fusion-bench 0.2.19__py3-none-any.whl → 0.2.21__py3-none-any.whl

fusion_bench/models/modeling_smile_gemma2/modeling_smile_gemma2.py

@@ -0,0 +1,986 @@
+from functools import partial
+from typing import Callable, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, HybridCache, StaticCache
+from transformers.generation import GenerationMixin
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.models.gemma2.modeling_gemma2 import (
+    _CHECKPOINT_FOR_DOC,
+    GEMMA2_INPUTS_DOCSTRING,
+    Gemma2RMSNorm,
+    Gemma2RotaryEmbedding,
+    apply_rotary_pos_emb,
+    eager_attention_forward,
+)
+from transformers.processing_utils import Unpack
+from transformers.utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    can_return_tuple,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+
+from fusion_bench.models.smile_moe.linear_from_hf_config import SmileLinear
+
+from .configuration_smile_gemma2 import SmileGemma2Config
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "SmileGemma2Config"
+
+
+class SmileGemma2MLP(nn.Module):
+    def __init__(self, config: SmileGemma2Config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        # * --- replace nn.Linear with SmileLinear ---
+        self.gate_proj = SmileLinear(config, self.hidden_size, self.intermediate_size)
+        self.up_proj = SmileLinear(config, self.hidden_size, self.intermediate_size)
+        self.down_proj = SmileLinear(config, self.intermediate_size, self.hidden_size)
+        # * --- end of replacement ---
+        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 SmileGemma2Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: SmileGemma2Config, layer_idx: int):
+        super().__init__()
+        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 = True
+
+        # * --- replace nn.Linear with SmileLinear ---
+        self.q_proj = SmileLinear(
+            config,
+            config.hidden_size,
+            config.num_attention_heads * self.head_dim,
+            bias=config.attention_bias,
+        )
+        self.k_proj = SmileLinear(
+            config,
+            config.hidden_size,
+            config.num_key_value_heads * self.head_dim,
+            bias=config.attention_bias,
+        )
+        self.v_proj = SmileLinear(
+            config,
+            config.hidden_size,
+            config.num_key_value_heads * self.head_dim,
+            bias=config.attention_bias,
+        )
+        self.o_proj = SmileLinear(
+            config,
+            config.num_attention_heads * self.head_dim,
+            config.hidden_size,
+            bias=config.attention_bias,
+        )
+        # * --- end of replacement ---
+        self.attn_logit_softcapping = self.config.attn_logit_softcapping
+        self.sliding_window = config.sliding_window if not bool(layer_idx % 2) else None
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: 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_value 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,
+                "sliding_window": self.sliding_window,
+            }
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs
+            )
+
+            # Here we need to slice as we use a static cache by default, but FA2 does not support it
+            if (
+                attention_mask is not None
+                and self.config._attn_implementation == "flash_attention_2"
+            ):
+                seq_len = attention_mask.shape[-1]
+                key_states, value_states = (
+                    key_states[:, :, :seq_len, :],
+                    value_states[:, :, :seq_len, :],
+                )
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            if self.config._attn_implementation == "sdpa" and kwargs.get(
+                "output_attentions", False
+            ):
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+                )
+            else:
+                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 SmileGemma2DecoderLayer(nn.Module):
+
+    def __init__(self, config: SmileGemma2Config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.config = config
+        self.is_sliding = not bool(layer_idx % 2)
+        self.self_attn = SmileGemma2Attention(config=config, layer_idx=layer_idx)
+        self.mlp = SmileGemma2MLP(config)
+        self.input_layernorm = Gemma2RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.post_attention_layernorm = Gemma2RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+        self.pre_feedforward_layernorm = Gemma2RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.post_feedforward_layernorm = Gemma2RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.sliding_window = config.sliding_window
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        last_cache_position: int = 0,
+        **kwargs,
+    ) -> Tuple[
+        torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
+    ]:
+        if (
+            self.is_sliding and attention_mask is not None
+        ):  # efficient SDPA and no padding
+            # In prefill, we may be larger than sliding window
+            effective_seq_len = max(cache_position.shape[0], self.sliding_window)
+            # For FA2, the mask is 2D and is of shape [bs, processed_tokens] (not [bs, max_cache_len]),
+            # thus we must slice from the right (at most `effective_seq_len` elements)
+            if self.config._attn_implementation == "flash_attention_2":
+                attention_mask = attention_mask[:, -effective_seq_len:]
+            # Otherwise, the mask is 4D of shape [bs, 1, query_len, max_cache_len] thus we must slice
+            # from the left, with an offset if we are beyond the sliding window
+            else:
+                min_dtype = torch.finfo(attention_mask.dtype).min
+                sliding_window_mask = torch.tril(
+                    torch.ones_like(attention_mask, dtype=torch.bool),
+                    diagonal=-self.sliding_window,
+                )
+                attention_mask = torch.where(
+                    sliding_window_mask, min_dtype, attention_mask
+                )
+                # In case we are beyond the sliding window, we need to correctly offset the mask slicing
+                # `last_cache_position` is equivalent to `cache_position[-1]` but without breaking dynamo
+                offset = last_cache_position - effective_seq_len
+                # Should only be used when beyond the sliding window (i.e. offset > 0)
+                offset = max(0, offset)
+                attention_mask = attention_mask[
+                    :, :, :, offset : offset + effective_seq_len
+                ]
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            position_embeddings=position_embeddings,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            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
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
+
+
+class SmileGemma2PreTrainedModel(PreTrainedModel):
+    config_class = SmileGemma2Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["SmileGemma2DecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+    _supports_attention_backend = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+class SmileGemma2Model(SmileGemma2PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Gemma2DecoderLayer`]
+
+    Args:
+        config: Gemma2Config
+    """
+
+    def __init__(self, config: SmileGemma2Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(
+            config.vocab_size, config.hidden_size, self.padding_idx
+        )
+        self.layers = nn.ModuleList(
+            [
+                SmileGemma2DecoderLayer(config, layer_idx)
+                for layer_idx in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = Gemma2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Gemma2RotaryEmbedding(config=config)
+        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
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(GEMMA2_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[HybridCache] = 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,
+        last_cache_position: Optional[int] = None,
+        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+    ) -> BaseModelOutputWithPast:
+        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 inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None and not self.training:
+            batch_size, seq_len, _ = inputs_embeds.shape
+            # NOTE: ideally, `HybridCache` should be initialized outside the model with `layer_device_map`
+            past_key_values = HybridCache(
+                self.config,
+                max_batch_size=batch_size,
+                max_cache_len=seq_len,
+                dtype=inputs_embeds.dtype,
+                device=self.device,
+            )
+
+        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,
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        # This is needed to correctly slice the mask without data-dependent slicing later on if using dynamo tracing
+        # (retrieving the same value from `cache_position` later on would crash dynamo)
+        if last_cache_position is None:
+            last_cache_position = 0
+            if attention_mask is not None:
+                # In case a 4d mask is passed directly without using `generate`, we have to rely on cache_position
+                # It will break dynamo tracing but there are no way around it (and it should never happen in practice)
+                last_cache_position = (
+                    attention_mask.shape[-1]
+                    if attention_mask.dim() == 2
+                    else cache_position[-1].item()
+                )
+        causal_mask = self._update_causal_mask(
+            attention_mask,
+            inputs_embeds,
+            cache_position,
+            past_key_values,
+            output_attentions,
+        )
+
+        # embed positions
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # normalized
+        # Gemma2 downcasts the below to float16, causing sqrt(3072)=55.4256 to become 55.5
+        # See https://github.com/huggingface/transformers/pull/29402
+        normalizer = torch.tensor(
+            self.config.hidden_size**0.5, dtype=hidden_states.dtype
+        )
+        hidden_states = hidden_states * normalizer
+
+        # 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,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
+                    hidden_states,
+                    position_embeddings,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    last_cache_position,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    position_embeddings=position_embeddings,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    last_cache_position=last_cache_position,
+                    **flash_attn_kwargs,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    @torch.no_grad()
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: HybridCache,
+        output_attentions: bool,
+    ):
+        # Flash Attention currently doesn't support static cache but Gemma2 work only with static cache.
+        # So we will pass in attention mask as is in any case, not only when ther's padding. Then we'll use its shape
+        # to cut out keys/values trailing 0 used in static cache. This workaround should be compile compatible
+        # as it doesn't cause dynamic control issues.
+        if self.config._attn_implementation == "flash_attention_2":
+            return attention_mask
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        sequence_length = input_tensor.shape[1]
+        if isinstance(past_key_values, (HybridCache, StaticCache)):
+            target_length = past_key_values.get_max_cache_shape()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if attention_mask is not None
+                else input_tensor.shape[1]
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        **kwargs,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
+                `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache,
+                to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            device (`torch.device`):
+                The device to place the 4D attention mask on.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length),
+                fill_value=min_dtype,
+                dtype=dtype,
+                device=device,
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(
+                target_length, device=device
+            ) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = (
+                    causal_mask.clone()
+                )  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[
+                    :, None, None, :
+                ].to(causal_mask.device)
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[
+                    :, :, :, :mask_length
+                ].masked_fill(padding_mask, min_dtype)
+
+        return causal_mask
+
+
+class SmileGemma2ForCausalLM(SmileGemma2PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = SmileGemma2Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @can_return_tuple
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    @add_start_docstrings_to_model_forward(GEMMA2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(
+        output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[HybridCache] = 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,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **loss_kwargs,
+    ) -> CausalLMOutputWithPast:
+        r"""
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+            logits_to_keep (`int` or `torch.Tensor`, *optional*):
+                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+                This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Gemma2ForCausalLM
+
+        >>> model = Gemma2ForCausalLM.from_pretrained("google/gemma-2-9b")
+        >>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b")
+
+        >>> prompt = "What is your favorite condiment?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "What is your favorite condiment?"
+        ```"""
+
+        if self.training and self.config._attn_implementation != "eager":
+            logger.warning_once(
+                "It is strongly recommended to train Gemma2 models with the `eager` attention implementation "
+                f"instead of `{self.config._attn_implementation}`. Use `eager` with `AutoModelForCausalLM.from_pretrained('<path-to-checkpoint>', attn_implementation='eager')`."
+            )
+        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
+        )
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs: BaseModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            cache_position=cache_position,
+            **loss_kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = (
+            slice(-logits_to_keep, None)
+            if isinstance(logits_to_keep, int)
+            else logits_to_keep
+        )
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        if self.config.final_logit_softcapping is not None:
+            logits = logits / self.config.final_logit_softcapping
+            logits = torch.tanh(logits)
+            logits = logits * self.config.final_logit_softcapping
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        use_cache=True,
+        logits_to_keep=None,
+        **kwargs,
+    ):
+        # Overwritten: has a special cache type, `HybridCache`
+
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            cache_position=cache_position,
+            position_ids=position_ids,
+            use_cache=use_cache,
+            logits_to_keep=logits_to_keep,
+            **kwargs,
+        )
+
+        # This is needed to correctly slice the mask without data-dependent slicing later on if using dynamo tracing
+        # (retrieving the same value from `cache_position` later on would crash dynamo)
+        model_inputs["last_cache_position"] = (
+            attention_mask.shape[-1] if attention_mask is not None else 0
+        )
+        if logits_to_keep is None:
+            _ = model_inputs.pop("logits_to_keep", None)
+
+        if (
+            isinstance(past_key_values, HybridCache)
+            and attention_mask.ndim == 2
+            and not self.config._attn_implementation == "flash_attention_2"
+        ):
+            if model_inputs["inputs_embeds"] is not None:
+                batch_size, sequence_length, _ = model_inputs["inputs_embeds"].shape
+                device = model_inputs["inputs_embeds"].device
+            else:
+                batch_size, sequence_length = model_inputs["input_ids"].shape
+                device = model_inputs["input_ids"].device
+
+            attention_mask = (
+                self.model._prepare_4d_causal_attention_mask_with_cache_position(
+                    attention_mask,
+                    sequence_length=sequence_length,
+                    target_length=past_key_values.get_max_cache_shape(),
+                    dtype=self.lm_head.weight.dtype,
+                    device=device,
+                    cache_position=cache_position,
+                    batch_size=batch_size,
+                )
+            )
+            model_inputs["attention_mask"] = attention_mask
+
+        return model_inputs
+
+
+class SmileGemma2ForSequenceClassification(SmileGemma2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = SmileGemma2Model(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(GEMMA2_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = 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,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> SequenceClassifierOutputWithPast:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+
+        transformer_outputs: BaseModelOutputWithPast = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        hidden_states = transformer_outputs.last_hidden_state
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError(
+                "Cannot handle batch sizes > 1 if no padding token is defined."
+            )
+        if self.config.pad_token_id is None:
+            last_non_pad_token = -1
+        elif input_ids is not None:
+            # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id
+            non_pad_mask = (input_ids != self.config.pad_token_id).to(
+                logits.device, torch.int32
+            )
+            token_indices = torch.arange(
+                input_ids.shape[-1], device=logits.device, dtype=torch.int32
+            )
+            last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)
+        else:
+            last_non_pad_token = -1
+            logger.warning_once(
+                f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+            )
+
+        pooled_logits = logits[
+            torch.arange(batch_size, device=logits.device), last_non_pad_token
+        ]
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                pooled_logits=pooled_logits,
+                config=self.config,
+            )
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+class SmileGemma2ForTokenClassification(SmileGemma2PreTrainedModel):
+    def __init__(self, config: SmileGemma2Config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = SmileGemma2Model(config)
+        if getattr(config, "classifier_dropout", None) is not None:
+            classifier_dropout = config.classifier_dropout
+        elif getattr(config, "hidden_dropout", None) is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.score = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(GEMMA2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = 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,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> TokenClassifierOutput:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+
+        outputs: BaseModelOutputWithPast = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        sequence_output = outputs.last_hidden_state
+        sequence_output = self.dropout(sequence_output)
+        logits = self.score(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.config)
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = [
+    "SmileGemma2PreTrainedModel",
+    "SmileGemma2Model",
+    "SmileGemma2ForCausalLM",
+    "SmileGemma2ForSequenceClassification",
+    "SmileGemma2ForTokenClassification",
+]