A high-throughput and memory-efficient inference and serving engine for LLMs
[Rust Frontend] Fix DeepSeek V3.2 continue_final_message rendering (#45155)
Signed-off-by: reidliu41 <reid201711@gmail.com>
Reid committed
Jun 11, 2026 at 17:34 UTC
0d657e44dcca844499b7adcd03ec590f933dfd29
2 files changed
+22
-3
rust/src/chat/src/renderer/deepseek_v32/encoding.rs
+4
-3
@@ -49,6 +49,7 @@ pub(super) fn render_request(request: &ChatRequest) -> Result<String> {
49
let last_user_render_index =
50
find_last_user_render_index(request.messages.as_slice(), render_offset);
51
let last_user_actual_index = find_last_user_actual_index(request.messages.as_slice());
52
+ let continue_final_message = request.chat_options.continue_final_message();
53
let mut prompt = String::from(BOS_TOKEN);
54
55
if request.tool_parsing_enabled() {
@@ -66,6 +67,7 @@ pub(super) fn render_request(request: &ChatRequest) -> Result<String> {
67
last_user_actual_index,
68
thinking_mode,
69
drop_thinking,
70
+ continue_final_message,
71
)?;
72
}
73
@@ -96,6 +98,7 @@ fn render_message(
98
last_user_actual_index: usize,
99
thinking_mode: ThinkingMode,
100
drop_thinking: bool,
101
+ continue_final_message: bool,
102
) -> Result<()> {
103
let render_index = message_index as isize + render_offset;
104
let opens_thinking = render_index == last_user_render_index;
@@ -125,9 +128,7 @@ fn render_message(
128
thinking_mode,
129
drop_thinking,
130
),
128
- // TODO: Respect `continue_final_message` and map it to DeepSeek's
129
- // prefix-style final-assistant continuation behavior.
130
- false,
131
+ continue_final_message && message_index + 1 == messages.len(),
132
),
133
ChatMessage::ToolResponse { content, .. } => render_tool_message(
134
out,
rust/src/chat/src/renderer/deepseek_v32/tests.rs
+18
@@ -404,6 +404,24 @@ fn assistant_after_last_user_requires_reasoning_or_tool_calls() {
404
expect!["chat template error: invalid DeepSeek V3.2 assistant message after last user message: expected reasoning or tool calls"]
405
.assert_eq(&error.to_report_string());
406
}
407
+
408
+#[test]
409
+fn continue_final_assistant_omits_final_eos() {
410
+ let mut request = ChatRequest {
411
+ messages: vec![
412
+ ChatMessage::user("write"),
413
+ ChatMessage::assistant_text("partial answer"),
414
+ ],
415
+ ..ChatRequest::for_test()
416
+ };
417
+ request.chat_options.generation_prompt_mode = GenerationPromptMode::ContinueFinalAssistant;
418
+
419
+ let rendered = render_request(&request);
420
+
421
+ expect!["<｜begin▁of▁sentence｜><｜User｜>write<｜Assistant｜></think>partial answer"]
422
+ .assert_eq(&rendered);
423
+}
424
+
425
#[test]
426
fn render_rejects_multimodal_input() {
427
let request = ChatRequest {