A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 19:00 UTC
Add chat prefix completion feature to DeepSeek v3.2 (#31147)
CHENYUE committed
Jan 5, 2026 at 11:20 UTC
346e56455a3b10787c7d6e951861317953723185
1 file changed
+9
-5
vllm/tokenizers/deepseek_v32_encoding.py
+9
-5
@@ -169,6 +169,7 @@ def render_message(
169
response_format = msg.get("response_format")
170
tool_calls = msg.get("tool_calls")
171
reasoning_content = msg.get("reasoning") or msg.get("reasoning_content")
172
+ is_prefix = msg.get("prefix", False)
173
174
if tools:
175
tools = tools_from_openai_format(tools)
@@ -273,11 +274,14 @@ def render_message(
274
+ thinking_end_token
275
)
276
276
- prompt += assistant_msg_template.format(
277
- reasoning=thinking_part,
278
- content=summary_content,
279
- tool_calls=tool_calls_content,
280
- )
277
+ if not tool_calls and is_prefix:
278
+ prompt += summary_content
279
+ else:
280
+ prompt += assistant_msg_template.format(
281
+ reasoning=thinking_part,
282
+ content=summary_content,
283
+ tool_calls=tool_calls_content,
284
+ )
285
else:
286
raise NotImplementedError(f"Unknown role: {role}")
287