@ class Gemma3nTextScaledWordEmbedding(nn.Embedding):
 
     def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int, embed_scale: float = 1.0):
         super().__init__(num_embeddings, embedding_dim, padding_idx)
+        self.scalar_embed_scale = embed_scale
         self.register_buffer("embed_scale", torch.tensor(embed_scale), persistent=False)
 
     def forward(self, input_ids: torch.Tensor):
@@ -1228,6 +1232,7 @@ def apply_rotary_pos_emb(
     return (x * cos) + (rotate_half(x) * sin)
 
 
+@use_kernelized_func(apply_rotary_pos_emb)
 class Gemma3nTextAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -1254,7 +1259,6 @@ class Gemma3nTextAttention(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.sliding_window = config.sliding_window if self.layer_type == "sliding_attention" else None
         self.is_sliding = self.layer_type == "sliding_attention"
 
@@ -1283,7 +1287,7 @@ class Gemma3nTextAttention(nn.Module):
         attention_mask: Optional[torch.Tensor] = None,
         past_key_values: Optional[Cache] = None,
         cache_position: Optional[torch.LongTensor] = None,
-        **kwargs: Unpack[FlashAttentionKwargs],
+        **kwargs: Unpack[TransformersKwargs],
     ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, self.config.head_dim)
@@ -1379,10 +1383,8 @@ class Gemma3nTextDecoderLayer(GradientCheckpointingLayer):
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
         past_key_values: Optional[Cache] = None,
-        output_attentions: Optional[bool] = False,
-        use_cache: Optional[bool] = False,
         cache_position: Optional[torch.LongTensor] = None,
-        **kwargs,
+        **kwargs: Unpack[TransformersKwargs],
     ) -> tuple[torch.Tensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
         predictions = self.altup.predict(hidden_states)
         active_prediction = predictions[self.config.altup_active_idx]
@@ -1390,14 +1392,12 @@ class Gemma3nTextDecoderLayer(GradientCheckpointingLayer):
         active_prediction_normed = self.input_layernorm(active_prediction)
         laurel_output = self.laurel(active_prediction_normed)
 
-        attn, self_attn_weights = self.self_attn(
+        attn, _ = self.self_attn(
             hidden_states=active_prediction_normed,
             attention_mask=attention_mask,
             position_ids=position_ids,
             position_embeddings=position_embeddings,
             past_key_values=past_key_values,
-            output_attentions=output_attentions,
-            use_cache=use_cache,
             cache_position=cache_position,
             **kwargs,
         )
@@ -1426,154 +1426,7 @@ class Gemma3nTextDecoderLayer(GradientCheckpointingLayer):
         first_prediction = self.post_per_layer_input_norm(first_prediction)
         corrected_predictions[1:] += first_prediction
 
-        outputs = (corrected_predictions,)
-
-        if output_attentions:
-            outputs += (self_attn_weights,)
-
-        return outputs
-
-
-class Gemma3nMLP(nn.Module):
-    def __init__(self, config):
-        super().__init__()
-        self.config = config
-        self.hidden_size = config.hidden_size
-        self.intermediate_size = config.intermediate_size
-        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
-        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_activation]
-
-    def forward(self, x):
-        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
-        return down_proj
-
-
-class Gemma3nAttention(nn.Module):
-    """Multi-headed attention from 'Attention Is All You Need' paper"""
-
-    def __init__(self, config: Gemma3nConfig, layer_idx: int):
-        super().__init__()
-        self.layer_type = config.layer_types[layer_idx] if hasattr(config, "layer_types") else None
-        self.config = config
-        self.layer_idx = layer_idx
-        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
-        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
-        self.scaling = config.query_pre_attn_scalar**-0.5
-        self.attention_dropout = self.config.attention_dropout
-        self.is_causal = not getattr(config, "use_bidirectional_attention", False)
-
-        self.q_proj = nn.Linear(
-            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
-        )
-        self.k_proj = nn.Linear(
-            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
-        )
-        self.v_proj = nn.Linear(
-            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
-        )
-        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.attn_logit_softcapping = self.config.attn_logit_softcapping
-        self.sliding_window = config.sliding_window if self.layer_type == "sliding_attention" else None
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        past_key_values: Optional[Cache] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        **kwargs: Unpack[FlashAttentionKwargs],
-    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
-        input_shape = hidden_states.shape[:-1]
-        hidden_shape = (*input_shape, -1, self.head_dim)
-
-        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
-        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
-        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
-
-        cos, sin = position_embeddings
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
-
-        if past_key_values is not None:
-            # sin and cos are specific to RoPE models; cache_position needed for the static cache
-            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
-            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
-
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
-
-        attn_output, attn_weights = attention_interface(
-            self,
-            query_states,
-            key_states,
-            value_states,
-            attention_mask,
-            dropout=self.attention_dropout if self.training else 0.0,
-            scaling=self.scaling,
-            sliding_window=self.sliding_window,
-            softcap=self.attn_logit_softcapping,
-            **kwargs,
-        )
-
-        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
-        attn_output = self.o_proj(attn_output)
-        return attn_output, attn_weights
-
-
-class Gemma3nDecoderLayer(GradientCheckpointingLayer):
-    def __init__(self, config: Gemma3nConfig, layer_idx: int):
-        super().__init__()
-        self.hidden_size = config.hidden_size
-        self.config = config
-        self.attention_type = config.layer_types[layer_idx]
-        self.self_attn = Gemma3nAttention(config=config, layer_idx=layer_idx)
-        self.mlp = Gemma3nMLP(config)
-        self.input_layernorm = Gemma3nRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.post_attention_layernorm = Gemma3nRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-
-        self.pre_feedforward_layernorm = Gemma3nRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.post_feedforward_layernorm = Gemma3nRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Cache] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        **kwargs,
-    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
-        residual = hidden_states
-
-        hidden_states = self.input_layernorm(hidden_states)
-
-        # Self Attention
-        hidden_states, _ = self.self_attn(
-            hidden_states=hidden_states,
-            position_embeddings=position_embeddings,
-            attention_mask=attention_mask,
-            position_ids=position_ids,
-            past_key_values=past_key_values,
-            cache_position=cache_position,
-            **kwargs,
-        )
-        hidden_states = self.post_attention_layernorm(hidden_states)
-        hidden_states = residual + hidden_states
-
-        residual = hidden_states
-        hidden_states = self.pre_feedforward_layernorm(hidden_states)
-        hidden_states = self.mlp(hidden_states)
-        hidden_states = self.post_feedforward_layernorm(hidden_states)
-        hidden_states = residual + hidden_states
-
-        return hidden_states
+        return corrected_predictions
 
 
 @auto_docstring
@@ -1590,8 +1443,8 @@ class Gemma3nPreTrainedModel(PreTrainedModel):
     _can_compile_fullgraph = True
     _supports_attention_backend = True
     _can_record_outputs = {
-        "hidden_states": Gemma3nDecoderLayer,
-        "attentions": Gemma3nAttention,
+        "hidden_states": Gemma3nTextDecoderLayer,
+        "attentions": Gemma3nTextAttention,
     }
     input_modalities = ("image", "text", "audio")
 
@@ -1602,8 +1455,38 @@ class Gemma3nPreTrainedModel(PreTrainedModel):
             init.ones_(module.weight)
         elif isinstance(module, Gemma3nAudioAttention):
             init.zeros_(module.per_dim_scale)
+            q_scale = module.head_dim**-0.5
+            r_softplus_0 = 1.0 / torch.nn.functional.softplus(torch.tensor(0.0))
+            init.copy_(module.q_scale, q_scale * r_softplus_0)
+            init.constant_(module.softcap, module.attention_logits_soft_cap)
+            init.copy_(module.local_causal_valid_mask, module.create_local_causal_valid_mask())
+        elif isinstance(module, Gemma3nTextScaledWordEmbedding):
+            init.constant_(module.embed_scale, module.scalar_embed_scale)
         elif isinstance(module, Gemma3nTextAltUp):
             init.zeros_(module.correct_output_scale)
+            init.constant_(module.router_input_scale, self.config.hidden_size**-1.0)
+        elif isinstance(module, Gemma3nAudioRelativePositionEmbedding):
+            min_timescale, max_timescale = 1.0, 1.0e4
+            num_timescales = module.channels // 2
+            log_timescale_increment = math.log(float(max_timescale) / float(min_timescale)) / max(
+                num_timescales - 1, 1
+            )
+            inv_timescales = min_timescale * torch.exp(torch.arange(num_timescales) * -log_timescale_increment)
+            init.copy_(module.inv_timescales, inv_timescales.float().unsqueeze(0).unsqueeze(0))
+        elif isinstance(module, Gemma3nTextModel):
+            init.constant_(module.per_layer_projection_scale, self.hidden_size**-0.5)
+            init.constant_(module.per_layer_input_scale, 1 / math.sqrt(2.0))
+        elif isinstance(module, Gemma3nRotaryEmbedding):
+            for layer_type in module.layer_types:
+                rope_init_fn = module.compute_default_rope_parameters
+                if module.rope_type[layer_type] != "default":
+                    rope_init_fn = ROPE_INIT_FUNCTIONS[module.rope_type[layer_type]]
+                curr_inv_freq, _ = rope_init_fn(module.config, layer_type=layer_type)
+                init.copy_(getattr(module, f"{layer_type}_inv_freq"), curr_inv_freq)
+                init.copy_(getattr(module, f"{layer_type}_original_inv_freq"), curr_inv_freq)
+
+        if hasattr(module, "gradient_clipping"):
+            init.constant_(module.gradient_clipping, self.config.gradient_clipping)
 
 
 class Gemma3nRotaryEmbedding(nn.Module):
@@ -1629,7 +1512,7 @@ class Gemma3nRotaryEmbedding(nn.Module):
                 rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type[layer_type]]
             curr_inv_freq, curr_attention_scaling = rope_init_fn(self.config, device, layer_type=layer_type)
             self.register_buffer(f"{layer_type}_inv_freq", curr_inv_freq, persistent=False)
-            setattr(self, f"{layer_type}_original_inv_freq", curr_inv_freq)
+            self.register_buffer(f"{layer_type}_original_inv_freq", curr_inv_freq.clone(), persistent=False)
             setattr(self, f"{layer_type}_attention_scaling", curr_attention_scaling)
 
     @staticmethod
@@ -1678,7 +1561,7 @@ class Gemma3nRotaryEmbedding(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() * attention_scaling
@@ -1741,7 +1624,7 @@ class Gemma3nTextModel(Gemma3nPreTrainedModel):
         # Initialize weights and apply final processing
         self.post_init()
 
-    @can_return_tuple
+    @check_model_inputs(tie_last_hidden_states=False)
     @auto_docstring
     def forward(
         self,
@@ -1752,8 +1635,6 @@ class Gemma3nTextModel(Gemma3nPreTrainedModel):
         past_key_values: Optional[Cache] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
         use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
         **kwargs: Unpack[TransformersKwargs],
     ) -> BaseModelOutputWithPast:
@@ -1761,37 +1642,21 @@ class Gemma3nTextModel(Gemma3nPreTrainedModel):
         per_layer_inputs (torch.Tensor, *optional*, defaults to None):
             Pre-computed per-layer embeddings. If None, they are derived from input_ids if provided.
         """
-        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
-
         if (input_ids is None) ^ (inputs_embeds is not None):
             raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
 
-        if self.gradient_checkpointing and self.training and use_cache:
-            logger.warning_once(
-                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
-            )
-            use_cache = False
-
         if input_ids is not None:
             inputs_embeds = self.embed_tokens(input_ids)
             per_layer_inputs = self.get_per_layer_inputs(input_ids)
 
         per_layer_inputs = self.project_per_layer_inputs(inputs_embeds, per_layer_inputs)
 
-        if use_cache and past_key_values is None and not self.training:
+        if use_cache and past_key_values is None:
             past_key_values = DynamicCache(config=self.config)
 
         if cache_position is None:
             past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
-            cache_position = torch.arange(
-                past_seen_tokens,
-                past_seen_tokens + inputs_embeds.shape[1],
-                device=inputs_embeds.device,
-            )
+            cache_position = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
 
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
@@ -1835,39 +1700,21 @@ class Gemma3nTextModel(Gemma3nPreTrainedModel):
         for layer_type in self.config.layer_types:
             position_embeddings[layer_type] = self.rotary_emb(hidden_states, position_ids, layer_type)
 
-        # decoder layers
-        all_hidden_states = () if output_hidden_states else None
-        all_self_attns = () if output_attentions else None
-
         for decoder_layer in self.layers[: self.config.num_hidden_layers]:
-            if output_hidden_states:
-                all_hidden_states += (hidden_states,)
-
             causal_mask = causal_mask_mapping[decoder_layer.attention_type]
             per_layer_input = per_layer_inputs[:, :, decoder_layer.layer_idx, :]
 
-            layer_outputs = decoder_layer(
+            hidden_states = decoder_layer(
                 hidden_states,
                 position_embeddings[decoder_layer.attention_type],
                 per_layer_input,
                 attention_mask=causal_mask,
                 position_ids=position_ids,
                 past_key_values=past_key_values,
-                output_attentions=output_attentions,
-                use_cache=use_cache,
                 cache_position=cache_position,
                 **kwargs,
             )
 
-            hidden_states = layer_outputs[0]
-
-            if output_attentions:
-                all_self_attns += (layer_outputs[1],)
-
-        # add hidden states from the last decoder layer (but before reprojecting to stay consistent with layer output)
-        if output_hidden_states:
-            all_hidden_states += (hidden_states,)
-
         # Per-layer inputs to single output
         target_magnitude = torch.mean(hidden_states[0] ** 2, dim=-1, keepdim=True) ** 0.5
         temp_hidden_states = [hidden_states[0]]
@@ -1887,8 +1734,6 @@ class Gemma3nTextModel(Gemma3nPreTrainedModel):
         return BaseModelOutputWithPast(
             last_hidden_state=hidden_states,
             past_key_values=past_key_values,
-            hidden_states=all_hidden_states,
-            attentions=all_self_attns,
         )
 
     def get_per_layer_inputs(self, input_ids: torch.LongTensor) -> torch.Tensor:
@@ -2175,7 +2020,7 @@ class Gemma3nModel(Gemma3nPreTrainedModel):
         use_cache: Optional[bool] = None,
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
-        **lm_kwargs,
+        **lm_kwargs: Unpack[TransformersKwargs],
     ) -> Gemma3nCausalLMOutputWithPast:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -2363,7 +2208,7 @@ class Gemma3nForConditionalGeneration(Gemma3nPreTrainedModel, GenerationMixin):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         logits_to_keep: Union[int, torch.Tensor] = 0,
-        **lm_kwargs,
+        **lm_kwargs: Unpack[TransformersKwargs],
     ) -> Gemma3nCausalLMOutputWithPast:
         r"""
         input_features_mask (torch.Tensor, *optional*, defaults to None):
@@ -2492,6 +2337,7 @@ class Gemma3nForConditionalGeneration(Gemma3nPreTrainedModel, GenerationMixin):
         use_cache=True,
         logits_to_keep=None,
         labels=None,
+        is_first_iteration=False,
         **kwargs,
     ):
         # Overwritten -- custom `position_ids` and `pixel_values` handling
@@ -2505,13 +2351,14 @@ class Gemma3nForConditionalGeneration(Gemma3nPreTrainedModel, GenerationMixin):
             use_cache=use_cache,
             logits_to_keep=logits_to_keep,
             token_type_ids=token_type_ids,
+            is_first_iteration=is_first_iteration,
             **kwargs,
         )
 
         # If we're in cached decoding stage, multimodal inputs should be None because input ids do not contain special
         # tokens anymore. Otherwise multimodal inputs should be passed to model.
         # NOTE: use_cache=False always needs pixel_values, input_features, and input_features_mask
-        if cache_position[0] == 0:
+        if is_first_iteration or not use_cache:
             model_inputs["pixel_values"] = pixel_values
             model_inputs["input_features"] = input_features
             model_inputs["input_features_mask"] = input_features_mask