batch : add n_used count (ggml-org#14512)

ggml-ci

Author: ggerganov

src/llama-batch.cpp

@@ -411,6 +411,10 @@ uint32_t llama_batch_allocr::get_n_outputs() const {
     return n_outputs;
 }
 
+uint32_t llama_batch_allocr::get_n_used() const {
+    return n_used;
+}
+
 std::vector<int32_t> & llama_batch_allocr::get_out_ids() {
     return out_ids;
 }
@@ -426,6 +430,8 @@ llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const {
 void llama_batch_allocr::split_reset() {
     out_ids.clear();
 
+    n_used = 0;
+
     used.clear();
     used.resize(get_n_tokens(), false);
 
@@ -450,6 +456,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
         idxs.push_back(cur_idx);
 
         used[cur_idx] = true;
+        ++n_used;
 
         ++cur_idx;
 
@@ -535,6 +542,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
             idxs_per_seq[s].push_back(idx);
 
             used[idx] = true;
+            ++n_used;
 
             ++cur_idx[s];
         }
@@ -576,6 +584,7 @@ llama_ubatch llama_batch_allocr::split_seq(uint32_t n_ubatch) {
         idxs.push_back(cur_idx);
 
         used[cur_idx] = true;
+        ++n_used;
 
         if (idxs.size() >= n_ubatch) {
             break;

src/llama-batch.h

@@ -56,6 +56,7 @@ class llama_batch_allocr {
 
     uint32_t get_n_tokens()  const;
     uint32_t get_n_outputs() const;
+    uint32_t get_n_used()    const;
 
     // the array of output indices in the order they were encountered during the ubatch splitting
     std::vector<int32_t> & get_out_ids();
@@ -127,6 +128,8 @@ class llama_batch_allocr {
     // batch indices of the output
     std::vector<int32_t> out_ids;
 
+    uint32_t n_used;
+
     // used[i] indicates if token i has already been used in a previous ubatch
     std::vector<bool> used;
 

src/llama-kv-cache-unified-iswa.cpp

@@ -113,6 +113,11 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
         auto sinfos_base = kv_base->prepare(ubatches);
         if (sinfos_base.empty()) {
             break;
@@ -144,6 +149,11 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
         auto sinfos_base = kv_base->prepare(ubatches);
         if (sinfos_base.empty()) {
             break;

src/llama-kv-cache-unified.cpp

@@ -353,6 +353,11 @@ llama_memory_context_ptr llama_kv_cache_unified::init_batch(
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
         auto sinfos = prepare(ubatches);
         if (sinfos.empty()) {
             break;

src/llama-memory-hybrid.cpp

@@ -80,6 +80,11 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
         // prepare the recurrent batches first
         if (!mem_recr->prepare(ubatches)) {
             // TODO: will the recurrent cache be in an undefined context at this point?

src/llama-memory-recurrent.cpp

@@ -377,7 +377,8 @@ llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr &
                 ubatch = balloc.split_equal(n_ubatch);
             }
 
-            if (ubatch.n_tokens == 0) {
+            if (balloc.get_n_used() < balloc.get_n_tokens()) {
+                // failed to find a suitable split
                 break;
             }