[Bugfix][Rust Frontend] Honor add_generation_prompt in DeepSeek V3.2/V4/V4.1 renderers (#56593)

Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Signed-off-by: YeonwooSung <neos960518@gmail.com>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
This commit is contained in:
Yeonwoo Sung
2026-09-15 10:03:26 +00:00
committed by GitHub
co-authored by Bugen Zhao
parent 073f883b56
commit 142020c6c2
5 changed files with 217 additions and 24 deletions
+28 -19
View File
@@ -224,21 +224,30 @@ pub(super) fn render_request_with_media_order(
}
}
if (is_user_like_entry(message, current_render_index as usize, dialect)
|| (dialect == DsDialect::V4 && matches!(message, ChatMessage::System { .. })))
&& next_rendered_entry_is_assistant_or_end(
if is_user_like_entry(message, current_render_index as usize, dialect)
|| (dialect == DsDialect::V4 && matches!(message, ChatMessage::System { .. }))
{
// Skip the opener only at the last rendered entry when
// add_generation_prompt is off. Inspect the next rendered
// entry, not the raw index: V4 merges trailing tool results.
let should_write_transition = match next_rendered_entry(
request.messages.as_slice(),
message_index,
drop_historical_developers,
last_user_like_message_index,
)
{
write_assistant_transition(
&mut out,
thinking_mode,
drop_thinking,
current_render_index >= last_user_render_index,
);
) {
Some(ChatMessage::Assistant { .. }) => true,
None => request.chat_options.add_generation_prompt(),
Some(_) => false,
};
if should_write_transition {
write_assistant_transition(
&mut out,
thinking_mode,
drop_thinking,
current_render_index >= last_user_render_index,
);
}
}
}
@@ -420,14 +429,17 @@ fn is_historical_developer(
&& last_user_like_message_index.is_some_and(|last_index| message_index < last_index)
}
/// Return whether the next rendered entry is assistant, or there is no next
/// entry.
fn next_rendered_entry_is_assistant_or_end(
/// Return the next rendered message after `message_index`.
///
/// Skips merged user-content tails and dropped historical developers so the
/// caller can tell a following assistant turn apart from the end of the
/// conversation.
fn next_rendered_entry(
messages: &[ChatMessage],
message_index: usize,
drop_historical_developers: bool,
last_user_like_message_index: Option<usize>,
) -> bool {
) -> Option<&ChatMessage> {
let mut next_index = message_index + 1;
while next_index < messages.len()
&& (is_following_user_content(messages, next_index)
@@ -441,10 +453,7 @@ fn next_rendered_entry_is_assistant_or_end(
next_index += 1;
}
messages
.get(next_index)
.map(|message| matches!(message, ChatMessage::Assistant { .. }))
.unwrap_or(true)
messages.get(next_index)
}
/// Render the tool preamble shown to the model for one DeepSeek dialect.
@@ -61,6 +61,7 @@ pub(super) fn render_request(request: &ChatRequest) -> Result<String> {
find_last_user_render_index(request.messages.as_slice(), render_offset);
let last_user_actual_index = find_last_user_actual_index(request.messages.as_slice());
let continue_final_message = request.chat_options.continue_final_message();
let add_generation_prompt = request.chat_options.add_generation_prompt();
let mut prompt = String::from(BOS_TOKEN);
if !initial_tools.is_empty() {
@@ -79,6 +80,7 @@ pub(super) fn render_request(request: &ChatRequest) -> Result<String> {
thinking_mode,
drop_thinking,
continue_final_message,
add_generation_prompt,
)?;
}
@@ -110,11 +112,13 @@ fn render_message(
thinking_mode: ThinkingMode,
drop_thinking: bool,
continue_final_message: bool,
add_generation_prompt: bool,
) -> Result<()> {
let render_index = message_index as isize + render_offset;
let opens_thinking = render_index == last_user_render_index;
let after_last_user_turn = render_index > last_user_render_index;
let after_or_at_last_user_turn = render_index >= last_user_render_index;
let add_generation_prompt = message_index + 1 != messages.len() || add_generation_prompt;
match message {
ChatMessage::System { content } => render_system_message(out, Some(content), &[]),
@@ -123,11 +127,13 @@ fn render_message(
content,
tools.as_deref().unwrap_or(&[]),
thinking_mode == ThinkingMode::Thinking && opens_thinking,
add_generation_prompt,
),
ChatMessage::User { content } => render_user_message(
out,
content,
thinking_mode == ThinkingMode::Thinking && opens_thinking,
add_generation_prompt,
),
ChatMessage::Assistant { content } => render_assistant_message(
out,
@@ -195,6 +201,7 @@ fn render_developer_message(
content: &ChatContent,
tools: &[ChatTool],
opens_thinking: bool,
add_generation_prompt: bool,
) -> Result<()> {
if content.is_empty() {
return Err(Error::ChatTemplate(
@@ -209,7 +216,7 @@ fn render_developer_message(
}
out.push_str("\n\n# The user's message is: ");
write_chat_content(out, content)?;
write_user_like_suffix(out, opens_thinking);
write_user_like_suffix(out, opens_thinking, add_generation_prompt);
Ok(())
}
@@ -219,17 +226,20 @@ fn render_user_message(
out: &mut String,
content: &ChatContent,
opens_thinking: bool,
add_generation_prompt: bool,
) -> Result<()> {
out.push_str("<|User|>");
write_chat_content(out, content)?;
write_user_like_suffix(out, opens_thinking);
write_user_like_suffix(out, opens_thinking, add_generation_prompt);
Ok(())
}
/// Shared trailing wrapper used by both real user turns and native developer
/// turns after their content has already been written.
// TODO: respect `add_generation_prompt` option
fn write_user_like_suffix(out: &mut String, opens_thinking: bool) {
fn write_user_like_suffix(out: &mut String, opens_thinking: bool, add_generation_prompt: bool) {
if !add_generation_prompt {
return;
}
out.push_str("<|Assistant|>");
if opens_thinking {
out.push_str(THINKING_START_TOKEN);
@@ -281,6 +281,44 @@ fn assistant_after_last_user_requires_reasoning_or_tool_calls() {
.assert_eq(&error.to_report_string());
}
#[test]
fn last_user_turn_appends_generation_prompt_by_default() {
let request = thinking_request(vec![ChatMessage::user("Hello")]);
let rendered = render_request(&request);
assert!(rendered.ends_with("<|User|>Hello<|Assistant|><think>"));
}
#[test]
fn last_user_turn_omits_generation_prompt_when_disabled() {
let mut request = thinking_request(vec![ChatMessage::user("Hello")]);
request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
let rendered = render_request(&request);
assert!(rendered.ends_with("<|User|>Hello"));
assert!(!rendered.contains("<|Assistant|>"));
assert!(!rendered.contains("<think>"));
}
#[test]
fn user_then_assistant_still_writes_suffix_without_generation_prompt() {
let mut request = ChatRequest {
messages: vec![
ChatMessage::user("Hello"),
ChatMessage::assistant_text("Hi there."),
],
..ChatRequest::for_test()
};
request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
let rendered = render_request(&request);
expect!["<|begin▁of▁sentence|><|User|>Hello<|Assistant|></think>Hi there.<|end▁of▁sentence|>"]
.assert_eq(&rendered);
}
#[test]
fn continue_final_assistant_omits_final_eos() {
let mut request = ChatRequest {
@@ -633,3 +633,94 @@ fn continue_final_assistant_omits_final_eos() {
expect!["<|begin▁of▁sentence|><|User|>write<|Assistant|></think>partial answer"]
.assert_eq(&rendered);
}
#[test]
fn last_user_turn_appends_generation_prompt_by_default() {
let request = thinking_request(vec![ChatMessage::user("Hello")]);
let rendered = render_request(&request);
assert!(rendered.ends_with("<|User|>Hello<|Assistant|><think>"));
}
#[test]
fn last_user_turn_omits_generation_prompt_when_disabled() {
let mut request = thinking_request(vec![ChatMessage::user("Hello")]);
request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
let rendered = render_request(&request);
assert!(rendered.ends_with("<|User|>Hello"));
assert!(!rendered.contains("<|Assistant|>"));
assert!(!rendered.contains("<think>"));
}
#[test]
fn user_then_assistant_still_writes_suffix_without_generation_prompt() {
let mut request = ChatRequest {
messages: vec![
ChatMessage::user("Hello"),
ChatMessage::assistant_text("Hi there."),
],
..ChatRequest::for_test()
};
request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
request
.chat_options
.template_kwargs
.insert("thinking".to_string(), Value::Bool(false));
let rendered = render_request(&request);
expect![
"<|begin▁of▁sentence|><|User|>Hello<|Assistant|></think>Hi there.<|end▁of▁sentence|>"
]
.assert_eq(&rendered);
}
#[test]
fn trailing_system_omits_generation_prompt_when_disabled() {
let mut request = ChatRequest {
messages: vec![
ChatMessage::system("rules"),
ChatMessage::user("question"),
ChatMessage::system("final rules"),
],
..ChatRequest::for_test()
};
request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
request
.chat_options
.template_kwargs
.insert("enable_thinking".to_string(), Value::Bool(false));
request.chat_options.reasoning_effort = Some(ReasoningEffort::Low);
let rendered = render_request(&request);
expect!["<|begin▁of▁sentence|>rules<|User|>questionfinal rules"].assert_eq(&rendered);
}
#[test]
fn last_merged_tool_result_omits_generation_prompt_when_disabled() {
let mut request = thinking_request(vec![
ChatMessage::user("question"),
ChatMessage::tool_response("result", "call-1"),
]);
request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
let rendered = render_request(&request);
expect!["<|begin▁of▁sentence|><|User|>question\n\n<tool_result>result</tool_result>"]
.assert_eq(&rendered);
assert!(!rendered.contains("<|Assistant|>"));
}
#[test]
fn last_developer_omits_generation_prompt_when_disabled() {
let mut request = thinking_request(vec![ChatMessage::developer("latest instruction", None)]);
request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
let rendered = render_request(&request);
expect!["<|begin▁of▁sentence|><|User|>latest instruction"].assert_eq(&rendered);
}
@@ -10,7 +10,9 @@ use super::DeepSeekV41ChatRenderer;
use crate::ChatRenderer;
use crate::event::{AssistantContentBlock, AssistantToolCall};
use crate::renderer::test_utils::{FixtureRequestOptions, fixture_chat_request};
use crate::request::{ChatContent, ChatContentPart, ChatMessage, ChatRequest, ReasoningEffort};
use crate::request::{
ChatContent, ChatContentPart, ChatMessage, ChatRequest, GenerationPromptMode, ReasoningEffort,
};
fn render(request: &ChatRequest) -> String {
DeepSeekV41ChatRenderer::new()
@@ -281,6 +283,49 @@ fn mid_system_advances_last_user_boundary_and_drops_old_reasoning() {
.assert_eq(&render(&request));
}
#[test]
fn last_user_turn_omits_generation_prompt_when_disabled() {
let mut request = request();
request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
expect!["<|begin▁of▁sentence|><|System|>Reasoning Effort: 50 (range 1-100, the higher the value, the more thorough the reasoning)\n\n<|User|>question"]
.assert_eq(&render(&request));
}
#[test]
fn last_mid_system_omits_generation_prompt_when_disabled() {
let mut request = ChatRequest {
messages: vec![
ChatMessage::system("sys"),
ChatMessage::user("q1"),
ChatMessage::assistant_blocks(vec![
AssistantContentBlock::Reasoning { text: "r1".into() },
AssistantContentBlock::Text { text: "a1".into() },
]),
ChatMessage::developer("old policy", None),
ChatMessage::system("mid sys"),
],
..ChatRequest::for_test()
};
request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
request
.chat_options
.template_kwargs
.insert("reasoning_effort".into(), json!(88));
expect!["<|begin▁of▁sentence|><|System|>Reasoning Effort: 88 (range 1-100, the higher the value, the more thorough the reasoning)\n\nsys<|User|>q1<|Assistant|></think>a1<|end▁of▁sentence|><|System|>mid sys"]
.assert_eq(&render(&request));
}
#[test]
fn chat_mode_omits_generation_prompt_when_disabled() {
let mut request = request();
request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
request.chat_options.template_kwargs.insert("thinking".into(), json!(false));
expect!["<|begin▁of▁sentence|><|User|>question"].assert_eq(&render(&request));
}
#[test]
fn joins_text_parts_with_reference_separator_and_preserves_literal_tags() {
let mut request = ChatRequest {