kv-cache : fix out-of-bounds view during reserve graph (#13547)

* kv-cache : fix reserve graph out-of-bounds access

ggml-ci

* cont : add comment

* cont : fix comments [no ci]

* cont : more correct comment [no ci]

src/llama-kv-cache.cpp

@@ -441,6 +441,13 @@ void llama_kv_cache_unified::defrag_sched(float thold) {
 
 void llama_kv_cache_unified::set_full() {
     n = size;
+
+    // when simulating a full KV cache, the specific value of the "head" pointer is not important because it does not
+    //   affect the shapes of the tensors in the compute graph - it only affects the offsets of the K/V views.
+    //   we should only guarantee that the head position won't cause out-of-bounds view of the K, V tensors, so
+    //   setting it to 0 is the simplest way to achieve that
+    // ref: https://github.com/ggml-org/llama.cpp/issues/13359
+    head = 0;
 }
 
 llama_sbatch llama_kv_cache_unified::sbatch_init(
@@ -1712,6 +1719,7 @@ void llama_kv_cache_recurrent::defrag_sched(float thold) {
 
 void llama_kv_cache_recurrent::set_full() {
     n = size;
+    head = 0;
 }
 
 llama_sbatch llama_kv_cache_recurrent::sbatch_init(

src/llama-kv-cache.h

@@ -171,11 +171,8 @@ public:
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
     void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
 
-    // Note: The value of head isn't only used to optimize searching
-    // for a free KV slot. llama_decode_impl also uses it, so it
-    // cannot be freely changed after a slot has been allocated.
-    uint32_t head = 0;
-    uint32_t size = 0;
+    uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
+    uint32_t size = 0; // total number of cells, shared across all sequences
     uint32_t used = 0; // used cells (i.e. at least one seq_id)
 
     // computed before each graph build
@@ -343,11 +340,8 @@ public:
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
     void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
 
-    // Note: The value of head isn't only used to optimize searching
-    // for a free KV slot. llama_decode_impl also uses it, so it
-    // cannot be freely changed after a slot has been allocated.
-    uint32_t head = 0;
-    uint32_t size = 0;
+    uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
+    uint32_t size = 0; // total number of cells, shared across all sequences
     uint32_t used = 0; // used cells (i.e. at least one seq_id)
 
     // computed before each graph build