Merge pull request ggml-org#370 from Okabintaro/fix-state-pickle

fix: Make LLamaState pickleable for disk cache

Author: abetlen

Diff for: llama_cpp/llama.py

@@ -141,7 +141,9 @@ def __getitem__(self, key: Sequence[int]) -> "LlamaState":
         if _key is None:
             raise KeyError("Key not found")
         value: "LlamaState" = self.cache.pop(_key)  # type: ignore
-        self.cache.push(_key, side="front")  # type: ignore
+        # NOTE: This puts an integer as key in cache, which breaks, 
+        # Llama.longest_token_prefix(k, key) above since k is not a tuple of ints/tokens
+        # self.cache.push(_key, side="front")  # type: ignore
         return value
 
     def __contains__(self, key: Sequence[int]) -> bool:
@@ -168,7 +170,7 @@ def __init__(
         eval_logits: Deque[List[float]],
         input_ids: npt.NDArray[np.intc],
         scores: npt.NDArray[np.single],
-        llama_state,  # type: llama_cpp.Array[llama_cpp.c_uint8]
+        llama_state: bytes,
         llama_state_size: int,
     ):
         self.eval_tokens = eval_tokens
@@ -1509,7 +1511,7 @@ def save_state(self) -> LlamaState:
             eval_logits=self.eval_logits.copy(),
             scores=self._scores.copy(),
             input_ids=self._input_ids.copy(),
-            llama_state=llama_state_compact,
+            llama_state=bytes(llama_state_compact),
             llama_state_size=n_bytes,
         )
 
@@ -1520,7 +1522,10 @@ def load_state(self, state: LlamaState) -> None:
         self._scores = state.scores.copy()
         self._input_ids = state.input_ids.copy()
         state_size = state.llama_state_size
-        if llama_cpp.llama_set_state_data(self.ctx, state.llama_state) != state_size:
+        LLamaStateArrayType = (llama_cpp.c_uint8 * state_size)
+        llama_state = LLamaStateArrayType.from_buffer_copy(state.llama_state)
+
+        if llama_cpp.llama_set_state_data(self.ctx, llama_state) != state_size:
             raise RuntimeError("Failed to set llama state data")
 
     def n_ctx(self) -> int: