Upload configuration_recast_llama.py with huggingface_hub

configuration_recast_llama.py

@@ -0,0 +1,78 @@
+# filename: configuration_recast_llama2.py
+from transformers import PretrainedConfig
+
+
+class RECAST7b_llama(PretrainedConfig):
+    model_type = "recast7b_llama"
+    attribute_map = {
+        "hidden_size": "hidden_size",
+        "num_attention_heads": "num_attention_heads",
+    }
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        hidden_size=4096,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,  # Llama 2 uses same as num_attention_heads
+        hidden_act="silu",
+        max_position_embeddings=4096,  # Standard Llama 2 context length
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,  # Llama 2 uses 1e-6
+        use_cache=True,
+        pad_token_id=0,  # Llama 2 uses 0
+        bos_token_id=1,  # Llama 2 uses 1
+        eos_token_id=2,  # Llama 2 uses 2
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,  # Standard RoPE, no complex scaling
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        # Template-specific configs
+        num_templates=2,
+        num_groups=16,  # Adjust based on your compression needs
+        coef_height=8,  # Adjust based on your compression needs
+        # num_cf=1,
+        torch_dtype="bfloat16",
+        **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
+        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.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.mlp_bias = mlp_bias
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.rope_theta = rope_theta
+        self.rope_scaling = None  # Llama 2 doesn't use rope scaling
+        self.torch_dtype = torch_dtype
+
+        # Template-specific configs
+        self.num_templates = num_templates
+        self.num_groups = num_groups
+        self.coef_height = coef_height
+        # self.num_cf = num_cf
+        # self.num_rank = num_rank
+
+        # Calculate coef_width automatically
+        layers_per_group = self.num_hidden_layers // self.num_groups
+        self.coef_width = self.coef_height * layers_per_group
+
+        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
+        )