@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 04f50ad9d1
Fork 0

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 15, 23:30 UTC

[Bugfix] deepseek_r1_reasoning_parser put reason content in wrong field in certain edge case (#13097)

LikeSundayLikeRain committed Feb 13, 2025 at 02:11 UTC 04f50ad9d1e1fd3f2d9594d5dbfd1d51d37bfea0
2 files changed +6 -6
tests/entrypoints/openai/reasoning_parsers/test_deepseekr1_reasoning_parser.py
+5 -5
@@ -24,10 +24,10 @@ COMPLETE_REASONING = {
24 "reasoning_content": "This is a reasoning section",
25 "content": None,
26 }
27 -NO_REASONING = {
27 +NO_CONTENT = {
28 "output": "This is content",
29 - "reasoning_content": None,
30 - "content": "This is content",
29 + "reasoning_content": "This is content",
30 + "content": None,
31 }
32 NO_REASONING_STREAMING = {
33 "output": "This is a reasoning section",
@@ -98,8 +98,8 @@ TEST_CASES = [
98 ),
99 pytest.param(
100 False,
101 - NO_REASONING,
102 - id="no_reasoning_token",
101 + NO_CONTENT,
102 + id="no_content_token",
103 ),
104 pytest.param(
105 True,
vllm/entrypoints/openai/reasoning_parsers/deepseek_r1_reasoning_parser.py
+1 -1
@@ -128,7 +128,7 @@ class DeepSeekR1ReasoningParser(ReasoningParser):
128 # Thus we assume the reasoning content is always at the start.
129 # Ref https://huggingface.co/deepseek-ai/DeepSeek-R1/commit/8a58a132790c9935686eb97f042afa8013451c9f
130 if self.think_end_token not in model_output:
131 - return None, model_output
131 + return model_output, None
132 else:
133 # Add a start token if it's missing to keep compatibility.
134 if self.think_start_token not in model_output: