"""DeepSeekV3 Config API."""
from typing import Optional, Union
from mindspore._checkparam import args_type_check
from mindformers.modules.transformer.moe import MoEConfig
from mindformers.modules.transformer.transformer import default_transformer_config, \
TransformerOpParallelConfig, default_moe_config
from mindformers.models.utils import convert_mstype
from mindformers.tools.register import MindFormerRegister, MindFormerModuleType
from mindformers.mindformer_book import MindFormerBook
from research.deepseek3.deepseek2_config import DeepseekV2Config
__all__ = ['DeepseekV3Config']
@MindFormerRegister.register(MindFormerModuleType.CONFIG)
class DeepseekV3Config(DeepseekV2Config):
"""DeepSeekV3 config class which defines the model size. It redefines the attention block and experts in moe blocks.
NOTE: If any nuisances, refer to in-between llama_config and the transformer original deepseek configuration.
Args:
batch_size (int, optional): Batch size for input data, use in predict. Default: ``1``.
seq_length (int, optional): The sequence length of input_ids. Default: ``2048``.
hidden_size (int, optional): Dimensionality of the encoder layers and the pooler layer. Default: ``4096``.
num_layers (int, optional): Number of hidden layers in the Transformer decoder. Default: ``32``.
num_heads (int, optional): Number of attention heads for each attention layer in the Transformer decoder.
Default: ``32``.
n_kv_heads (int, optional): Define multi group head attention heads number. Default: ``None``.
max_position_embedding (int, optional): Customize the maximum sequence length that the model can handle.
Default: ``32768``.
intermediate_size (int, optional): Customize the number of dimension of the intermediate layer.
Default: ``None``.
multiple_of (Optional[int]): Define SwiGLU hidden layer size multiples, default 256.
ffn_dim_multiplier (Optional[int]): Define ffn layer dim multiples, default None.
kv_lora_rank (int, optional): kv_lora_rank for Multi-Latent-Attention. Default: ``512``.
q_lora_rank (int, optional): q_lora_rank for Multi-Latent-Attention. Default: ``1536``.
qk_rope_head_dim (int, optional): qk_rope_head_dim for Multi-Latent-Attention. Default: ``64``.
v_head_dim (int, optional): v_head_dim for Multi-Latent-Attention. Default: ``128``.
qk_nope_head_dim (int, optional): qk_nope_head_dim for Multi-Latent-Attention. Default: ``128``.
vocab_size (int, optional): Vocabulary size of the llama model. Default: ``32000``.
rms_norm_eps (float, optional): The epsilon value of the denominator. Default: ``1e-5``.
bos_token_id (int, optional): The id of the *beginning-of-sequence* token. Default: ``1``.
eos_token_id (int, optional): The id of the *end-of-sequence* token. Default: ``2``.
pad_token_id (int, optional): The id of the *padding* token. Default: ``0``.
ignore_token_id (int, optional): The id of the *ignoring* token. Default: ``-100``.
theta (float, optional): Frequency factors for sine and cosine functions in RoPE. Default: ``10000.0``.
compute_dtype (str, optional): Linear layer compute dtype. Default: ``float16``.
layernorm_compute_type (str, optional): Layernorm compute dtype. Default: ``float32``.
softmax_compute_type (str, optional): Softmax compute dtype. Default: ``float32``.
rotary_dtype (str, optional): RoPE compute dtype. Default: ``float32``.
param_init_type (str, optional): Parameter initial dtype. Default: ``float16``.
qkv_has_bias (bool, optional): Whether the Query, Key, and Value projection has bias. Default: ``False``.
parallel_config (Union[dict, TransformerOpParallelConfig], optional): The parallel configuration.
Default: ``default_transformer_config`` , an instance of `TransformerOpParallelConfig` with default args.
moe_config (Union[dict, MoEConfig], optional): The MoE configuration. Default: ``default_moe_config`` ,
an instance of `MoEConfig` with default args.
use_past (bool, optional): Whether the model should use the past last key/values attentions
(if applicable to the model) to speed up decoding. Default: ``False``.
extend_method (str, optional): The extent method of seq length in inference. Default: ``None``.
scaling_factor (float, optional): Scaling factor to adjust the weights of the frequency factors in the sine
and cosine functions. Default: ``1.0``.
is_dynamic (bool, optional): Whether to use dynamic shape. Default: ``False``.
use_flash_attention (bool, optional): Whether to enable flash attention ops. Default: ``False``.
fine_grain_interleave (int, optional): Set the number of fine-grained interleave. Default: ``1``.
pp_interleave_num (int, optional): Set the number of pipeline interleave. Default: ``1``.
offset (int, optional): Offset of transformer layer when set pipeline stage number. Default: ``0``.
checkpoint_name_or_path (str, optional): checkpoint path or name used to load to the network. Default: ``None``.
repetition_penalty (float, optional): The parameter for repetition penalty. 1.0 means no penalty.
See `this paper <https://arxiv.org/pdf/1909.05858.pdf>`_ for more details. Default: ``1.0``.
max_decode_length (int, optional): The maximum length the generated tokens can have. Corresponds to the
length of the input prompt + `max_new_tokens`. Its effect is overridden by `max_new_tokens`, if also set.
Default: ``1024``.
block_size (int, optional): The maximum number of tokens in one block can have when using paged attention.
Default: ``16``.
num_blocks (int, optional): The maximum number of blocks when using paged attention. Default: ``512``.
top_k (int, optional): The number of highest probability vocabulary tokens to keep for top-k-filtering.
Default: ``5``.
top_p (float, optional): If set to float < 1, only the smallest set of most probable tokens with probabilities
that add up to `top_p` or higher are kept for generation. Default: ``1.0``.
do_sample (bool, optional): Whether to use sampling; use greedy decoding otherwise. Default: ``True``.
mtp_depth (`int`): The depth for multi-token prediction.
mtp_loss_factor (`float`): The loss factor for multi-token prediction.
return_extra_loss (`bool`): Whether to use extra loss for moe modules.
Returns:
Class, DeepseekV3Config.
"""
model_type = "deepseekv3"
_support_list = MindFormerBook.get_config_support_list()['deepseekv3']
@args_type_check(parallel_config=(dict, TransformerOpParallelConfig))
def __init__(self,
batch_size: int = 1,
seq_length: int = 2048,
hidden_size: int = 4096,
num_layers: int = 32,
num_heads: int = 32,
n_kv_heads: Optional[int] = None,
max_position_embeddings: int = 32768,
intermediate_size: int = 18432,
multiple_of: int = 256,
ffn_dim_multiplier: Optional[int] = None,
kv_lora_rank: int = 512,
q_lora_rank: int = 1536,
qk_rope_head_dim: int = 64,
v_head_dim: int = 128,
qk_nope_head_dim: int = 128,
vocab_size: int = 129280,
rms_norm_eps: float = 1e-6,
bos_token_id: int = 1,
eos_token_id: int = 2,
pad_token_id: int = 0,
ignore_token_id: int = -100,
theta: float = 10000.0,
compute_dtype: str = "bfloat16",
layernorm_compute_type: str = "float32",
softmax_compute_type: str = "float32",
rotary_dtype: str = "float32",
param_init_type: str = "bfloat16",
init_method_std=0.006,
qkv_has_bias=False,
qkv_concat=False,
ffn_concat=False,
parallel_config: Union[dict, TransformerOpParallelConfig] = default_transformer_config,
moe_config: Union[dict, MoEConfig] = default_moe_config,
use_past: bool = False,
extend_method: str = "YARN",
scaling_factor: dict = None,
is_dynamic: bool = False,
use_flash_attention: bool = False,
fine_grain_interleave: int = 1,
pp_interleave_num: int = 1,
offset: int = 0,
checkpoint_name_or_path="",
repetition_penalty: float = 1.0,
max_decode_length: int = 1024,
block_size: int = 16,
num_blocks: int = 512,
top_k: int = 5,
top_p: float = 1.0,
do_sample: bool = True,
mtp_depth: int = 0,
mtp_loss_factor: float = 0.3,
return_extra_loss: bool = True,
input_sliced_sig: bool = False,
use_fused_rope=False,
use_fused_swiglu=False,
enable_fa_var_len=False,
return_hidden_states=False,
**kwargs):
super(DeepseekV3Config, self).__init__(**kwargs)
if isinstance(parallel_config, dict):
parallel_config = TransformerOpParallelConfig(**parallel_config)
if isinstance(moe_config, dict):
moe_config = MoEConfig(**moe_config)
self.batch_size = batch_size
self.seq_length = seq_length
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_layers = num_layers
self.num_heads = num_heads
self.max_position_embeddings = max_position_embeddings if max_position_embeddings else seq_length
self.intermediate_size = intermediate_size
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.n_kv_heads = n_kv_heads
self.multiple_of = multiple_of
self.ffn_dim_multiplier = ffn_dim_multiplier
self.rms_norm_eps = rms_norm_eps
self.qkv_concat = qkv_concat
self.ffn_concat = ffn_concat
self.param_init_type = convert_mstype(param_init_type)
self.qkv_has_bias = qkv_has_bias
self.layernorm_compute_type = convert_mstype(layernorm_compute_type)
self.softmax_compute_type = convert_mstype(softmax_compute_type)
self.rotary_dtype = convert_mstype(rotary_dtype)
self.compute_dtype = convert_mstype(compute_dtype)
self.parallel_config = parallel_config
self.moe_config = moe_config
self.checkpoint_name_or_path = checkpoint_name_or_path
self.bos_token_id = bos_token_id
self.eos_token_id = eos_token_id
self.pad_token_id = pad_token_id
self.ignore_token_id = ignore_token_id
self.use_past = use_past
self.extend_method = extend_method
self.scaling_factor = scaling_factor
self.is_dynamic = is_dynamic
self.use_flash_attention = use_flash_attention
self.fine_grain_interleave = fine_grain_interleave
self.pp_interleave_num = pp_interleave_num
self.offset = offset
self.repetition_penalty = repetition_penalty
self.max_decode_length = max_decode_length
self.top_k = top_k
self.top_p = top_p
self.do_sample = do_sample
self.theta = theta
self.block_size = block_size
self.num_blocks = num_blocks
self.return_extra_loss = return_extra_loss
self.mtp_depth = mtp_depth
self.mtp_loss_factor = mtp_loss_factor
self.input_sliced_sig = input_sliced_sig
self.mlp_has_gate = True
self.hidden_act = "silu"
self.init_method_std = init_method_std
self.use_fused_rope = use_fused_rope
self.use_fused_swiglu = use_fused_swiglu
self.enable_fa_var_len = enable_fa_var_len
self.return_hidden_states = return_hidden_states