modeling_gemmagain.py

# coding=utf-8
# Copyright 2025 Google Inc. HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Gemma3 - Gemma3 text model with layer looping support (wrapper approach).

This model allows running the same physical layers multiple times in sequence,
enabling parameter-efficient deep networks. Compatible with standard Gemma3 weights.

This version uses wrapper layers that pass the virtual layer index at runtime,
giving each virtual layer position unique RoPE encodings and cache slots.
"""
import copy
from typing import Callable, Optional, Union

import torch
import torch.nn as nn
from torch.nn import CrossEntropyLoss

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache, DynamicLayer
from transformers.generation import GenerationMixin
from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
from transformers.utils.deprecation import deprecate_kwarg

try:
    from .configuration_gemmagain import Gemma3Config
except ImportError:
    from configuration_gemmagain import Gemma3Config


logger = logging.get_logger(__name__)


class Gemma3TextScaledWordEmbedding(nn.Embedding):
    """
    This module overrides nn.Embeddings' forward by multiplying with embeddings scale.
    """

    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.register_buffer("embed_scale", torch.tensor(embed_scale), persistent=False)

    def forward(self, input_ids: torch.Tensor):
        return super().forward(input_ids) * self.embed_scale.to(self.weight.dtype)


class Gemma3MLP(nn.Module):
    def __init__(self, config: Gemma3Config):
        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 Gemma3RMSNorm(nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.zeros(dim))

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        output = self._norm(x.float())
        # Gemma3 uses (x * w).to(dtype) instead of x.to(dtype) * w
        output = output * (1.0 + self.weight.float())
        return output.type_as(x)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.eps}"


class Gemma3RotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor

    def __init__(self, config: Gemma3Config, device=None):
        super().__init__()
        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    @torch.no_grad()
    @dynamic_rope_update
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        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):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors."""
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """Repeat KV heads for GQA."""
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    dropout: float = 0.0,
    scaling: Optional[float] = None,
    softcap: Optional[float] = None,
    **kwargs,
) -> tuple[torch.Tensor, torch.Tensor]:
    if scaling is None:
        scaling = module.head_dim**-0.5

    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling

    if softcap is not None:
        attn_weights = attn_weights / softcap
        attn_weights = torch.tanh(attn_weights)
        attn_weights = attn_weights * softcap
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()
    return attn_output, attn_weights


class Gemma3Attention(nn.Module):
    """Multi-headed attention with support for virtual layer index."""

    def __init__(self, config: Gemma3Config, layer_idx: int):
        super().__init__()
        self.is_sliding = config.layer_types[layer_idx] == "sliding_attention"
        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 self.config.use_bidirectional_attention

        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.attn_logit_softcapping = self.config.attn_logit_softcapping
        self.sliding_window = config.sliding_window if self.is_sliding else None

        self.q_norm = Gemma3RMSNorm(dim=config.head_dim, eps=config.rms_norm_eps)
        self.k_norm = Gemma3RMSNorm(dim=config.head_dim, eps=config.rms_norm_eps)

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: torch.Tensor,
        attention_mask: Optional[torch.Tensor],
        past_key_values: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        real_layer_idx: Optional[int] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, Optional[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)

        query_states = self.q_norm(query_states)
        key_states = self.k_norm(key_states)

        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:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            # Use real_layer_idx for cache slot - each virtual layer gets its own slot
            slot_idx = real_layer_idx if real_layer_idx is not None else self.layer_idx
            key_states, value_states = past_key_values.update(key_states, value_states, slot_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,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class Gemma3DecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Gemma3Config, layer_idx: int):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.layer_idx = layer_idx
        self.attention_type = config.layer_types[layer_idx]
        self.self_attn = Gemma3Attention(config=config, layer_idx=layer_idx)
        self.mlp = Gemma3MLP(config)
        self.input_layernorm = Gemma3RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Gemma3RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
        self.pre_feedforward_layernorm = Gemma3RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
        self.post_feedforward_layernorm = Gemma3RMSNorm(self.hidden_size, eps=config.rms_norm_eps)

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings_global: torch.Tensor,
        position_embeddings_local: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        real_layer_idx: Optional[int] = None,
        **kwargs,
    ) -> torch.Tensor:
        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Apply global RoPE to non-sliding layers, local RoPE to sliding layers
        if self.self_attn.is_sliding:
            position_embeddings = position_embeddings_local
        else:
            position_embeddings = position_embeddings_global

        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,
            use_cache=use_cache,
            cache_position=cache_position,
            real_layer_idx=real_layer_idx,
            **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


@auto_docstring
class Gemma3PreTrainedModel(PreTrainedModel):
    config_class = Gemma3Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Gemma3DecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": Gemma3DecoderLayer,
        "attentions": Gemma3Attention,
    }

    def _init_weights(self, module):
        super()._init_weights(module)
        # Initialize RMSNorm weights to 0 (Gemma3 uses 1 + weight)
        if "RMSNorm" in module.__class__.__name__:
            module.weight.data.zero_()


def _expand_layer_sequence(layer_sequence, num_hidden_layers):
    """Expand layer_sequence config into a flat list of layer indices."""
    l_seq = []
    for item in layer_sequence:
        if isinstance(item, int):
            l_seq.append(item)
        elif isinstance(item, list):
            if len(item) == 2:
                start, end = item
                l_seq += list(range(start, min(end, num_hidden_layers)))
            elif len(item) == 3:
                start, end, repeats = item
                l_seq += list(range(start, min(end, num_hidden_layers))) * repeats
            else:
                raise ValueError(f"Invalid layer_sequence item: {item}")
        else:
            raise ValueError(f"Invalid layer_sequence item type: {type(item)}")
    return l_seq


def _bidirectional_window_overlay(sliding_window: int) -> Callable[[int, int, int, int], bool]:
    """Enables a bidirectional mask within the sliding window."""
    def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
        return abs(q_idx - kv_idx) < sliding_window
    return inner_mask


@auto_docstring
class Gemma3Model(Gemma3PreTrainedModel):
    def __init__(self, config: Gemma3Config):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = Gemma3TextScaledWordEmbedding(
            config.vocab_size, config.hidden_size, self.padding_idx, embed_scale=config.hidden_size**0.5
        )
        # Physical layers - the actual parameters
        self.layers = nn.ModuleList(
            [Gemma3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Gemma3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Gemma3RotaryEmbedding(config=config)
        self.gradient_checkpointing = False

        # Create local RoPE with different theta
        local_config = copy.deepcopy(config)
        local_config.rope_theta = config.rope_local_base_freq
        local_config.rope_scaling = {"rope_type": "default"}
        self.rotary_emb_local = Gemma3RotaryEmbedding(config=local_config)

        # Build the virtual layer execution schedule as plain data (FSDP-compatible)
        # Each entry is (physical_layer_idx, virtual_layer_idx)
        layer_indices = _expand_layer_sequence(config.layer_sequence, config.num_hidden_layers)
        self._execution_schedule = [(phys_idx, virt_idx) for virt_idx, phys_idx in enumerate(layer_indices)]

        # For compatibility, expose the layer sequence info
        self._layer_sequence = layer_indices
        self._num_cache_slots = len(self._execution_schedule)

        self.post_init()

    @auto_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,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        # Handle use_cache default
        effective_use_cache = use_cache if use_cache is not None else self.config.use_cache

        if effective_use_cache and not self.training:
            if past_key_values is None:
                # Create cache with enough slots for the full virtual layer sequence
                # DynamicCache doesn't pre-allocate layers, so we need to do it manually
                past_key_values = DynamicCache()
                for _ in range(self._num_cache_slots):
                    past_key_values.layers.append(DynamicLayer())
            elif isinstance(past_key_values, DynamicCache) and len(past_key_values.layers) < self._num_cache_slots:
                # Extend cache if created externally with fewer slots
                while len(past_key_values.layers) < self._num_cache_slots:
                    past_key_values.layers.append(DynamicLayer())

        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)

        # Prepare attention masks
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            mask_kwargs = {
                "config": self.config,
                "input_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "cache_position": cache_position,
                "past_key_values": past_key_values,
                "position_ids": position_ids,
            }
            sliding_mask_kwargs = mask_kwargs.copy()

            if self.config.use_bidirectional_attention:
                mask_kwargs["or_mask_function"] = lambda *args: torch.tensor(True, dtype=torch.bool)
                sliding_mask_kwargs["or_mask_function"] = _bidirectional_window_overlay(self.config.sliding_window)

            causal_mask_mapping = {
                "full_attention": create_causal_mask(**mask_kwargs),
                "sliding_attention": create_sliding_window_causal_mask(**sliding_mask_kwargs),
            }

        hidden_states = inputs_embeds
        position_embeddings_global = self.rotary_emb(hidden_states, position_ids)
        position_embeddings_local = self.rotary_emb_local(hidden_states, position_ids)

        # Execute virtual layers using the schedule (FSDP-compatible)
        for physical_idx, virtual_idx in self._execution_schedule:
            layer = self.layers[physical_idx]
            hidden_states = layer(
                hidden_states,
                position_embeddings_global=position_embeddings_global,
                position_embeddings_local=position_embeddings_local,
                attention_mask=causal_mask_mapping[layer.attention_type],
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                cache_position=cache_position,
                real_layer_idx=virtual_idx,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
        )


@auto_docstring
class Gemma3ForCausalLM(Gemma3PreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

    def __init__(self, config: Gemma3Config):
        super().__init__(config)
        self.model = Gemma3Model(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        self.post_init()

    @can_return_tuple
    @auto_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,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> CausalLMOutputWithPast:
        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,
            cache_position=cache_position,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        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:
            # Standard loss calculation
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


__all__ = [
    "Gemma3ForCausalLM",
    "Gemma3Model",
    "Gemma3PreTrainedModel",
]