llama : refactor session file management (#8699)

* llama : refactor session file management

* llama : saving and restoring state checks for overflow

The size of the buffers should now be given to the functions working
with them, otherwise a truncated file could cause out of bound reads.

* llama : stream from session file instead of copying into a big buffer

Loading session files should no longer cause a memory usage spike.

* llama : llama_state_get_size returns the actual size instead of max

This is a breaking change, but makes that function *much* easier
to keep up to date, and it also makes it reflect the behavior
of llama_state_seq_get_size.

* llama : share code between whole and seq_id-specific state saving

Both session file types now use a more similar format.

* llama : no longer store all hparams in session files

Instead, the model arch name is stored.
The layer count and the embedding dimensions of the KV cache
are still verified when loading.
Storing all the hparams is not necessary.

* llama : fix uint64_t format type

* llama : various integer type cast and format string fixes

Some platforms use "%lu" and others "%llu" for uint64_t.
Not sure how to handle that, so casting to size_t when displaying errors.

* llama : remove _context suffix for llama_data_context

* llama : fix session file loading

llama_state_get_size cannot be used to get the max size anymore.

* llama : more graceful error handling of invalid session files

* llama : remove LLAMA_MAX_RNG_STATE

It's no longer necessary to limit the size of the RNG state,
because the max size of session files is not estimated anymore.

* llama : cast seq_id in comparison with unsigned n_seq_max

examples/save-load-state/save-load-state.cpp

@@ -47,7 +47,7 @@ int main(int argc, char ** argv) {
     // save state (rng, logits, embedding and kv_cache) to file
     {
         std::vector<uint8_t> state_mem(llama_state_get_size(ctx));
-        const size_t written = llama_state_get_data(ctx, state_mem.data());
+        const size_t written = llama_state_get_data(ctx, state_mem.data(), state_mem.size());
 
         FILE *fp_write = fopen("dump_state.bin", "wb");
         fwrite(state_mem.data(), 1, written, fp_write);
@@ -99,13 +99,16 @@ int main(int argc, char ** argv) {
 
     // load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector<uint8_t> state_mem(llama_state_get_size(ctx2));
+        std::vector<uint8_t> state_mem;
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        fseek(fp_read, 0, SEEK_END);
+        state_mem.resize(ftell(fp_read));
+        fseek(fp_read, 0, SEEK_SET);
         const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
         fclose(fp_read);
 
-        if (read != llama_state_set_data(ctx2, state_mem.data())) {
+        if (read != llama_state_set_data(ctx2, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
@@ -159,13 +162,16 @@ int main(int argc, char ** argv) {
 
     // load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector<uint8_t> state_mem(llama_state_get_size(ctx3));
+        std::vector<uint8_t> state_mem;
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        fseek(fp_read, 0, SEEK_END);
+        state_mem.resize(ftell(fp_read));
+        fseek(fp_read, 0, SEEK_SET);
         const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
         fclose(fp_read);
 
-        if (read != llama_state_set_data(ctx3, state_mem.data())) {
+        if (read != llama_state_set_data(ctx3, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx3);
             llama_free_model(model);
@@ -182,7 +188,7 @@ int main(int argc, char ** argv) {
     {
         // save kv of seq 0
         std::vector<uint8_t> seq_store(llama_state_seq_get_size(ctx3, 0));
-        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), 0);
+        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), seq_store.size(), 0);
         if (ncopy != seq_store.size()) {
             fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
             llama_free(ctx3);
@@ -196,7 +202,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1
-        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), 1);
+        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), seq_store.size(), 1);
         if (nset != seq_store.size()) {
             fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
             llama_free(ctx3);

include/llama.h

@@ -33,17 +33,15 @@
 
 #define LLAMA_DEFAULT_SEED 0xFFFFFFFF
 
-#define LLAMA_MAX_RNG_STATE (64*1024)
-
 #define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 #define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 7
+#define LLAMA_SESSION_VERSION 8
 
 #define LLAMA_STATE_SEQ_MAGIC   LLAMA_FILE_MAGIC_GGSQ
-#define LLAMA_STATE_SEQ_VERSION 1
+#define LLAMA_STATE_SEQ_VERSION 2
 
 #ifdef __cplusplus
 extern "C" {
@@ -691,10 +689,11 @@ extern "C" {
     // State / sessions
     //
 
-    // Returns the maximum size in bytes of the state (rng, logits, embedding
-    // and kv_cache) - will often be smaller after compacting tokens
-    LLAMA_API size_t llama_state_get_size(const struct llama_context * ctx);
-    LLAMA_API DEPRECATED(size_t llama_get_state_size(const struct llama_context * ctx),
+    // Returns the *actual* size in bytes of the state
+    // (rng, logits, embedding and kv_cache)
+    // Only use when saving the state, not when restoring it, otherwise the size may be too small.
+    LLAMA_API size_t llama_state_get_size(struct llama_context * ctx);
+    LLAMA_API DEPRECATED(size_t llama_get_state_size(struct llama_context * ctx),
         "use llama_state_get_size instead");
 
     // Copies the state to the specified destination address.
@@ -702,7 +701,8 @@ extern "C" {
     // Returns the number of bytes copied
     LLAMA_API size_t llama_state_get_data(
             struct llama_context * ctx,
-                         uint8_t * dst);
+                         uint8_t * dst,
+                          size_t   size);
     LLAMA_API DEPRECATED(size_t llama_copy_state_data(
             struct llama_context * ctx,
                          uint8_t * dst),
@@ -712,7 +712,8 @@ extern "C" {
     // Returns the number of bytes read
     LLAMA_API size_t llama_state_set_data(
             struct llama_context * ctx,
-                   const uint8_t * src);
+                   const uint8_t * src,
+                          size_t   size);
     LLAMA_API DEPRECATED(size_t llama_set_state_data(
             struct llama_context * ctx,
                    const uint8_t * src),
@@ -754,6 +755,7 @@ extern "C" {
     LLAMA_API size_t llama_state_seq_get_data(
             struct llama_context * ctx,
                          uint8_t * dst,
+                          size_t   size,
                     llama_seq_id   seq_id);
 
     // Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence
@@ -763,6 +765,7 @@ extern "C" {
     LLAMA_API size_t llama_state_seq_set_data(
             struct llama_context * ctx,
                    const uint8_t * src,
+                          size_t   size,
                     llama_seq_id   dest_seq_id);
 
     LLAMA_API size_t llama_state_seq_save_file(