configuration_mobilellm_p1.py

from transformers.configuration_utils import PretrainedConfig

class MobileLLMP1TextConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`MobileLLMP1TextModel`]. It is used to instantiate a
    MobileLLMP1 text 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 MobileLLMP1 1B model.

    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 202048):
            Vocabulary size of the Llama4 text model. Defines the maximum number of different tokens that can be represented
            by the `inputs_ids` passed when calling [`Llama4TextModel`].
        hidden_size (`int`, *optional*, defaults to 5120):
            Dimensionality of the embeddings and hidden states.
        intermediate_size (`int`, *optional*, defaults to 8192):
            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
        intermediate_size_mlp (`int`, *optional*, defaults to 16384): TODO
        num_hidden_layers (`int`, *optional*, defaults to 48):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 40):
            Number of attention heads for each attention layer in the Transformer encoder.
        num_key_value_heads (`int`, *optional*, defaults to 8):
            This is the number of key_value heads that should be used to implement Grouped Query Attention. If not
            specified, will default to `num_attention_heads`.
        head_dim (`int`, *optional*, defaults to 128): TODO
        hidden_act (`str` or `Callable`, *optional*, defaults to `"silu"`):
            The non-linear activation function (function or string) in the encoder and pooler.
        max_position_embeddings (`int`, *optional*, defaults to 131072):
            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.
        sliding_window (`int`, *optional*, defaults to 512):
            In MobileLLMP1, every 4 out of 5 layers use sliding window attention. This is the size of the sliding window.
        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
            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.
        pad_token_id (`int`, *optional*, defaults to 128004):
            The id of the padding token.
        bos_token_id (`int`, *optional*, defaults to 1):
            The id of the beginning of sentence token.
        eos_token_id (`int`, *optional*, defaults to 2):
            The id of the end of sentence token.
        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
            Whether to tie weight embeddings
        rope_theta (`float`, *optional*, defaults to `500000.0`):
            The base period of the RoPE embeddings.
        attention_dropout (`int`, *optional*, defaults to 0.0): TODO
        num_experts_per_tok (`int`, *optional*, defaults to 1): TODO
        num_local_experts (`int`, *optional*, defaults to 16): TODO
        moe_layers (`int`, *optional*): TODO
        interleave_moe_layer_step (`int`, *optional*, defaults to 1): TODO
        use_qk_norm (`int`, *optional*, defaults to `True`): TODO
        output_router_logits (`int`, *optional*, defaults to `False`): TODO
        router_aux_loss_coef (`int`, *optional*, defaults to 0.001): TODO
        router_jitter_noise (`int`, *optional*, defaults to 0.0): TODO
        rope_scaling (`Dict`, *optional*):
            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
            accordingly.
            Expected contents:
                `rope_type` (`str`):
                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
                    'llama3'], with 'default' being the original RoPE implementation.
                `factor` (`float`, *optional*):
                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
                    original maximum pre-trained length.
                `original_max_position_embeddings` (`int`, *optional*):
                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
                    pretraining.
                `attention_factor` (`float`, *optional*):
                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
                    computation. If unspecified, it defaults to value recommended by the implementation, using the
                    `factor` field to infer the suggested value.
                `beta_fast` (`float`, *optional*):
                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
                    ramp function. If unspecified, it defaults to 32.
                `beta_slow` (`float`, *optional*):
                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
                    ramp function. If unspecified, it defaults to 1.
                `short_factor` (`list[float]`, *optional*):
                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                    size divided by the number of attention heads divided by 2
                `long_factor` (`list[float]`, *optional*):
                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                    size divided by the number of attention heads divided by 2
                `low_freq_factor` (`float`, *optional*):
                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
                `high_freq_factor` (`float`, *optional*):
                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
            <TODO>
            <TODO>
        no_rope_layers (`list[int]`, *optional*):
            List with at least the same length as the number of layers in the model.
            A `1` at an index position indicates that the corresponding layer will use RoPE,
            while a `0` indicates that it's a NoPE layer.
        no_rope_layer_interval (`int`, *optional*, defaults to 4):
            If `no_rope_layers` is `None`, it will be created using a NoPE layer every
            `no_rope_layer_interval` layers.
        attention_chunk_size (`int`, *optional*, defaults to 8192):
            <TODO>
        layer_types (`list`, *optional*):
            Attention pattern for each layer.
        attn_temperature_tuning (`bool`, *optional*, defaults to `True`):
            Whether to dynamically scale the attention temperature for each query token based on sequence length.
            Recommended for long sequences (e.g., >32k tokens) to maintain stable output results.
        floor_scale (`int`, *optional*, defaults to 8192): TODO
        attn_scale (`int`, *optional*, defaults to 0.1): TODO

    Example:
    """

    model_type = "llama4_text"
    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.*.feed_forward.shared_expert.gate_proj": "local_colwise",
        "layers.*.feed_forward.shared_expert.up_proj": "local_colwise",
        "layers.*.feed_forward.shared_expert.down_proj": "local_rowwise",
        "layers.*.feed_forward.experts.gate_up_proj": "local_packed_rowwise",  # row because not linear
        "layers.*.feed_forward.experts.down_proj": "local_colwise",  # col because not linear
        "layers.*.feed_forward.experts": "local",
        "layers.*.feed_forward.gate_proj": "local_colwise",
        "layers.*.feed_forward.up_proj": "local_colwise",
        "layers.*.feed_forward.down_proj": "local_rowwise",
        "layers.*.feed_forward": "gather",
    }
    base_model_ep_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.*.feed_forward.experts.gate_up_proj": "grouped_gemm",  # row because not linear
        "layers.*.feed_forward.experts.down_proj": "grouped_gemm",  # col because not linear
        "layers.*.feed_forward.experts": "gather",  # all reduce
        "layers.*.feed_forward.gate_proj": "local_colwise",
        "layers.*.feed_forward.up_proj": "local_colwise",
        "layers.*.feed_forward.down_proj": "local_rowwise",
        "layers.*.feed_forward.router": "ep_router",
    }

    def __init__(
        self,
        vocab_size=202048,
        hidden_size=1280,
        intermediate_size=6144,
        intermediate_size_mlp=6144,
        num_hidden_layers=30,
        num_attention_heads=20,
        num_key_value_heads=4,
        head_dim=64,
        hidden_act="silu",
        max_position_embeddings=131072,
        initializer_range=0.02,
        rms_norm_eps=1e-5,
        use_cache=True,
        pad_token_id=None,
        sliding_window=512,
        bos_token_id=1,
        eos_token_id=2,
        tie_word_embeddings=True,
        rope_theta=500000,
        attention_dropout=0.0,
        num_experts_per_tok=1,
        num_local_experts=16,
        moe_layers=None,
        interleave_moe_layer_step=1,
        use_qk_norm=False,
        output_router_logits=False,
        router_aux_loss_coef=0.001,
        router_jitter_noise=0.0,
        rope_scaling=None,
        no_rope_layers=None,
        no_rope_layer_interval=4,
        attention_chunk_size=8192,
        layer_types=None,
        attn_temperature_tuning=True,
        floor_scale=8192,
        attn_scale=0.1,
        **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.attn_temperature_tuning = attn_temperature_tuning
        self.attn_scale = attn_scale
        self.floor_scale = floor_scale
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.intermediate_size_mlp = intermediate_size_mlp
        self.num_hidden_layers = num_hidden_layers
        self.sliding_window = sliding_window
        self.num_attention_heads = num_attention_heads
        self.rope_scaling = rope_scaling
        self.attention_bias = False
        # 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.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout
        self.head_dim = (
            head_dim
            if head_dim is not None
            else self.hidden_size // self.num_attention_heads
        )
        self.use_qk_norm = use_qk_norm

        self.num_experts_per_tok = num_experts_per_tok
        self.num_local_experts = num_local_experts

        self.output_router_logits = output_router_logits
        self.router_aux_loss_coef = router_aux_loss_coef
        self.router_jitter_noise = router_jitter_noise
        self.layer_types = layer_types

        # Backwards compatibility
        if no_rope_layers == []:
            no_rope_layers = None

        default_no_rope_layers = [
            int((layer_idx + 1) % no_rope_layer_interval != 0)
            for layer_idx in range(self.num_hidden_layers)
        ]

        self.no_rope_layers = (
            no_rope_layers if no_rope_layers else default_no_rope_layers
        )

        # If no pattern set, use our default pattern
        if self.layer_types is None:
            self.layer_types = [
                "sliding_attention" if bool((i) % 4) else "full_attention"
                for i in range(self.num_hidden_layers)
            ] + [
                "full_attention"
            ]  # Last layer is always full attention

        self.interleave_moe_layer_step = interleave_moe_layer_step
        self.moe_layers = (
            moe_layers
            if moe_layers is not None
            else list(
                range(
                    interleave_moe_layer_step - 1,
                    num_hidden_layers,
                    interleave_moe_layer_step,
                )
            )
        )