bug fix

firecoperana · firecoperana · commit 8bb26c48e7c2 · 2026-01-25T11:33:30.000-06:00
diff --git a/examples/server/server-context.cpp b/examples/server/server-context.cpp
@@ -2829,17 +2829,10 @@ inline void rewind_context(server_slot& slot, int32_t n_rewind) {
         }
     }
 
-
     slot.token_buffer.resize(n_keep_rewind);
-    slot.cache_tokens.keep_first(slot.cache_tokens.size() - n_rewind);
-
-	// to be used for common batch add
-    if (n_keep_rewind >= 1) {
-        slot.sampled = slot.token_buffer.at(n_keep_rewind - 1).tok; 
-    }
-    else {
-        slot.sampled = slot.cache_tokens[slot.cache_tokens.size() - 1];
-    }
+    size_t n_keep = slot.cache_tokens.size() - n_rewind;
+    slot.sampled = slot.cache_tokens[n_keep];
+    slot.cache_tokens.keep_first(n_keep);
 }
 
 void server_context::buffer_and_check_string_ban(server_slot & slot, completion_token_output & result) {

Original file line number	Diff line number	Diff line change
`@@ -2829,17 +2829,10 @@ inline void rewind_context(server_slot& slot, int32_t n_rewind) {`
`2829`	`2829`	`}`
`2830`	`2830`	`}`
`2831`	`2831`
`2832`		`-`
`2833`	`2832`	`slot.token_buffer.resize(n_keep_rewind);`
`2834`		`- slot.cache_tokens.keep_first(slot.cache_tokens.size() - n_rewind);`
`2835`		`-`
`2836`		`- // to be used for common batch add`
`2837`		`- if (n_keep_rewind >= 1) {`
`2838`		`- slot.sampled = slot.token_buffer.at(n_keep_rewind - 1).tok;`
`2839`		`- }`
`2840`		`- else {`
`2841`		`- slot.sampled = slot.cache_tokens[slot.cache_tokens.size() - 1];`
`2842`		`- }`
	`2833`	`+ size_t n_keep = slot.cache_tokens.size() - n_rewind;`
	`2834`	`+ slot.sampled = slot.cache_tokens[n_keep];`
	`2835`	`+ slot.cache_tokens.keep_first(n_keep);`
`2843`	`2836`	`}`
`2844`	`2837`
`2845`	`2838`	`void server_context::buffer_and_check_string_ban(server_slot & slot, completion_token_output & result) {`