llama: restore a kv_cache in case of failed computation

src/llama.cpp

@@ -2811,6 +2811,22 @@ struct llama_kv_cache {
     }
 };
 
+// saves the kv_cache state for future recovery
+// used to preserve the kv_cache state before searching for a slot
+struct llama_kv_slot_restorer {
+    struct llama_kv_cache_state {
+        uint32_t head = 0;
+        uint32_t size = 0;
+        uint32_t used = 0;
+        uint32_t n    = 0;
+    } old_state;
+
+    std::vector<llama_kv_cell> recurrent_cells;    // for     recurrent models only
+    std::pair<uint32_t, uint32_t> slot_boundaries; // for non-recurrent models only
+
+    bool restore = false;
+};
+
 struct llama_control_vector {
     std::vector<struct ggml_tensor *> tensors; // per layer
     std::vector<ggml_context_ptr> ctxs;
@@ -3508,11 +3524,19 @@ static bool llama_kv_cache_init(
 // to the first cell of the slot.
 static bool llama_kv_cache_find_slot(
            struct llama_kv_cache & cache,
-       const struct llama_ubatch & batch) {
+       const struct llama_ubatch & batch,
+   struct llama_kv_slot_restorer * slot_restorer = nullptr) {
     const uint32_t n_tokens = batch.n_tokens;
     const uint32_t n_seqs   = batch.n_seqs;
     const uint32_t n_seq_tokens = batch.n_seq_tokens;
 
+    if (slot_restorer != nullptr) {
+        slot_restorer->old_state.head  = cache.head;
+        slot_restorer->old_state.size  = cache.size;
+        slot_restorer->old_state.used  = cache.used;
+        slot_restorer->old_state.n     = cache.n;
+    }
+
     if (cache.recurrent) {
         // For recurrent state architectures (like Mamba or RWKV),
         // each cache cell can store the state for a whole sequence.
@@ -3521,6 +3545,11 @@ static bool llama_kv_cache_find_slot(
         // can only process batches with an equal number of new tokens in each sequence
         GGML_ASSERT(batch.equal_seqs);
 
+        if (slot_restorer != nullptr) {
+            slot_restorer->recurrent_cells = cache.cells;
+            slot_restorer->restore = true;
+        }
+
         int32_t min = cache.size - 1;
         int32_t max = 0;
 
@@ -3709,6 +3738,11 @@ static bool llama_kv_cache_find_slot(
         }
     }
 
+    if (slot_restorer != nullptr) {
+        slot_restorer->slot_boundaries = std::make_pair(cache.head, cache.head + n_tokens);
+        slot_restorer->restore = true;
+    }
+
     for (uint32_t s = 0; s < n_seqs; s++) {
         for (uint32_t i = 0; i < n_seq_tokens; ++i) {
             uint32_t k = s*n_seq_tokens + i;
@@ -3998,6 +4032,23 @@ static uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams)
     return cparams.flash_attn ? 256u : 32u;
 }
 
+static void llama_kv_cache_slot_restore(
+        const struct llama_kv_slot_restorer & restorer,
+                      struct llama_kv_cache & cache) {
+    if (restorer.restore) {
+        cache.head  = restorer.old_state.head;
+        cache.size  = restorer.old_state.size;
+        cache.used  = restorer.old_state.used;
+        cache.n     = restorer.old_state.n;
+
+        if (cache.recurrent) {
+            cache.cells = restorer.recurrent_cells;
+        } else {
+            llama_kv_cache_seq_rm(cache, -1, restorer.slot_boundaries.first, restorer.slot_boundaries.second + 1);
+        }
+    }
+}
+
 //
 // model loading and saving
 //
@@ -17256,6 +17307,7 @@ static int llama_decode_internal(
     lctx.n_queued_tokens += n_tokens_all;
 
     auto & kv_self = lctx.kv_self;
+    llama_kv_slot_restorer kv_slot_restorer;
 
     const int64_t n_embd  = hparams.n_embd;
     const int64_t n_vocab = hparams.n_vocab;
@@ -17340,7 +17392,7 @@ static int llama_decode_internal(
                 kv_self.head = 0;
             }
 
-            if (!llama_kv_cache_find_slot(kv_self, ubatch)) {
+            if (!llama_kv_cache_find_slot(kv_self, ubatch, &kv_slot_restorer)) {
                 return 1;
             }
 
@@ -17390,9 +17442,9 @@ static int llama_decode_internal(
         llama_set_inputs(lctx, ubatch);
 
         const auto compute_status = llama_graph_compute(lctx, gf, n_threads, threadpool);
+        if (compute_status != GGML_STATUS_SUCCESS) {
+            llama_kv_cache_slot_restore(kv_slot_restorer, kv_self);
             switch (compute_status) {
-            case GGML_STATUS_SUCCESS:
-                break;
                 case GGML_STATUS_ABORTED:
                     return 2;
                 case GGML_STATUS_ALLOC_FAILED:
@@ -17401,6 +17453,7 @@ static int llama_decode_internal(
                 default:
                     return -3;
             }
+        }
 
         // update the kv ring buffer
         {