Create configuration_helion.py

configuration_helion.py

@@ -0,0 +1,103 @@
+"""Helion model configuration."""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class HelionConfig(PretrainedConfig):
+    """
+    Configuration class for Helion model.
+    
+    Args:
+        vocab_size (int, optional): Vocabulary size. Defaults to 32768.
+        hidden_size (int, optional): Dimensionality of hidden layers. Defaults to 4096.
+        intermediate_size (int, optional): Dimensionality of MLP. Defaults to 14336.
+        num_hidden_layers (int, optional): Number of decoder layers. Defaults to 32.
+        num_attention_heads (int, optional): Number of attention heads. Defaults to 32.
+        num_key_value_heads (int, optional): Number of key-value heads for GQA. Defaults to 8.
+        hidden_act (str, optional): Activation function. Defaults to "silu".
+        max_position_embeddings (int, optional): Maximum sequence length. Defaults to 8192.
+        initializer_range (float, optional): Standard deviation for weight initialization. Defaults to 0.02.
+        rms_norm_eps (float, optional): Epsilon for RMS normalization. Defaults to 1e-6.
+        use_cache (bool, optional): Whether to use KV cache. Defaults to True.
+        pad_token_id (int, optional): Padding token ID. Defaults to None.
+        bos_token_id (int, optional): Beginning of sequence token ID. Defaults to 1.
+        eos_token_id (int, optional): End of sequence token ID. Defaults to 2.
+        tie_word_embeddings (bool, optional): Tie input/output embeddings. Defaults to False.
+        rope_theta (float, optional): Base for RoPE. Defaults to 10000.0.
+        rope_scaling (dict, optional): RoPE scaling config. Defaults to None.
+        attention_bias (bool, optional): Add bias to attention projections. Defaults to False.
+        attention_dropout (float, optional): Dropout for attention. Defaults to 0.0.
+        mlp_bias (bool, optional): Add bias to MLP. Defaults to False.
+    """
+    
+    model_type = "helion"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    
+    def __init__(
+        self,
+        vocab_size=32768,
+        hidden_size=4096,
+        intermediate_size=14336,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=8,
+        hidden_act="silu",
+        max_position_embeddings=8192,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        residual_dropout=0.0,
+        embedding_dropout=0.0,
+        use_sliding_window=False,
+        sliding_window=None,
+        use_flash_attention_2=True,
+        pretraining_tp=1,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        
+        # Grouped Query Attention
+        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.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.mlp_bias = mlp_bias
+        self.residual_dropout = residual_dropout
+        self.embedding_dropout = embedding_dropout
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window
+        self.use_flash_attention_2 = use_flash_attention_2
+        self.pretraining_tp = pretraining_tp
+        
+        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,
+        )