Add configuration_expivme_diffusion.py

configuration_expivme_diffusion.py

@@ -0,0 +1,35 @@
+"""HuggingFace Transformers config for ExpIvme-DiffusionConversate-v1-Instruct."""
+
+from transformers import PretrainedConfig
+
+
+class ExpIvmeDiffusionConfig(PretrainedConfig):
+    model_type = "expivme_diffusion"
+
+    def __init__(self, vocab_size=16004, hidden_dim=896, n_layers=12, n_heads=14,
+                 context_len=1024, ffn_mult=4.0, rope_theta=10_000.0, norm_eps=1e-5,
+                 tie_embeddings=True, dropout=0.0, mask_token_id=16000,
+                 user_token_id=16001, assistant_token_id=16002, endturn_token_id=16003, **kwargs):
+        self.vocab_size = vocab_size
+        self.hidden_dim = hidden_dim
+        self.n_layers = n_layers
+        self.n_heads = n_heads
+        self.context_len = context_len
+        self.ffn_mult = ffn_mult
+        self.rope_theta = rope_theta
+        self.norm_eps = norm_eps
+        self.dropout = dropout
+        self.mask_token_id = mask_token_id
+        self.user_token_id = user_token_id
+        self.assistant_token_id = assistant_token_id
+        self.endturn_token_id = endturn_token_id
+        self.max_position_embeddings = context_len
+        self.num_hidden_layers = n_layers
+        self.num_attention_heads = n_heads
+        self.hidden_size = hidden_dim
+        kwargs.setdefault("tie_word_embeddings", tie_embeddings)
+        super().__init__(**kwargs)
+
+    @property
+    def head_dim(self):
+        return self.hidden_dim // self.n_heads