reinstate tool call id logic, keep track of previously generated ids

Author: ochafik

common/chat.cpp

@@ -92,7 +92,9 @@ std::vector<common_chat_msg_diff> common_chat_msg_diff::compute_diffs(const comm
             auto & diff = diffs.emplace_back();
             diff.tool_call_index = idx;
             diff.tool_call_delta.name = newf.name;
-            diff.tool_call_delta.id = newf.id;
+            if (pref.id != newf.id) {
+                diff.tool_call_delta.id = newf.id;
+            }
             diff.tool_call_delta.arguments = args_diff;
         }
     }

common/chat.h

@@ -42,6 +42,18 @@ struct common_chat_msg {
     bool empty() const {
         return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() && tool_name.empty() && tool_call_id.empty();
     }
+    void ensure_tool_call_ids_set(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
+        for (auto i = 0u; i < tool_calls.size(); i++) {
+            if (ids_cache.size() <= i) {
+                auto id = tool_calls[i].id;
+                if (id.empty()) {
+                    id = gen_tool_call_id();
+                }
+                ids_cache.push_back(id);
+            }
+            tool_calls[i].id = ids_cache[i];
+        }
+    }
     bool operator==(const common_chat_msg & other) const {
         return role == other.role
             && content == other.content

examples/server/server.cpp

@@ -1282,6 +1282,7 @@ struct server_slot {
     llama_token sampled;
 
     common_chat_format chat_format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    std::vector<std::string> generated_tool_call_ids;
 
     // stats
     size_t n_sent_text        = 0; // number of sent text character
@@ -1313,6 +1314,7 @@ struct server_slot {
         generated_token_probs.clear();
         generated_msg = {};
         json_schema = json();
+        generated_tool_call_ids.clear();
     }
 
     bool is_non_causal() const {
@@ -2356,14 +2358,12 @@ struct server_context {
             /* is_partial= */ true,
             slot.params.oaicompat_chat_syntax);
         if (!new_msg.empty()) {
+            new_msg.ensure_tool_call_ids_set(slot.generated_tool_call_ids, gen_tool_call_id);
             slot.generated_msg = new_msg;
         }
         res->oaicompat_previous_msg = previous_msg;
         res->oaicompat_new_msg      = new_msg.empty() ? previous_msg : new_msg;
 
-        // res->previous_content = slot.generated_text.substr(0, slot.generated_text.size() - tkn.text_to_send.size());
-        // res->oaicompat_chat_format = slot.params.oaicompat_chat_format;
-
         // populate res.probs_output
         if (slot.params.sampling.n_probs > 0) {
             res->prob_output = tkn; // copy the token probs
@@ -2409,6 +2409,7 @@ struct server_context {
             res->content,
             /* is_partial= */ slot.stop == STOP_TYPE_LIMIT,
             slot.params.oaicompat_chat_syntax);
+        res->oaicompat_msg.ensure_tool_call_ids_set(slot.generated_tool_call_ids, gen_tool_call_id);
         res->oaicompat_chat_syntax = slot.params.oaicompat_chat_syntax;
 
         // populate res.probs_output