Merge branch 'custom_rope' into main

Author: abetlen

llama_cpp/llama.py

@@ -205,6 +205,8 @@ def __init__(
         model_path: str,
         # NOTE: These parameters are likely to change in the future.
         n_ctx: int = 512,
+        rope_freq_base: float = 10000.0,
+        rope_freq_scale: float = 1.0,
         n_parts: int = -1,
         n_gpu_layers: int = 0,
         seed: int = 1337,
@@ -228,6 +230,8 @@ def __init__(
         Args:
             model_path: Path to the model.
             n_ctx: Maximum context size.
+            rope_freq_base: RoPE base frequency.
+            rope_freq_scale: RoPE frequency scale.
             n_parts: Number of parts to split the model into. If -1, the number of parts is automatically determined.
             seed: Random seed. -1 for random.
             f16_kv: Use half-precision for key/value cache.
@@ -256,6 +260,8 @@ def __init__(
 
         self.params = llama_cpp.llama_context_default_params()
         self.params.n_ctx = n_ctx
+        self.params.rope_freq_base = rope_freq_base
+        self.params.rope_freq_scale = rope_freq_scale
         self.params.n_gpu_layers = n_gpu_layers
         self.params.seed = seed
         self.params.f16_kv = f16_kv

llama_cpp/llama_cpp.py

@@ -188,6 +188,8 @@ class llama_context_params(Structure):
     _fields_ = [
         ("seed", c_uint32),
         ("n_ctx", c_int32),
+        ("rope_freq_base", c_float),
+        ("rope_freq_scale", c_float),
         ("n_batch", c_int32),
         ("n_gpu_layers", c_int32),
         ("main_gpu", c_int32),