cont : minor

Author: ggerganov

common/speculative.cpp

@@ -1252,22 +1252,22 @@ struct common_speculative_session::impl {
     }
 
     bool accept(llama_tokens ids) {
-        has_partial = false;
-
         LOG_WRN("%s: n_draft=%zu, ids.size=%zu\n", __func__, draft.size(), ids.size());
 
+        has_partial = false;
+
         if (ids.size() < draft.size() + 1) {
             // the main model rejected some tokens
-            LOG_DBG("%s: partial acceptance: %zu < %zu\n", __func__, draft.size(), draft.size());
-
             if (params.use_checkpoints) {
-                // we shorten the draft and retry
+                // shorten the draft to the number of accepted tokens
                 draft.resize(ids.size() - 1);
 
                 has_partial = true;
 
                 return false;
             }
+
+            LOG_DBG("%s: partial acceptance: %zu < %zu\n", __func__, draft.size(), draft.size());
         }
 
         draft = std::move(ids);

tools/server/server-context.cpp

@@ -84,10 +84,8 @@ struct server_slot {
     mtmd_context * mctx = nullptr;
 
     // speculative decoding
-    server_prompt_checkpoint spec_ckpt;
-
     std::vector<int32_t> spec_i_batch;
-
+    server_prompt_checkpoint spec_ckpt;
     std::unique_ptr<common_speculative_session> spec = nullptr;
 
     // TODO: move members that belong to the task (such as `generated_text`, `has_new_line`) to task_results_state
@@ -204,8 +202,8 @@ struct server_slot {
         n_sent_text    = 0;
 
         if (can_speculate()) {
-            spec_ckpt = {};
             spec_i_batch.clear();
+            spec_ckpt.clear();
             spec->clear();
         }
         generated_tokens.clear();
@@ -2950,9 +2948,7 @@ struct server_context_impl {
                     common_sampler_ptr smpl_save(common_sampler_clone(slot.smpl.get()));
 
                     GGML_ASSERT(slot.spec_i_batch.size() == n_draft + 1);
-
                     auto accepted = common_sampler_sample_and_accept_n(slot.smpl.get(), slot.ctx, slot.spec_i_batch, draft);
-
                     slot.spec_i_batch.clear();
 
                     if (!slot.spec->accept(std::move(accepted))) {

tools/server/server-task.h

@@ -580,6 +580,13 @@ struct server_prompt_checkpoint {
     bool empty() const {
         return data.empty();
     }
+
+    void clear() {
+        pos_min = 0;
+        pos_max = 0;
+        n_tokens = 0;
+        data.clear();
+    }
 };
 
 struct server_prompt {