transformers 5.0.0__py3-none-any.whl → 5.0.0rc0__py3-none-any.whl

transformers/models/gemma3n/modular_gemma3n.py

@@ -1,3 +1,4 @@
+# coding=utf-8
 # Copyright 2025 Google Inc. HuggingFace Inc. team. All rights reserved.
 #
 #
@@ -14,8 +15,7 @@
 # limitations under the License.
 import math
 from collections.abc import Callable, Sequence
-from dataclasses import dataclass
-from typing import Any
+from typing import Any, Optional, Union
 
 import torch
 import torch.nn as nn
@@ -26,12 +26,12 @@ from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...configuration_utils import PreTrainedConfig, layer_type_validation
 from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
-from ...modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling
-from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, RopeParameters
+from ...modeling_flash_attention_utils import FlashAttentionKwargs
+from ...modeling_outputs import BaseModelOutputWithPast
+from ...modeling_rope_utils import RopeParameters
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging, torch_compilable_check
-from ...utils.generic import check_model_inputs
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
 from ..auto import AutoModel
 from ..gemma2.configuration_gemma2 import Gemma2Config
 from ..gemma2.modeling_gemma2 import (
@@ -45,7 +45,6 @@ from ..gemma3.modeling_gemma3 import (
     Gemma3DecoderLayer,
     Gemma3ForCausalLM,
     Gemma3RMSNorm,
-    Gemma3RotaryEmbedding,
     Gemma3TextModel,
     Gemma3TextScaledWordEmbedding,
 )
@@ -153,8 +152,6 @@ class Gemma3nTextConfig(Gemma2Config, PreTrainedConfig):
             The sparsity factor used to extract the top-k activations for a given layer. The provided Sequence must
             explicitly provide a sparsity value for each layer in the model. By default, the first 10 layers are
             sparse with a sparsity factor of 0.95 and the rest are dense.
-        tie_word_embeddings (`bool`, *optional*, defaults to `True`):
-            Whether to tie weight embeddings
 
     ```python
     >>> from transformers import Gemma3nTextModel, Gemma3nTextConfig
@@ -179,7 +176,7 @@ class Gemma3nTextConfig(Gemma2Config, PreTrainedConfig):
         vocab_size_per_layer_input: int = 262_144,
         hidden_size: int = 2048,
         hidden_size_per_layer_input: int = 256,
-        intermediate_size: int | Sequence[int] = 16_384,
+        intermediate_size: Union[int, Sequence[int]] = 16_384,
         num_hidden_layers: int = 35,
         num_attention_heads: int = 8,
         num_key_value_heads: int = 2,
@@ -192,11 +189,11 @@ class Gemma3nTextConfig(Gemma2Config, PreTrainedConfig):
         pad_token_id: int = 0,
         eos_token_id: int = 1,
         bos_token_id: int = 2,
-        rope_parameters: RopeParameters | dict[str, RopeParameters] | None = None,
+        rope_parameters: Optional[RopeParameters | dict[str, RopeParameters]] = None,
         attention_bias: bool = False,
         attention_dropout: float = 0.0,
         sliding_window: int = 512,
-        layer_types: Sequence[str] | None = None,
+        layer_types: Optional[Sequence[str]] = None,
         final_logit_softcapping: float = 30.0,
         altup_active_idx: int = 0,
         altup_coef_clip: float = 120.0,
@@ -204,8 +201,7 @@ class Gemma3nTextConfig(Gemma2Config, PreTrainedConfig):
         altup_num_inputs: int = 4,
         num_kv_shared_layers: int = 15,
         laurel_rank: int = 64,
-        activation_sparsity_pattern: float | Sequence[float] | None = None,
-        tie_word_embeddings: bool | None = True,
+        activation_sparsity_pattern: Optional[Union[float, Sequence[float]]] = None,
         **kwargs,
     ):
         if isinstance(intermediate_size, Sequence) and (intsize_len := len(intermediate_size)) != num_hidden_layers:
@@ -216,9 +212,6 @@ class Gemma3nTextConfig(Gemma2Config, PreTrainedConfig):
         elif not isinstance(intermediate_size, Sequence):
             intermediate_size = [intermediate_size] * num_hidden_layers
 
-        self.pad_token_id = pad_token_id
-        self.bos_token_id = bos_token_id
-        self.eos_token_id = eos_token_id
         self.vocab_size = vocab_size
         self.vocab_size_per_layer_input = vocab_size_per_layer_input
         self.max_position_embeddings = max_position_embeddings
@@ -268,11 +261,12 @@ class Gemma3nTextConfig(Gemma2Config, PreTrainedConfig):
             )
         self.activation_sparsity_pattern = activation_sparsity_pattern
         self.rope_parameters = rope_parameters
-        self.pad_token_id = pad_token_id
-        self.bos_token_id = bos_token_id
-        self.eos_token_id = eos_token_id
-        self.tie_word_embeddings = tie_word_embeddings
-        PreTrainedConfig.__init__(**kwargs)
+        PreTrainedConfig.__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs,
+        )
 
     def convert_rope_params_to_dict(self, ignore_keys_at_rope_validation=None, **kwargs):
         rope_scaling = kwargs.pop("rope_scaling", None)
@@ -494,7 +488,7 @@ class Gemma3nVisionConfig(TimmWrapperConfig):
         vocab_size: int = 128,
         vocab_offset: int = 262_144,
         rms_norm_eps: float = 1e-06,
-        model_args: dict | None = None,
+        model_args: Optional[dict] = None,
         **kwargs,
     ):
         self.architecture = architecture
@@ -544,8 +538,7 @@ class Gemma3nConfig(PreTrainedConfig):
             The audio token index to encode the audio prompt.
         initializer_range (`float`, *optional*, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        tie_word_embeddings (`bool`, *optional*, defaults to `True`):
-            Whether to tie weight embeddings
+
 
     Example:
 
@@ -580,19 +573,18 @@ class Gemma3nConfig(PreTrainedConfig):
 
     def __init__(
         self,
-        text_config: Gemma3nTextConfig | dict[str, Any] | None = None,
-        vision_config: Gemma3nVisionConfig | dict[str, Any] | None = None,
-        audio_config: Gemma3nAudioConfig | dict[str, Any] | None = None,
-        audio_soft_tokens_per_image: int | None = 188,
-        vision_soft_tokens_per_image: int | None = 256,
-        boi_token_id: int | None = 255_999,
-        eoi_token_id: int | None = 262_144,
-        image_token_id: int | None = 262_145,
-        boa_token_id: int | None = 256_000,
-        eoa_token_id: int | None = 262_272,
-        audio_token_id: int | None = 262_273,
-        initializer_range: float | None = 0.02,
-        tie_word_embeddings: bool | None = True,
+        text_config: Optional[Union[Gemma3nTextConfig, dict[str, Any]]] = None,
+        vision_config: Optional[Union[Gemma3nVisionConfig, dict[str, Any]]] = None,
+        audio_config: Optional[Union[Gemma3nAudioConfig, dict[str, Any]]] = None,
+        audio_soft_tokens_per_image: int = 188,
+        vision_soft_tokens_per_image: int = 256,
+        boi_token_id: int = 255_999,
+        eoi_token_id: int = 262_144,
+        image_token_id: int = 262_145,
+        boa_token_id: int = 256_000,
+        eoa_token_id: int = 262_272,
+        audio_token_id: int = 262_273,
+        initializer_range: float = 0.02,
         **kwargs,
     ):
         super().__init__(**kwargs)
@@ -628,18 +620,6 @@ class Gemma3nConfig(PreTrainedConfig):
         self.eoa_token_id = eoa_token_id
         self.audio_token_id = audio_token_id
         self.initializer_range = initializer_range
-        self.tie_word_embeddings = tie_word_embeddings
-
-
-@dataclass
-@auto_docstring
-class Gemma3nAudioEncoderModelOutput(BaseModelOutputWithPooling):
-    r"""
-    audio_mel_mask (`torch.FloatTensor`, *optional*):
-        A torch.BoolTensor of shape `(batch_size, num_frames)`
-    """
-
-    audio_mel_mask: torch.BoolTensor | None = None
 
 
 class Gemma3nModelOutputWithPast(PaligemmaModelOutputWithPast):
@@ -657,7 +637,7 @@ class Gemma3nModelOutputWithPast(PaligemmaModelOutputWithPast):
         audio_hidden_states of the model produced by the audio encoder and after projecting the last hidden state.
     """
 
-    audio_hidden_states: torch.FloatTensor | None = None
+    audio_hidden_states: Optional[torch.FloatTensor] = None
 
 
 class Gemma3nCausalLMOutputWithPast(PaliGemmaCausalLMOutputWithPast):
@@ -679,7 +659,7 @@ class Gemma3nCausalLMOutputWithPast(PaliGemmaCausalLMOutputWithPast):
         audio_hidden_states of the model produced by the audio encoder and after projecting the last hidden state.
     """
 
-    audio_hidden_states: torch.FloatTensor | None = None
+    audio_hidden_states: Optional[torch.FloatTensor] = None
 
 
 class Gemma3nRMSNorm(Gemma3RMSNorm):
@@ -902,16 +882,6 @@ 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,
@@ -922,7 +892,13 @@ 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
-        return 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 _pad_dim1(self, x: torch.Tensor, pad_left: int, pad_right: int) -> torch.Tensor:
         batch, _, *tail_shape = x.shape
@@ -1496,12 +1472,10 @@ class Gemma3nAudioEncoder(PreTrainedModel):
         self.conformer = nn.ModuleList(
             [Gemma3nAudioConformerBlock(config) for _ in range(config.conf_num_hidden_layers)]
         )
-        self.post_init()
 
-    @check_model_inputs
     def forward(
-        self, audio_mel: torch.Tensor, audio_mel_mask: torch.BoolTensor, **kwargs: Unpack[TransformersKwargs]
-    ) -> tuple | Gemma3nAudioEncoderModelOutput:
+        self, audio_mel: torch.Tensor, audio_mel_mask: torch.BoolTensor
+    ) -> tuple[torch.Tensor, torch.BoolTensor]:
         """Encodes a batch of MELs.
 
         Args:
@@ -1552,10 +1526,7 @@ class Gemma3nAudioEncoder(PreTrainedModel):
             current_mask = current_mask[:, :: self.config.conf_reduction_factor]
 
         audio_encodings = audio_encodings.masked_fill(current_mask.unsqueeze(-1), 0.0)
-        return Gemma3nAudioEncoderModelOutput(
-            last_hidden_state=audio_encodings,
-            audio_mel_mask=current_mask,
-        )
+        return audio_encodings, current_mask
 
 
 # ==== Language Model ====
@@ -1686,7 +1657,7 @@ class Gemma3nTextAltUp(nn.Module):
         innovation = activated - predictions[self.config.altup_active_idx]  # (batch, num_tokens, hidden_size)
         innovation = innovation.repeat(self.config.altup_num_inputs, 1, 1, 1)  # Repeat on dim0 to match predictions
 
-        if self.training and self.config.altup_coef_clip is not None:
+        if self.config.altup_coef_clip is not None:
             self.correction_coefs.weight.data.clamp_(-self.config.altup_coef_clip, self.config.altup_coef_clip)
 
         # all_coefs adapted from jax.numpy.einsum("...p,pi->...i", ...)
@@ -1712,13 +1683,21 @@ class Gemma3nTextAltUp(nn.Module):
         return self.forward(corrected)
 
 
-def apply_rotary_pos_emb(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor, unsqueeze_dim: int = 1):
+def apply_rotary_pos_emb(
+    x: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+    unsqueeze_dim: int = 1,
+):
     """Applies Rotary Position Embedding to the query and key tensors.
 
     Args:
         x (`torch.Tensor`): The tensor to embed.
         cos (`torch.Tensor`): The cosine part of the rotary embedding.
         sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
         unsqueeze_dim (`int`, *optional*, defaults to 1):
             The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
             sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
@@ -1760,11 +1739,11 @@ class Gemma3nTextAttention(Gemma3Attention):
         self,
         hidden_states: torch.Tensor,
         position_embeddings: torch.Tensor = None,
-        attention_mask: torch.Tensor | None = None,
-        past_key_values: Cache | None = None,
-        cache_position: torch.LongTensor | None = None,
-        **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[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.config.head_dim)
 
@@ -1848,24 +1827,28 @@ class Gemma3nTextDecoderLayer(Gemma3DecoderLayer):
         hidden_states: torch.Tensor,
         position_embeddings: torch.Tensor = None,
         per_layer_input: torch.Tensor = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        cache_position: torch.LongTensor | None = None,
-        **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple[torch.Tensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
+        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,
+    ) -> tuple[torch.Tensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
         predictions = self.altup.predict(hidden_states)
         active_prediction = predictions[self.config.altup_active_idx]
 
         active_prediction_normed = self.input_layernorm(active_prediction)
         laurel_output = self.laurel(active_prediction_normed)
 
-        attn, _ = self.self_attn(
+        attn, self_attn_weights = 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,
         )
@@ -1894,17 +1877,18 @@ class Gemma3nTextDecoderLayer(Gemma3DecoderLayer):
         first_prediction = self.post_per_layer_input_norm(first_prediction)
         corrected_predictions[1:] += first_prediction
 
-        return corrected_predictions
+        outputs = (corrected_predictions,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
 
 
 class Gemma3nPreTrainedModel(Gemma2PreTrainedModel):
     config: Gemma3nConfig
     input_modalities = ("image", "text", "audio")
     _no_split_modules = ["Gemma3nTextDecoderLayer"]
-    _can_record_outputs = {
-        "hidden_states": Gemma3nTextDecoderLayer,
-        "attentions": Gemma3nTextAttention,
-    }
 
     @torch.no_grad()
     def _init_weights(self, module):
@@ -1913,42 +1897,8 @@ class Gemma3nPreTrainedModel(Gemma2PreTrainedModel):
             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(Gemma3RotaryEmbedding):
-    pass
 
 
 @auto_docstring(custom_intro="The base Gemma 3n language model without a language modeling head.")
@@ -2002,7 +1952,7 @@ class Gemma3nTextModel(Gemma3TextModel):
     def project_per_layer_inputs(
         self,
         inputs_embeds: torch.Tensor,
-        per_layer_inputs: torch.Tensor | None = None,
+        per_layer_inputs: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
         per_layer_projection: torch.Tensor = self.per_layer_model_projection(inputs_embeds)
         per_layer_projection *= self.per_layer_projection_scale.to(
@@ -2026,40 +1976,57 @@ class Gemma3nTextModel(Gemma3TextModel):
             dtype=inputs_embeds.dtype, device=per_layer_projection.device
         )
 
-    # Last hidden states should be before reprojecting, to stay consistent with the other layer outputs
-    @check_model_inputs(tie_last_hidden_states=False)
+    @can_return_tuple
     @auto_docstring
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,
-        per_layer_inputs: torch.Tensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        use_cache: bool | None = None,
-        cache_position: torch.LongTensor | None = None,
+        input_ids: Optional[torch.LongTensor] = None,
+        per_layer_inputs: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        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:
         r"""
         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:
+        if use_cache and past_key_values is None and not self.training:
             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(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
+            cache_position = torch.arange(
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device,
+            )
 
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
@@ -2103,21 +2070,39 @@ class Gemma3nTextModel(Gemma3TextModel):
         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, :]
 
-            hidden_states = decoder_layer(
+            layer_outputs = 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]]
@@ -2137,6 +2122,8 @@ class Gemma3nTextModel(Gemma3TextModel):
         return BaseModelOutputWithPast(
             last_hidden_state=hidden_states,
             past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
         )
 
 
@@ -2150,7 +2137,7 @@ class Gemma3nMultimodalEmbedder(nn.Module):
 
     def __init__(
         self,
-        multimodal_config: Gemma3nAudioConfig | Gemma3nVisionConfig,
+        multimodal_config: Union[Gemma3nAudioConfig, Gemma3nVisionConfig],
         text_config: Gemma3nTextConfig,
     ):
         super().__init__()
@@ -2169,8 +2156,8 @@ class Gemma3nMultimodalEmbedder(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,
-        inputs_embeds: torch.Tensor | None = None,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
         """Embeds token ids or soft tokens for multimodal content into language model space.
 
@@ -2213,34 +2200,37 @@ class Gemma3nModel(PaliGemmaModel):
         self.embed_vision = Gemma3nMultimodalEmbedder(config.vision_config, config.text_config)
         self.embed_audio = Gemma3nMultimodalEmbedder(config.audio_config, config.text_config)
 
-    @can_return_tuple
-    @auto_docstring(custom_intro="Projects the last hidden state from the vision model into language model space.")
-    def get_image_features(
-        self,
-        pixel_values: torch.FloatTensor,
-        **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | BaseModelOutputWithPooling:
-        vision_outputs = self.vision_tower(pixel_values=pixel_values, do_pooling=False, return_dict=True, **kwargs)
-        last_hidden_state = vision_outputs.last_hidden_state
+    def get_image_features(self, pixel_values: torch.Tensor) -> torch.Tensor:
+        """
+        Projects the last hidden state from the vision model into language model space.
+
+        Args:
+            pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`)
+               The tensors corresponding to the input images.
+
+        Returns:
+            image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`).
+        """
+        vision_outputs = self.vision_tower(
+            pixel_values=pixel_values, do_pooling=False, return_dict=True
+        ).last_hidden_state
         # Convert from (batch, channels, height, width) to (batch, height * width, channels) where:
         # height == width and height * width == Gemma3nConfig.vision_soft_tokens_per_image.
-        last_hidden_state = last_hidden_state.reshape(
-            last_hidden_state.shape[0],
+        vision_outputs = vision_outputs.reshape(
+            vision_outputs.shape[0],
             self.config.vision_config.hidden_size,
             self.config.vision_soft_tokens_per_image,
         ).permute(0, 2, 1)
         # Normalize and embed the soft tokens into language model space.
-        last_hidden_state *= self.config.vision_config.hidden_size**0.5
-        vision_outputs.pooler_output = self.embed_vision(inputs_embeds=last_hidden_state)
-
-        return vision_outputs
+        vision_outputs *= self.config.vision_config.hidden_size**0.5
+        return self.embed_vision(inputs_embeds=vision_outputs)
 
     def get_placeholder_mask(
         self,
-        input_ids: torch.LongTensor | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        image_features: torch.FloatTensor | None = None,
-        audio_features: torch.FloatTensor | None = None,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        image_features: Optional[torch.FloatTensor] = None,
+        audio_features: Optional[torch.FloatTensor] = None,
     ):
         """
         Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
@@ -2263,18 +2253,16 @@ class Gemma3nModel(PaliGemmaModel):
 
         n_image_tokens = special_image_mask.sum()
         special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
-        if image_features is not None:
-            torch_compilable_check(
-                inputs_embeds[special_image_mask].numel() == image_features.numel(),
-                f"Image features and image tokens do not match, tokens: {n_image_tokens}, features: {image_features.shape[0] * image_features.shape[1]}",
+        if image_features is not None and inputs_embeds[special_image_mask].numel() != image_features.numel():
+            raise ValueError(
+                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {image_features.shape[0] * image_features.shape[1]}"
             )
 
         n_audio_tokens = special_audio_mask.sum()
         special_audio_mask = special_audio_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
-        if audio_features is not None:
-            torch_compilable_check(
-                inputs_embeds[special_audio_mask].numel() == audio_features.numel(),
-                f"Audio features and audio tokens do not match, tokens: {n_audio_tokens}, features: {audio_features.shape[0] * audio_features.shape[1]}",
+        if audio_features is not None and inputs_embeds[special_audio_mask].numel() != audio_features.numel():
+            raise ValueError(
+                f"Audio features and image tokens do not match: tokens: {n_audio_tokens}, features {audio_features.shape[0] * audio_features.shape[1]}"
             )
 
         return special_image_mask, special_audio_mask
@@ -2282,21 +2270,21 @@ class Gemma3nModel(PaliGemmaModel):
     @can_return_tuple
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,  # text inputs
-        pixel_values: torch.FloatTensor | None = None,  # vision inputs
-        input_features: torch.FloatTensor | None = None,  # audio inputs
-        attention_mask: torch.Tensor | None = None,
-        input_features_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        token_type_ids: torch.LongTensor | None = None,
-        cache_position: torch.LongTensor | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        labels: torch.LongTensor | None = None,
-        use_cache: bool | None = None,
-        output_attentions: bool | None = None,
-        output_hidden_states: bool | None = None,
-        **lm_kwargs: Unpack[TransformersKwargs],
+        input_ids: Optional[torch.LongTensor] = None,  # text inputs
+        pixel_values: Optional[torch.FloatTensor] = None,  # vision inputs
+        input_features: Optional[torch.FloatTensor] = None,  # audio inputs
+        attention_mask: Optional[torch.Tensor] = None,
+        input_features_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        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,
+        **lm_kwargs,
     ) -> Gemma3nCausalLMOutputWithPast:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -2308,8 +2296,7 @@ class Gemma3nModel(PaliGemmaModel):
 
         ```python
         >>> from PIL import Image
-        >>> import httpx
-        >>> from io import BytesIO
+        >>> import requests
         >>> from transformers import AutoProcessor, Gemma3nForConditionalGeneration
 
         >>> model = Gemma3nForConditionalGeneration.from_pretrained("google/gemma3n2-3b-mix-224")
@@ -2317,8 +2304,7 @@ class Gemma3nModel(PaliGemmaModel):
 
         >>> prompt = "Where is the cat standing?"
         >>> url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
-        >>> with httpx.stream("GET", url) as response:
-        ...     image = Image.open(BytesIO(response.read()))
+        >>> image = Image.open(requests.get(url, stream=True).raw)
 
         >>> inputs = processor(images=image, text=prompt,  return_tensors="pt")
 
@@ -2368,7 +2354,7 @@ class Gemma3nModel(PaliGemmaModel):
 
         # Merge text and images
         if pixel_values is not None:
-            image_features = self.get_image_features(pixel_values, return_dict=True).pooler_output
+            image_features = self.get_image_features(pixel_values)
             image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
             special_image_mask, _ = self.get_placeholder_mask(
                 input_ids, inputs_embeds=inputs_embeds, image_features=image_features
@@ -2377,9 +2363,7 @@ class Gemma3nModel(PaliGemmaModel):
 
         # Merge text and audio
         if input_features is not None and input_features_mask is not None:
-            audio_features = self.get_audio_features(input_features, ~input_features_mask, return_dict=True)
-            audio_features = audio_features.pooler_output
-            audio_mask = audio_features.audio_mel_mask
+            audio_features, audio_mask = self.get_audio_features(input_features, ~input_features_mask)
 
             # The Gemma3nProcessor expects all audio will be 30s in length and inserts 188 audio soft tokens into the
             # text to account for this. However, the audio preprocessing and encoder do not gurarantee they will
@@ -2425,27 +2409,23 @@ class Gemma3nModel(PaliGemmaModel):
             audio_hidden_states=audio_features if input_features is not None else None,
         )
 
-    @can_return_tuple
-    @auto_docstring(custom_intro="Projects the last hidden state from the audio encoder into language model space.")
     def get_audio_features(
-        self,
-        input_features: torch.Tensor,
-        input_features_mask: torch.Tensor,
-        **kwargs: Unpack[TransformersKwargs],
-    ) -> tuple | Gemma3nAudioEncoderModelOutput:
-        r"""
-        input_features (`torch.FloatTensor]` of shape `(num_images, seq_length, num_features)`):
-            The tensors corresponding to the input audio.
-        input_features_mask (`torch.FloatTensor]` of shape `(num_images, seq_length)`):
-            The attention mask for the input audio.
+        self, input_features: torch.Tensor, input_features_mask: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         """
-        audio_outputs: Gemma3nAudioEncoderModelOutput = self.audio_tower(
-            input_features, input_features_mask, return_dict=True, **kwargs
-        )
-        audio_embeds = self.embed_audio(inputs_embeds=audio_outputs.last_hidden_state)
-        audio_outputs.pooler_output = audio_embeds
+        Projects the last hidden state from the audio encoder into language model space.
 
-        return audio_outputs
+        Args:
+            input_features (`torch.FloatTensor]` of shape `(num_images, seq_length, num_features)`):
+               The tensors corresponding to the input audio.
+            input_features_mask (`torch.FloatTensor]` of shape `(num_images, seq_length)`):
+               The attention mask for the input audio.
+
+        Returns:
+            audio_features (`torch.Tensor`): Audio feature tensor of shape `(num_images, audio_length, embed_dim)`).
+        """
+        audio_outputs, audio_mask = self.audio_tower(input_features, input_features_mask)
+        return self.embed_audio(inputs_embeds=audio_outputs), audio_mask
 
 
 @auto_docstring(
@@ -2461,22 +2441,22 @@ class Gemma3nForConditionalGeneration(PaliGemmaForConditionalGeneration):
     @auto_docstring
     def forward(
         self,
-        input_ids: torch.LongTensor | None = None,  # text inputs
-        pixel_values: torch.FloatTensor | None = None,  # vision inputs
-        input_features: torch.FloatTensor | None = None,  # audio inputs
-        attention_mask: torch.Tensor | None = None,
-        input_features_mask: torch.Tensor | None = None,
-        position_ids: torch.LongTensor | None = None,
-        past_key_values: Cache | None = None,
-        token_type_ids: torch.LongTensor | None = None,
-        cache_position: torch.LongTensor | None = None,
-        inputs_embeds: torch.FloatTensor | None = None,
-        labels: torch.LongTensor | None = None,
-        use_cache: bool | None = None,
-        output_attentions: bool | None = None,
-        output_hidden_states: bool | None = None,
-        logits_to_keep: int | torch.Tensor = 0,
-        **lm_kwargs: Unpack[TransformersKwargs],
+        input_ids: Optional[torch.LongTensor] = None,  # text inputs
+        pixel_values: Optional[torch.FloatTensor] = None,  # vision inputs
+        input_features: Optional[torch.FloatTensor] = None,  # audio inputs
+        attention_mask: Optional[torch.Tensor] = None,
+        input_features_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        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,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **lm_kwargs,
     ) -> Gemma3nCausalLMOutputWithPast:
         r"""
         input_features_mask (torch.Tensor, *optional*, defaults to None):
@@ -2491,8 +2471,7 @@ class Gemma3nForConditionalGeneration(PaliGemmaForConditionalGeneration):
 
         ```python
         >>> from PIL import Image
-        >>> import httpx
-        >>> from io import BytesIO
+        >>> import requests
         >>> from transformers import AutoProcessor, Gemma3ForConditionalGeneration
 
         >>> model = Gemma3ForConditionalGeneration.from_pretrained("google/gemma-3-4b-it")
@@ -2606,7 +2585,6 @@ class Gemma3nForConditionalGeneration(PaliGemmaForConditionalGeneration):
         use_cache=True,
         logits_to_keep=None,
         labels=None,
-        is_first_iteration=False,
         **kwargs,
     ):
         # Overwritten -- custom `position_ids` and `pixel_values` handling
@@ -2620,14 +2598,13 @@ class Gemma3nForConditionalGeneration(PaliGemmaForConditionalGeneration):
             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 is_first_iteration or not use_cache:
+        if cache_position[0] == 0:
             model_inputs["pixel_values"] = pixel_values
             model_inputs["input_features"] = input_features
             model_inputs["input_features_mask"] = input_features_mask