#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/granitemoe_swa/modular_granitemoe_swa.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_granitemoe_swa.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 IBM and the 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.
from huggingface_hub.dataclasses import strict

from ...configuration_utils import PreTrainedConfig
from ...modeling_rope_utils import RopeParameters
from ...utils import auto_docstring


@auto_docstring(checkpoint="ibm-granite/granite-swash-3b-a600m")
@strict
class GraniteMoeSWAConfig(PreTrainedConfig):
    r"""
    shared_intermediate_size (`int`, *optional*, defaults to 0):
        intermediate size for shared experts. Defaults to `0`, which disables the shared experts.
    sliding_window (`int`, *optional*, defaults to 128):
        Size of the sliding attention window used by layers whose `layer_types` entry is
        `"sliding_attention"`.
    layer_types (`list[str]`, *optional*):
        Per-layer attention type, each either `"full_attention"` or `"sliding_attention"`. When
        `None`, every fourth layer (`i % 4 == 0`) uses full attention and the rest use sliding
        window attention.
    layer_rope_theta (`list[float]`, *optional*):
        Per-layer RoPE base (`theta`) frequency. `0` sets NoPE (no positional embedding) for
        that layer. Overrides global `rope_parameters["rope_theta"]`, which is only used when
        this list is not provided or specified (`layer_rope_theta = None`).

    ```python
    >>> from transformers import GraniteMoeSWAModel, GraniteMoeSWAConfig

    >>> # Initializing a GraniteMoeSWA configuration
    >>> configuration = GraniteMoeSWAConfig()

    >>> # Initializing a model from the configuration
    >>> model = GraniteMoeSWAModel(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "granitemoe_swa"
    keys_to_ignore_at_inference = ["past_key_values"]

    vocab_size: int = 32000
    hidden_size: int = 4096
    intermediate_size: int = 11008
    num_hidden_layers: int = 32
    num_attention_heads: int = 32
    num_key_value_heads: int | None = None
    hidden_act: str = "silu"
    max_position_embeddings: int = 2048
    initializer_range: float = 0.02
    rms_norm_eps: float = 1e-6
    use_cache: bool = True
    pad_token_id: int | None = None
    bos_token_id: int | None = 1
    eos_token_id: int | list[int] | None = 2
    tie_word_embeddings: bool = False
    rope_parameters: RopeParameters | dict | None = None
    attention_bias: bool = False
    attention_dropout: float | int | None = 0.0
    embedding_multiplier: float | int | None = 1.0
    logits_scaling: float | int | None = 1.0
    residual_multiplier: float | int | None = 1.0
    attention_multiplier: float | int | None = 1.0
    num_local_experts: int | None = 8
    num_experts_per_tok: int | None = 2
    output_router_logits: bool | None = False
    router_aux_loss_coef: float | None = 0.001
    shared_intermediate_size: int = 0
    # Attention shards like Granite (+ per-head `sinks` colwise to track the head-sharding); the
    # routed experts shard tensor-parallel (packed gate/up colwise, down rowwise, `moe_tp_experts`)
    # with the router replicated. The optional shared expert (`shared_mlp`, off by default) is left
    # replicated -- it is small and its full output sums consistently with the all-reduced MoE output.
    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.*.self_attn.sinks": "colwise",
        "layers.*.block_sparse_moe.experts.gate_up_proj": "packed_colwise",
        "layers.*.block_sparse_moe.experts.down_proj": "rowwise",
        "layers.*.block_sparse_moe.experts": "moe_tp_experts",
    }
    # Expert-parallel plan: shard the routed experts across ranks (each rank owns a slice of the
    # experts) with the router driving the dispatch. The optional shared expert is left replicated.
    base_model_ep_plan = {
        "layers.*.block_sparse_moe.router": "ep_router",
        "layers.*.block_sparse_moe.experts.gate_up_proj": "grouped_gemm",
        "layers.*.block_sparse_moe.experts.down_proj": "grouped_gemm",
        "layers.*.block_sparse_moe.experts": "moe_tp_experts",
    }

    sliding_window: int | None = 128
    layer_types: list[str] | None = None
    layer_rope_theta: list[float | int] | None = None

    def __post_init__(self, **kwargs):
        if self.layer_types is None:
            self.layer_types = [
                "full_attention" if i % 4 == 0 else "sliding_attention" for i in range(self.num_hidden_layers)
            ]
        if self.num_key_value_heads is None:
            self.num_key_value_heads = self.num_attention_heads

        super().__post_init__(**kwargs)

        # Per-layer RoPE base theta (0 => NoPE). Default: global rope_theta.
        # Run after super post_init so that rope_theta is reliably set.
        if self.layer_rope_theta is None:
            self.layer_rope_theta = [self.rope_parameters["rope_theta"]] * self.num_hidden_layers


__all__ = ["GraniteMoeSWAConfig"]