llama : add api for getting/setting the complete state: rng, logits, embedding and kv_cache (#1105)

* reserve correct size for logits

* add functions to get and set the whole llama state:

including rng, logits, embedding and kv_cache

* remove unused variables

* remove trailing whitespace

* fix comment

1 files changed, 12 insertions, 0 deletions

diff --git a/llama.h b/llama.h
index e95ff73..f68a0cb 100644
--- a/llama.h
+++ b/llama.h
@@ -129,6 +129,18 @@ extern "C" {
                           size_t   n_size,
                              int   n_token_count);
 
+    // Returns the size in bytes of the state (rng, logits, embedding and kv_cache)
+    LLAMA_API size_t llama_get_state_size(struct llama_context * ctx);
+
+    // Copies the state to the specified destination address.
+    // Destination needs to have allocated enough memory.
+    // Returns the number of bytes copied
+    LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dest);
+
+    // Set the state reading from the specified address
+    // Returns the number of bytes read
+    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src);
+
     // Run the llama inference to obtain the logits and probabilities for the next token.
     // tokens + n_tokens is the provided batch of new tokens to process
     // n_past is the number of tokens to use from previous eval calls