configuration_blockffn.py

# coding=utf-8
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# 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.
"""BlockFFN model configuration"""

from transformers import PretrainedConfig
from transformers.modeling_rope_utils import rope_config_validation


class BlockFFNConfig(PretrainedConfig):

    model_type = "blockffn"
    keys_to_ignore_at_inference = ["past_key_values"]
    # Default tensor parallel plan for base model `BlockFFNModel`
    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=32000,
        hidden_size=4096,
        ffn_hidden_size=11008,
        num_layers=32,
        num_attention_heads=32,
        num_query_groups=None,
        hidden_act="silu",
        max_position_embeddings=2048,
        initializer_range=0.02,
        norm_epsilon=1e-6,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=1,
        eos_token_id=2,
        pretraining_tp=1,
        tie_word_embeddings=False,
        rope_theta=10000.0,
        rope_scaling=None,
        attention_bias=False,
        attention_dropout=0.0,
        mlp_bias=False,
        head_dim=None,
        use_mup=True,
        mup_emb_scale=12,
        mup_depth_scale=1.4,
        mup_base_hidden_size=256,
        num_experts=180,
        moe_ffn_hidden_size=128,
        moe_shared_expert_intermediate_size=128,
        moe_layer_freq="([0]*3+[1]*29)",
        moe_router_dtype="fp32",
        router_act_func="relu",
        router_norm_type="simple",
        router_norm_fixed=False,
        router_norm_scalar=False,
        router_norm_init_var=0.1,
        expert_act_func="norm_silu",
        expert_act_norm_type="normal",
        use_blockffn=False,
        router_type="topk",
        moe_router_topk=0,
        moe_router_topp=0,
        moe_router_enable_expert_bias=False,
        moe_router_score_function="sigmoid",
        moe_router_topk_scaling_factor=2.5,
        expert_not_gated=False,
        moe_router_pre_softmax=False,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.ffn_hidden_size = ffn_hidden_size
        self.num_layers = num_layers
        self.num_attention_heads = num_attention_heads

        # for backward compatibility
        if num_query_groups is None:
            num_query_groups = num_attention_heads

        self.num_query_groups = num_query_groups
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.norm_epsilon = norm_epsilon
        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.mlp_bias = mlp_bias
        self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads
        self.use_mup = use_mup
        self.mup_emb_scale = mup_emb_scale
        self.mup_depth_scale = mup_depth_scale
        self.mup_base_hidden_size = mup_base_hidden_size

        self.num_experts = num_experts
        self.moe_ffn_hidden_size = moe_ffn_hidden_size
        self.moe_shared_expert_intermediate_size = moe_shared_expert_intermediate_size
        self.moe_layer_freq = moe_layer_freq if isinstance(moe_layer_freq, (str, list)) else ([0] * num_layers)
        self.moe_router_dtype = moe_router_dtype
        self.router_act_func = router_act_func
        self.router_norm_type = router_norm_type
        self.router_norm_fixed = router_norm_fixed
        self.router_norm_scalar = router_norm_scalar
        self.router_norm_init_var = router_norm_init_var
        self.expert_act_func = expert_act_func
        self.expert_act_norm_type = expert_act_norm_type

        self.use_blockffn = use_blockffn
        self.router_type = router_type
        self.moe_router_topk = moe_router_topk
        self.moe_router_topp = moe_router_topp
        self.moe_router_enable_expert_bias = moe_router_enable_expert_bias
        self.moe_router_score_function = moe_router_score_function
        self.moe_router_topk_scaling_factor = moe_router_topk_scaling_factor
        self.expert_not_gated = expert_not_gated
        self.moe_router_pre_softmax = moe_router_pre_softmax

        # Validate the correctness of rotary position embeddings parameters
        # BC: if there is a 'type' field, copy it it to 'rope_type'.
        if self.rope_scaling is not None and "type" in self.rope_scaling:
            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
        rope_config_validation(self)

        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 mup_width_scale(self):
        return (self.hidden_size / self.mup_base_hidden_size) if (self.use_mup and self.mup_base_hidden_size > 0) else 1


__all__ = ["BlockFFNConfig"]