From: Michael Wand <redacted>
Date: Sun, 14 Jun 2026 09:52:15 +0000 (-0700)
Subject: cli : fix not copying preserved tokens (#24258)
X-Git-Tag: upstream/0.0.10438~807
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=6e14286edaa60a223292c8a996506905b2f66f66;p=pkg%2Fggml%2Fsources%2Fllama.cpp

cli : fix not copying preserved tokens (#24258)
---

diff --git a/tools/cli/cli.cpp b/tools/cli/cli.cpp
index 3ed345bf0..c03894b4b 100644
--- a/tools/cli/cli.cpp
+++ b/tools/cli/cli.cpp
@@ -97,11 +97,18 @@ struct cli_context {
                 task.params.chat_parser_params.parser.load(chat_params.parser);
             }
 
+            // Copy the preserved tokens into the sampling params
+            const llama_vocab * vocab = llama_model_get_vocab(
+                llama_get_model(ctx_server.get_llama_context()));
+            for (const auto & token : chat_params.preserved_tokens) {
+                auto ids = common_tokenize(vocab, token, false, true);
+                if (ids.size() == 1) {
+                    task.params.sampling.preserved_tokens.insert(ids[0]);
+                }
+            }
+
             // reasoning budget sampler
             if (!chat_params.thinking_end_tag.empty()) {
-                const llama_vocab * vocab = llama_model_get_vocab(
-                    llama_get_model(ctx_server.get_llama_context()));
-
                 task.params.sampling.reasoning_budget_tokens = defaults.sampling.reasoning_budget_tokens;
                 task.params.sampling.generation_prompt = chat_params.generation_prompt;