configuration_rize.py

# Adapted from an upstream configuration file.

from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging

logger = logging.get_logger(__name__)

RIZE_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
class RizeConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`RizeModel`]. It is used to instantiate an Rize
    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
    defaults will yield a similar configuration to that of the Rize-V3.

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.


    Args:
        vocab_size (`int`, *optional*, defaults to 129280):
            Vocabulary size of the Rize model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`RizeModel`]
        hidden_size (`int`, *optional*, defaults to 4096):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 11008):
            Dimension of the MLP representations.
        moe_intermediate_size (`int`, *optional*, defaults to 1407):
            Dimension of the MoE representations.
        num_hidden_layers (`int`, *optional*, defaults to 32):
            Number of hidden layers in the Transformer decoder.
        num_nextn_predict_layers (`int`, *optional*, defaults to 1):
            Number of nextn predict layers in the RizeV3 Model.
        num_attention_heads (`int`, *optional*, defaults to 32):
            Number of attention heads for each attention layer in the Transformer decoder.
        n_shared_experts (`int`, *optional*, defaults to None):
            Number of shared experts, None means dense model.
        n_routed_experts (`int`, *optional*, defaults to None):
            Number of routed experts, None means dense model.
        routed_scaling_factor (`float`, *optional*, defaults to 1.0):
            Scaling factor or routed experts.
        topk_method (`str`, *optional*, defaults to `gready`):
            Topk method used in routed gate.
        n_group (`int`, *optional*, defaults to None):
            Number of groups for routed experts.
        topk_group (`int`, *optional*, defaults to None):
            Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).
        num_experts_per_tok (`int`, *optional*, defaults to None):
            Number of selected experts, None means dense model.
        moe_layer_freq (`int`, *optional*, defaults to 1):
            The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
        first_k_dense_replace (`int`, *optional*, defaults to 0):
            Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
                                                            \--k dense layers--/
        norm_topk_prob (`bool`, *optional*, defaults to False):
            Whether to normalize the weights of the routed experts.
        scoring_func (`str`, *optional*, defaults to 'softmax'):
            Method of computing expert weights.
        aux_loss_alpha (`float`, *optional*, defaults to 0.001):
            Auxiliary loss weight coefficient.
        seq_aux = (`bool`, *optional*, defaults to True):
            Whether to compute the auxiliary loss for each individual sample.
        num_key_value_heads (`int`, *optional*):
            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
            by meanpooling all the original heads within that group. For more details checkout [this
            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
            `num_attention_heads`.
        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
            The non-linear activation function (function or string) in the decoder.
        max_position_embeddings (`int`, *optional*, defaults to 2048):
            The maximum sequence length that this model might ever be used with.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
            The epsilon used by the rms normalization layers.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models). Only
            relevant if `config.is_decoder=True`.
        pad_token_id (`int`, *optional*):
            Padding token id.
        bos_token_id (`int`, *optional*, defaults to 1):
            Beginning of stream token id.
        eos_token_id (`int`, *optional*, defaults to 2):
            End of stream token id.
        pretraining_tp (`int`, *optional*, defaults to 1):
            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
            issue](https://github.com/pytorch/pytorch/issues/76232).
        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
            Whether to tie weight embeddings
        rope_theta (`float`, *optional*, defaults to 10000.0):
            The base period of the RoPE embeddings.
        rope_scaling (`Dict`, *optional*):
            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
            `max_position_embeddings` to the expected new maximum.
        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
            Whether to use a bias in the query, key, value and output projection layers during self-attention.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        use_linear_ce (`bool`, *optional*, defaults to `True`):
            Whether to use cut-cross-entropy (linear CE) for causal LM loss computation when labels are provided.
        linear_ce_impl (`str`, *optional*, defaults to `"cce_exact"`):
            Implementation name passed to cut-cross-entropy when `use_linear_ce=True`.

    ```python
    >>> from transformers import RizeModel, RizeConfig

    >>> # Initializing a Rize-V3 style configuration
    >>> configuration = RizeConfig()

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

    model_type = "rize"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=129280,
        hidden_size=7168,
        intermediate_size=18432,
        moe_intermediate_size = 2048,
        num_hidden_layers=61,
        num_nextn_predict_layers=1,
        num_attention_heads=128,
        num_key_value_heads=128,
        n_shared_experts = 1,
        n_routed_experts = 256,
        ep_size = 1,
        routed_scaling_factor = 2.5,
        kv_lora_rank = 512,
        q_lora_rank = 1536,
        qk_rope_head_dim = 64,
        v_head_dim = 128,
        qk_nope_head_dim = 128,
        topk_method = 'noaux_tc',
        n_group = 8,
        topk_group = 4,
        num_experts_per_tok = 8,
        moe_layer_freq = 1,
        first_k_dense_replace = 3,
        norm_topk_prob = True,
        scoring_func = 'sigmoid',
        aux_loss_alpha = 0.001,
        seq_aux = True,
        auxfree_bias_lr = 0.0,
        hidden_act="silu",
        max_position_embeddings=4096,
        initializer_range=0.02,
        rms_norm_eps=1e-6,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=0,
        eos_token_id=1,
        pretraining_tp=1,
        tie_word_embeddings=False,
        rope_theta=10000.0,
        rope_scaling=None,
        attention_bias=False,
        attention_dropout=0.0,
        use_linear_ce=True,
        linear_ce_impl="cce_exact",
        attention_arch="auto",
        linear_attn_config=None,
        block_diag_causal_mask=False,
        reset_position_ids_per_sample=False,
        moe_router_active_only=True,
        prompt_loss_weight=0.0,
        freeze_router_on_sft=False,
        global_lbl_enabled=False,
        global_lbl_sync_across_ranks=False,
        global_lbl_buffer_across_ga=False,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.moe_intermediate_size = moe_intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_nextn_predict_layers = num_nextn_predict_layers
        self.num_attention_heads = num_attention_heads
        self.n_shared_experts = n_shared_experts
        self.n_routed_experts = n_routed_experts
        self.ep_size = ep_size
        self.routed_scaling_factor = routed_scaling_factor
        self.kv_lora_rank = kv_lora_rank
        self.q_lora_rank = q_lora_rank
        self.qk_rope_head_dim = qk_rope_head_dim
        self.v_head_dim = v_head_dim
        self.qk_nope_head_dim = qk_nope_head_dim
        self.topk_method = topk_method
        self.n_group = n_group
        self.topk_group = topk_group
        self.num_experts_per_tok = num_experts_per_tok
        self.moe_layer_freq = moe_layer_freq
        self.first_k_dense_replace = first_k_dense_replace
        self.norm_topk_prob = norm_topk_prob
        self.scoring_func = scoring_func
        self.aux_loss_alpha = aux_loss_alpha
        self.seq_aux = seq_aux
        self.auxfree_bias_lr = auxfree_bias_lr
        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads

        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.pretraining_tp = pretraining_tp
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.rope_scaling = rope_scaling
        self.attention_bias = attention_bias
        self.attention_dropout = attention_dropout
        self.use_linear_ce = use_linear_ce
        self.linear_ce_impl = linear_ce_impl
        self.block_diag_causal_mask = bool(block_diag_causal_mask)
        self.reset_position_ids_per_sample = bool(reset_position_ids_per_sample)
        self.moe_router_active_only = bool(moe_router_active_only)
        self.prompt_loss_weight = float(prompt_loss_weight)
        self.freeze_router_on_sft = bool(freeze_router_on_sft)
        self.global_lbl_enabled = bool(global_lbl_enabled)
        self.global_lbl_sync_across_ranks = bool(global_lbl_sync_across_ranks)
        self.global_lbl_buffer_across_ga = bool(global_lbl_buffer_across_ga)

        # ---- Kimi-Linear / hybrid attention knobs (optional) ----
        # attention_arch: 'auto' (default; infer from linear_attn_config), 'standard', or 'kimi_linear'
        self.attention_arch = attention_arch
        # linear_attn_config: dict or None
        self.linear_attn_config = linear_attn_config

        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,
        )

    @property
    def is_linear_attn(self) -> bool:
        """Whether this config enables the Kimi-Linear style hybrid attention.

        Notes:
          - This is intentionally a *property* (not a stored boolean) so that it works
            for both new configs (attention_arch) and older checkpoints that might only
            carry linear_attn_config.
        """
        arch = getattr(self, "attention_arch", None)
        la_cfg = getattr(self, "linear_attn_config", None)
        if arch is None:
            return la_cfg is not None
        # allow direct boolean override (e.g. programmatic configs)
        if isinstance(arch, bool):
            return bool(arch) and la_cfg is not None

        arch_s = str(arch).lower().strip()
        if arch_s in ("auto", "infer"):
            return la_cfg is not None
        if arch_s in ("standard", "full", "mla", "default", "none", ""):
            return False
        if arch_s in ("kimi_linear", "kda", "linear_attn", "linear", "hybrid"):
            return la_cfg is not None

        # Unknown value: be conservative (disable)
        return False

    @is_linear_attn.setter
    def is_linear_attn(self, value: bool) -> None:
        """Back-compat setter. Allows `config.is_linear_attn = True/False`."""
        if bool(value):
            # Only flip attention_arch if it is currently unset or standard-ish.
            cur = getattr(self, "attention_arch", "standard")
            if cur is None or str(cur).lower().strip() in ("standard", "full", "default", "none", ""):
                self.attention_arch = "kimi_linear"
        else:
            self.attention_arch = "standard"

    def is_kda_layer(self, layer_idx: int) -> bool:
        """Return True if `layer_idx` (0-indexed) should use KDA (linear attention).

        Layer index convention:
          - Internally we expect `layer_idx` to be 0-indexed (as used by `range(num_hidden_layers)`).
          - In config.linear_attn_config, `kda_layers` / `full_attn_layers` may be either:
              * 0-indexed (0..num_hidden_layers-1), OR
              * 1-indexed (1..num_hidden_layers) like the official Kimi-Linear configs.
            We auto-detect the convention.
        """
        if not self.is_linear_attn:
            return False

        cfg = getattr(self, "linear_attn_config", None)
        if not isinstance(cfg, dict):
            return False

        if layer_idx is None:
            return False
        try:
            layer_idx = int(layer_idx)
        except Exception:
            return False
        if layer_idx < 0 or layer_idx >= int(getattr(self, "num_hidden_layers", 0) or 0):
            return False

        kda_layers = cfg.get("kda_layers", None)
        full_layers = cfg.get("full_attn_layers", None)

        # Normalize lists (best-effort)
        kda = [int(x) for x in (kda_layers or [])] if isinstance(kda_layers, (list, tuple)) else []
        full = [int(x) for x in (full_layers or [])] if isinstance(full_layers, (list, tuple)) else []

        # Auto-detect index base for the config lists
        all_idx = kda + full
        one_indexed = False
        if all_idx:
            n = int(getattr(self, "num_hidden_layers", 0) or 0)
            # Strong signals:
            if 0 in all_idx:
                one_indexed = False
            elif n in all_idx:
                one_indexed = True
            else:
                # Heuristic: if everything is within [1, n], treat as 1-indexed
                mn, mx = min(all_idx), max(all_idx)
                if mn >= 1 and mx <= n:
                    one_indexed = True

        query_idx = layer_idx + (1 if one_indexed else 0)

        if kda:
            return query_idx in set(kda)
        if full:
            return query_idx not in set(full)

        # If neither list is provided, default to "no KDA"
        return False