Merge pull request #522 from bretello/llama2-70b-support

Llama2 70b support

README.md

@@ -135,6 +135,14 @@ For instance, if you want to work with larger contexts, you can expand the conte
 llm = Llama(model_path="./models/7B/ggml-model.bin", n_ctx=2048)
 ```
 
+### Loading llama-2 70b
+
+Llama2 70b must set the `n_gqa` parameter (grouped-query attention factor) to 8 when loading:
+
+```python
+llm = Llama(model_path="./models/7B/ggml-model.bin", n_gqa=8)
+```
+
 ## Web Server
 
 `llama-cpp-python` offers a web server which aims to act as a drop-in replacement for the OpenAI API.

llama_cpp/llama.py

@@ -216,6 +216,7 @@ def __init__(
         embedding: bool = False,
         n_threads: Optional[int] = None,
         n_batch: int = 512,
+        n_gqa: Optional[int] = None,  # must be 8 for llama2 70b
         last_n_tokens_size: int = 64,
         lora_base: Optional[str] = None,
         lora_path: Optional[str] = None,
@@ -260,6 +261,8 @@ def __init__(
 
         self.params = llama_cpp.llama_context_default_params()
         self.params.n_ctx = n_ctx
+        if n_gqa is not None:
+            self.params.n_gqa = n_gqa
         self.params.n_gpu_layers = n_gpu_layers
         self.params.seed = seed
         self.params.f16_kv = f16_kv