refactor: address PR review feedback

- Fix n_samplers type hint to ctypes.c_size_t
- Map flash_attn=True exactly to LLAMA_FLASH_ATTN_TYPE_ENABLED instead of relying on bool()
- Raise NotImplementedError instead of returning stub when C symbol is missing

Author: codavidgarcia

llama_cpp/_ctypes_extensions.py

@@ -118,7 +118,14 @@ def decorator(f: F) -> F:
                     return func
                 except AttributeError:
                     # Symbol not found in shared library (deprecated/removed)
-                    return f
+                    @functools.wraps(f)
+                    def stub(*args: Any, **kwargs: Any) -> Any:
+                        raise NotImplementedError(
+                            f"Symbol '{name}' not found in shared library. The C API might "
+                            "have been removed or deprecated."
+                        )
+
+                    return stub  # type: ignore
             else:
                 return f
 

llama_cpp/llama.py

@@ -2096,7 +2096,10 @@ def __getstate__(self):
             logits_all=self._logits_all,
             embedding=self.context_params.embeddings,
             offload_kqv=self.context_params.offload_kqv,
-            flash_attn=bool(self.context_params.flash_attn_type),
+            flash_attn=(
+                self.context_params.flash_attn_type
+                == llama_cpp.LLAMA_FLASH_ATTN_TYPE_ENABLED
+            ),
             op_offload=self.context_params.op_offload,
             swa_full=self.context_params.swa_full,
             # Sampling Params

llama_cpp/llama_cpp.py

@@ -874,7 +874,7 @@ class llama_context_params(ctypes.Structure):
         swa_full: bool
         kv_unified: bool
         samplers: ctypes.c_void_p
-        n_samplers: int
+        n_samplers: ctypes.c_size_t
 
     _fields_ = [
         ("n_ctx", ctypes.c_uint32),