Create configuration_deepseek.py

myr1/configuration_deepseek.py

@@ -0,0 +1,205 @@
+# configuration_deepseek.py
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+# This can remain empty if no pre-trained configs are being listed.
+DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+
+
+class DeepseekV3Config(PretrainedConfig):
+    r"""
+    Configuration class for the DeepSeek-V3 model architecture. Inherits from `PretrainedConfig`.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 129280):
+            Vocabulary size of the DeepSeek model.
+        hidden_size (`int`, *optional*, defaults to 7168):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 18432):
+            Dimension of the MLP representations in dense layers.
+        moe_intermediate_size (`int`, *optional*, defaults to 2048):
+            Dimension of the MLP representations used by MoE experts.
+        num_hidden_layers (`int`, *optional*, defaults to 61):
+            Number of hidden layers in the Transformer decoder.
+        num_nextn_predict_layers (`int`, *optional*, defaults to 1):
+            Number of "next-n predict" layers in the DeepSeekV3 model.
+        num_attention_heads (`int`, *optional*, defaults to 128):
+            Number of attention heads per attention layer.
+        num_key_value_heads (`int`, *optional*):
+            The number of key-value heads, used for GQA or MQA. Defaults to `num_attention_heads` if `None`.
+        n_shared_experts (`int`, *optional*, defaults to 1):
+            Number of shared experts. If None, indicates no shared experts (dense model).
+        n_routed_experts (`int`, *optional*, defaults to 256):
+            Number of routed experts. If None, indicates no routed experts (dense model).
+        ep_size (`int`, *optional*, defaults to 1):
+            The world-size used for expert parallelism. Typically set to the distributed world size if MoE is distributed.
+        routed_scaling_factor (`float`, *optional*, defaults to 2.5):
+            Scaling factor for routed experts' output weights.
+        kv_lora_rank (`int`, *optional*, defaults to 512):
+            The LoRA rank for Key and Value projections.
+        q_lora_rank (`int`, *optional*, defaults to 1536):
+            The LoRA rank for Query projections.
+        qk_rope_head_dim (`int`, *optional*, defaults to 64):
+            The dimension of the "rotary-embedded" portion of the Q/K heads.
+        v_head_dim (`int`, *optional*, defaults to 128):
+            The dimension of the Value heads.
+        qk_nope_head_dim (`int`, *optional*, defaults to 128):
+            The dimension of the Q/K heads that do *not* use rotary embeddings.
+        topk_method (`str`, *optional*, defaults to "noaux_tc"):
+            The gating TopK method in MoE (e.g. "noaux_tc", "topk_gating").
+        n_group (`int`, *optional*, defaults to 8):
+            Number of groups used in gating for MoE experts.
+        topk_group (`int`, *optional*, defaults to 4):
+            Number of selected groups for each token (MoE gating).
+        num_experts_per_tok (`int`, *optional*, defaults to 8):
+            Number of selected experts per token in the MoE gating.
+        moe_layer_freq (`int`, *optional*, defaults to 1):
+            Frequency of MoE layers among the transformer layers (1 = every layer is MoE).
+        first_k_dense_replace (`int`, *optional*, defaults to 3):
+            How many initial layers remain dense before MoE layers start appearing.
+        norm_topk_prob (`bool`, *optional*, defaults to True):
+            Whether to normalize the top-k gating probabilities.
+        scoring_func (`str`, *optional*, defaults to "sigmoid"):
+            The scoring function used for gating. For instance, "sigmoid" or "softmax".
+        aux_loss_alpha (`float`, *optional*, defaults to 0.001):
+            Scaling factor for any auxiliary MoE loss (e.g. load balancing).
+        seq_aux (`bool`, *optional*, defaults to True):
+            If True, auxiliary loss is computed per sample.
+        hidden_act (`str`, *optional*, defaults to "silu"):
+            Activation function used in MLP layers.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            Maximum sequence length the model can handle.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            Standard deviation of the truncated normal initializer.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-6):
+            Epsilon for RMS norm layers.
+        use_cache (`bool`, *optional*, defaults to True):
+            Whether the model returns `past_key_values`.
+        pad_token_id (`int`, *optional*):
+            Padding token id. If `None`, the model does not use a special padding token.
+        bos_token_id (`int`, *optional*, defaults to 0):
+            Beginning-of-sequence token id.
+        eos_token_id (`int`, *optional*, defaults to 1):
+            End-of-sequence token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Tensor parallelism rank used during pretraining for reproducibility.
+        tie_word_embeddings (`bool`, *optional*, defaults to False):
+            Whether to tie input and output embeddings.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            Base period for RoPE embeddings.
+        rope_scaling (`dict`, *optional*, defaults to None):
+            Dictionary for RoPE scaling parameters. (e.g. {"type":"yarn","factor":40,...})
+        attention_bias (`bool`, *optional*, defaults to False):
+            Whether to include bias terms in Q/K/V/out projections.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            Dropout probability for attention probabilities.
+        _attn_implementation (`str`, *optional*, defaults to "flash_attention_2"): # New: Attention Implementation Type
+            String indicating the attention implementation. Can be "eager", "flash_attention_2", or "sparse_attention" (if implemented).
+        **kwargs:
+            Additional arguments passed to `PretrainedConfig`.
+    """
+
+    model_type = "deepseek_v3"
+    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=None,  # Will be set to num_attention_heads if None
+        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,
+        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,
+        _attn_implementation="flash_attention_2", # New: Default to flash_attention_2
+        **kwargs,
+    ):
+        # Set defaults
+        self.vocab_size = vocab_size
+        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
+        # default num_key_value_heads = num_attention_heads if not specified
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_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.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        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._attn_implementation = _attn_implementation # New: set attention implementation type
+        self.pretraining_tp = pretraining_tp
+
+        # Pass everything to PretrainedConfig
+        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,
+        )