llama : defragment via non-overlapping moves

Author: ggerganov

llama.cpp

@@ -8028,7 +8028,7 @@ static int llama_decode_internal(
 }
 
 // copy the KV cache to the host memory and reshuffle the cells to the beginning of the cache
-// this way we eliminate any empty segments that may have been left by previous KV cache operations
+// this way we eliminate any empty holes that may have been left by previous KV cache operations
 //
 // TODO: optimizations are possible:
 //       - multiple threads
@@ -8045,36 +8045,81 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
     const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa();
     const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa();
     const uint32_t n_kv         = llama_kv_cache_cell_max(kv_self);
+    const uint32_t n_used       = kv_self.used;
 
     const uint32_t kv_size = kv_self.size;
 
+    assert(n_used <= n_kv);
+
     const int64_t t_start = ggml_time_us();
 
     std::vector<uint8_t> buf_k;
     std::vector<uint8_t> buf_v;
 
-    // the destination cell in the new KV cache
-    uint32_t id = 0;
-
     // number of cells moved
     uint32_t n_moves = 0;
 
     // determine which KV cells to move where
     std::vector<uint32_t> ids(n_kv, n_kv);
 
-    for (uint32_t i0 = 0; i0 < n_kv; ++i0) {
+    for (uint32_t i0 = 0; i0 < n_used; ++i0) {
         const auto & cell0 = kv_self.cells[i0];
 
         if (!cell0.is_empty()) {
-            ids[i0] = id;
+            ids[i0] = i0;
+
+            continue;
+        }
+
+        // found a hole - fill it with data from the end of the cache
+
+        // determine the size of the hole
+        uint32_t nh = 1;
+        while (i0 + nh < n_used && kv_self.cells[i0 + nh].is_empty()) {
+            nh++;
+        }
+
+        // starting from the end, find nh non-empty cells
+        uint32_t nf = 0;
+        uint32_t is = n_kv - 1;
+        for (; is > i0; --is) {
+            const auto & cell1 = kv_self.cells[is];
+
+            if (cell1.is_empty() || ids[is] != n_kv) {
+                continue;
+            }
 
-            if (i0 != id) {
-                kv_self.cells[id] = cell0;
-                n_moves++;
+            // non-empty cell which is not yet moved
+            nf++;
+            if (nf == nh) {
+                break;
+            }
+        }
+
+        GGML_ASSERT(nf == nh && "KV defrag bug: nf != nh");
+
+        nf = 0;
+
+        // go back and move the nf cells to the hole
+        for (uint32_t i1 = is; i1 < n_kv; ++i1) {
+            const auto & cell1 = kv_self.cells[i1];
+
+            if (cell1.is_empty() || ids[i1] != n_kv) {
+                continue;
             }
 
-            id++;
+            ids[i1] = i0 + nf;
+
+            // move the cell meta data
+            kv_self.cells[i0 + nf] = cell1;
+
+            n_moves++;
+            nf++;
         }
+
+        LLAMA_LOG_INFO("(tmp log) KV defrag: move [%u, %u) to [%u, %u)\n", is, n_kv, i0, i0 + nh);
+
+        i0 += nh - 1;
     }
 
     if (n_moves == 0) {
@@ -8083,11 +8128,11 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
 
     LLAMA_LOG_INFO("(tmp log) KV defrag cell moves: %u\n", n_moves);
 
-    kv_self.head = id;
-    kv_self.used = id;
+    kv_self.head = n_used;
+    kv_self.used = n_used;
 
     // zero the rest of the cells
-    for (uint32_t i = id; i < n_kv; ++i) {
+    for (uint32_t i = n_used; i < n_kv; ++i) {
         kv_self.cells[i] = llama_kv_cell();
     }