configuration_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.
"""Gemmagain model configuration - Gemma3 with layer looping support"""

from transformers.configuration_utils import PretrainedConfig, layer_type_validation
from transformers.modeling_rope_utils import rope_config_validation
from transformers.utils import logging


logger = logging.get_logger(__name__)


class GemmagainConfig(PretrainedConfig):
    r"""
    Configuration class for Gemmagain - a Gemma3 text model with layer looping support.

    This extends Gemma3TextConfig to add the `layer_sequence` parameter which controls
    how layers are executed, allowing layers to be repeated multiple times.

    Args:
        vocab_size (`int`, *optional*, defaults to 262208):
            Vocabulary size of the model.
        hidden_size (`int`, *optional*, defaults to 2560):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 10240):
            Dimension of the MLP representations.
        num_hidden_layers (`int`, *optional*, defaults to 34):
            Number of hidden layers in the Transformer decoder.
        num_attention_heads (`int`, *optional*, defaults to 8):
            Number of attention heads for each attention layer.
        num_key_value_heads (`int`, *optional*, defaults to 4):
            Number of key_value heads for GQA.
        head_dim (`int`, *optional*, defaults to 256):
            The attention head dimension.
        hidden_activation (`str`, *optional*, defaults to `"gelu_pytorch_tanh"`):
            The activation function.
        max_position_embeddings (`int`, *optional*, defaults to 131072):
            Maximum sequence length.
        layer_sequence (`list`, *optional*):
            Order to execute layers. Defaults to all layers once.
            Flexible format - each item can be:
            - An integer: single layer index (e.g., 5 means layer 5)
            - A 2-element list [start, end]: range of layers (e.g., [4, 20] means layers 4-19)
            - A 3-element list [start, end, repeats]: range repeated N times
            Examples:
            - [[0, 34, 1]]: all 34 layers once
            - [[0, 10], [10, 28, 2], [28, 34]]: layers 0-9, then 10-27 twice, then 28-33
        layer_types (`list`, *optional*):
            Attention pattern for each layer ("sliding_attention" or "full_attention").
        sliding_window (`int`, *optional*, defaults to 1024):
            Size of the sliding window for sliding attention layers.
        rope_theta (`float`, *optional*, defaults to 1000000.0):
            Base period for RoPE embeddings (global attention).
        rope_local_base_freq (`float`, *optional*, defaults to 10000.0):
            Base period for RoPE embeddings (local/sliding attention).
        query_pre_attn_scalar (`float`, *optional*, defaults to 256):
            Scaling factor for attention scores.
        rms_norm_eps (`float`, *optional*, defaults to 1e-6):
            Epsilon for RMS normalization.
        attention_bias (`bool`, *optional*, defaults to False):
            Whether to use bias in attention projections.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            Dropout ratio for attention.
        final_logit_softcapping (`float`, *optional*):
            Softcapping for final logits.
        attn_logit_softcapping (`float`, *optional*):
            Softcapping for attention logits.
        rope_scaling (`dict`, *optional*):
            RoPE scaling configuration.
        use_bidirectional_attention (`bool`, *optional*, defaults to False):
            If True, use bidirectional attention instead of causal.
    """

    model_type = "gemma3"
    keys_to_ignore_at_inference = ["past_key_values"]
    base_model_tp_plan = {
        "layers.*.self_attn.q_proj": "colwise",
        "layers.*.self_attn.k_proj": "colwise",
        "layers.*.self_attn.v_proj": "colwise",
        "layers.*.self_attn.o_proj": "rowwise",
        "layers.*.mlp.gate_proj": "colwise",
        "layers.*.mlp.up_proj": "colwise",
        "layers.*.mlp.down_proj": "rowwise",
    }
    base_model_pp_plan = {
        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
        "norm": (["hidden_states"], ["hidden_states"]),
    }

    def __init__(
        self,
        vocab_size=262_208,
        hidden_size=2560,
        intermediate_size=10240,
        num_hidden_layers=34,
        num_attention_heads=8,
        num_key_value_heads=4,
        head_dim=256,
        hidden_activation="gelu_pytorch_tanh",
        max_position_embeddings=131_072,
        initializer_range=0.02,
        rms_norm_eps=1e-6,
        use_cache=True,
        pad_token_id=0,
        eos_token_id=1,
        bos_token_id=2,
        tie_word_embeddings=True,
        rope_theta=1_000_000.0,
        attention_bias=False,
        attention_dropout=0.0,
        query_pre_attn_scalar=256,
        sliding_window=1024,
        layer_types=None,
        layer_sequence=None,
        final_logit_softcapping=None,
        attn_logit_softcapping=None,
        rope_scaling=None,
        rope_local_base_freq=10_000.0,
        use_bidirectional_attention=False,
        **kwargs,
    ):
        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.head_dim = head_dim
        self.num_key_value_heads = num_key_value_heads
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_bias = attention_bias
        self.attention_dropout = attention_dropout
        self.hidden_activation = hidden_activation
        self.query_pre_attn_scalar = query_pre_attn_scalar
        self.sliding_window = sliding_window
        self.final_logit_softcapping = final_logit_softcapping
        self.attn_logit_softcapping = attn_logit_softcapping
        self.use_bidirectional_attention = use_bidirectional_attention

        if use_bidirectional_attention:
            self.sliding_window = (self.sliding_window // 2) + 1

        self.rope_local_base_freq = rope_local_base_freq
        self.rope_scaling = rope_scaling
        rope_config_validation(self)

        # Layer sequence for looping - defaults to all layers once
        if layer_sequence is None:
            layer_sequence = [[0, num_hidden_layers, 1]]
        self.layer_sequence = layer_sequence

        # Layer types (sliding vs full attention)
        self._sliding_window_pattern = kwargs.get("sliding_window_pattern", 6)
        self.layer_types = layer_types
        if self.layer_types is None:
            self.layer_types = [
                "sliding_attention" if bool((i + 1) % self._sliding_window_pattern) else "full_attention"
                for i in range(self.num_hidden_layers)
            ]
        layer_type_validation(self.layer_types, self.num_hidden_layers)


__all__ = ["GemmagainConfig"]