| """ | |
| DeepSeek model configuration | |
| """ | |
| from transformers.configuration_utils import PretrainedConfig | |
| from transformers.utils import logging | |
| logger = logging.get_logger(__name__) | |
| DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {} | |
| class DeepSeekConfig(PretrainedConfig): | |
| r""" | |
| This is the configuration class to store the configuration of a [`DeepSeekModel`]. It is used to instantiate a | |
| DeepSeek 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 DeepSeek-V3 | |
| [deepseek-ai/DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3) architecture. | |
| 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 50256): | |
| Vocabulary size of the DeepSeek model. Defines the number of different tokens that can be represented by the | |
| `inputs_ids` passed when calling [`DeepSeekModel`] | |
| hidden_size (`int`, *optional*, defaults to 1024): | |
| Dimension of the hidden representations. | |
| intermediate_size (`int`, *optional*, defaults to 4096): | |
| Dimension of the MLP representations for dense layers. | |
| moe_intermediate_size (`int`, *optional*, defaults to 704): | |
| Dimension of the MLP representations for MoE layers. | |
| num_hidden_layers (`int`, *optional*, defaults to 6): | |
| Number of hidden layers in the Transformer decoder. | |
| num_dense_layers (`int`, *optional*, defaults to 1): | |
| Number of dense (non-MoE) layers in the model. | |
| num_attention_heads (`int`, *optional*, defaults to 8): | |
| Number of attention heads for each attention layer in the Transformer decoder. | |
| num_routed_experts (`int`, *optional*, defaults to 4): | |
| Number of routed experts in MoE layers. | |
| num_shared_experts (`int`, *optional*, defaults to 2): | |
| Number of shared experts in MoE layers. | |
| num_activated_experts (`int`, *optional*, defaults to 2): | |
| Number of experts activated per token in MoE layers. | |
| num_expert_groups (`int`, *optional*, defaults to 1): | |
| Number of expert groups in MoE layers. | |
| num_limited_groups (`int`, *optional*, defaults to 1): | |
| Number of limited groups in MoE layers. | |
| score_func (`str`, *optional*, defaults to `"softmax"`): | |
| Scoring function for expert selection. Can be "softmax" or "sigmoid". | |
| route_scale (`float`, *optional*, defaults to 1.0): | |
| Scaling factor for routing weights. | |
| q_lora_rank (`int`, *optional*, defaults to 0): | |
| Rank of LoRA adaptation for query projection. 0 means no LoRA. | |
| kv_lora_rank (`int`, *optional*, defaults to 256): | |
| Rank of LoRA adaptation for key-value projection. | |
| qk_nope_head_dim (`int`, *optional*, defaults to 64): | |
| Dimension of query-key heads without positional encoding. | |
| qk_rope_head_dim (`int`, *optional*, defaults to 32): | |
| Dimension of query-key heads with rotary positional encoding. | |
| v_head_dim (`int`, *optional*, defaults to 64): | |
| Dimension of value heads. | |
| original_seq_len (`int`, *optional*, defaults to 512): | |
| Original sequence length used during pretraining. | |
| rope_theta (`float`, *optional*, defaults to 10000.0): | |
| Base frequency for rotary positional encoding. | |
| rope_factor (`float`, *optional*, defaults to 40): | |
| Scaling factor for RoPE frequency adjustment. | |
| beta_fast (`int`, *optional*, defaults to 32): | |
| Fast beta parameter for YaRN RoPE scaling. | |
| beta_slow (`int`, *optional*, defaults to 1): | |
| Slow beta parameter for YaRN RoPE scaling. | |
| mscale (`float`, *optional*, defaults to 1.0): | |
| Scale factor for attention logits when using extended context. | |
| max_position_embeddings (`int`, *optional*, defaults to 256): | |
| The maximum sequence length that this model might ever be used with. | |
| max_batch_size (`int`, *optional*, defaults to 2): | |
| The maximum batch size that this model might ever be used with for caching. | |
| 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-3): | |
| 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*): | |
| The id of the padding token. | |
| bos_token_id (`int`, *optional*, defaults to 2): | |
| The id of the "beginning-of-sequence" token. | |
| eos_token_id (`int`, *optional*, defaults to 3): | |
| The id of the "end-of-sequence" token. | |
| tie_word_embeddings (`bool`, *optional*, defaults to `False`): | |
| Whether to tie weight embeddings | |
| ```python | |
| >>> from transformers import DeepSeekModel, DeepSeekConfig | |
| >>> # Initializing a DeepSeek configuration | |
| >>> configuration = DeepSeekConfig() | |
| >>> # Initializing a model from the configuration | |
| >>> model = DeepSeekModel(configuration) | |
| >>> # Accessing the model configuration | |
| >>> configuration = model.config | |
| ```""" | |
| model_type = "deepseek" | |
| keys_to_ignore_at_inference = ["past_key_values"] | |
| def __init__( | |
| self, | |
| vocab_size=50256, | |
| hidden_size=1024, | |
| intermediate_size=4096, | |
| moe_intermediate_size=704, | |
| num_hidden_layers=6, | |
| num_dense_layers=1, | |
| num_attention_heads=8, | |
| num_routed_experts=4, | |
| num_shared_experts=2, | |
| num_activated_experts=2, | |
| num_expert_groups=1, | |
| num_limited_groups=1, | |
| score_func="softmax", | |
| route_scale=1.0, | |
| q_lora_rank=0, | |
| kv_lora_rank=256, | |
| qk_nope_head_dim=64, | |
| qk_rope_head_dim=32, | |
| v_head_dim=64, | |
| original_seq_len=512, | |
| rope_theta=10000.0, | |
| rope_factor=40, | |
| beta_fast=32, | |
| beta_slow=1, | |
| mscale=1.0, | |
| max_position_embeddings=256, | |
| max_batch_size=2, | |
| initializer_range=0.02, | |
| rms_norm_eps=1e-3, | |
| use_cache=True, | |
| pad_token_id=0, | |
| bos_token_id=2, | |
| eos_token_id=3, | |
| tie_word_embeddings=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_dense_layers = num_dense_layers | |
| self.num_attention_heads = num_attention_heads | |
| self.num_routed_experts = num_routed_experts | |
| self.num_shared_experts = num_shared_experts | |
| self.num_activated_experts = num_activated_experts | |
| self.num_expert_groups = num_expert_groups | |
| self.num_limited_groups = num_limited_groups | |
| self.score_func = score_func | |
| self.route_scale = route_scale | |
| self.q_lora_rank = q_lora_rank | |
| self.kv_lora_rank = kv_lora_rank | |
| self.qk_nope_head_dim = qk_nope_head_dim | |
| self.qk_rope_head_dim = qk_rope_head_dim | |
| self.v_head_dim = v_head_dim | |
| self.original_seq_len = original_seq_len | |
| self.rope_theta = rope_theta | |
| self.rope_factor = rope_factor | |
| self.beta_fast = beta_fast | |
| self.beta_slow = beta_slow | |
| self.mscale = mscale | |
| self.max_batch_size = max_batch_size | |
| self.initializer_range = initializer_range | |
| self.rms_norm_eps = rms_norm_eps | |
| self.use_cache = use_cache | |
| self.tie_word_embeddings = tie_word_embeddings | |
| 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, | |
| ) |