ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

server: update deepseek reasoning format (pass reasoning_content as diffs) (#13933)

Browse source 
* server: update deepseek reasoning format (now in reasoning_content diffs), add legacy option for compat
* update unit/test_tool_call.py::test_thoughts
This commit is contained in:
Olivier Chafik 2025-06-02 10:15:44 -07:00 • committed by GitHub
parent bfd322796c
commit c9bbc77931
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
8 changed files with 30 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

2
tools/server/server.cpp
View file

@ -360,7 +360,7 @@ struct server_task {
params.oaicompat_chat_syntax.format = defaults.oaicompat_chat_syntax.format;
}
params.oaicompat_chat_syntax.reasoning_format = params_base.reasoning_format;
params.oaicompat_chat_syntax.reasoning_in_content = params.stream;
params.oaicompat_chat_syntax.reasoning_in_content = params.stream && (params_base.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY);
params.oaicompat_chat_syntax.thinking_forced_open = json_value(data, "thinking_forced_open", false);
params.oaicompat_chat_syntax.parse_tool_calls = json_value(data, "parse_tool_calls", false);
}

13
tools/server/tests/unit/test_tool_call.py
View file

@ -499,13 +499,12 @@ def do_test_calc_result(server: ServerProcess, result_override: str | None, n_pr
@pytest.mark.slow
@pytest.mark.parametrize("n_predict,reasoning_format,stream,expect_reasoning_content,expect_content,hf_repo,template_override", [
(128, 'deepseek', CompletionMode.NORMAL, None, "^The sum of 102 and 7 is 109[\\s\\S]*", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(128, None, CompletionMode.NORMAL, None, "^The sum of 102 and 7 is 109[\\s\\S]*", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(1024, 'deepseek', CompletionMode.NORMAL, "I need to calculate the sum of 102 and 7[\\s\\S]*", "To find the sum of[\\s\\S]*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, 'deepseek', CompletionMode.STREAMED, None, "^<think>I need to calculate [\\s\\S]*?</think>To find the sum of [\\s\\S]*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, 'deepseek', CompletionMode.NORMAL, "First, I [\\s\\S]*", "To find the sum of[\\s\\S]*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
(1024, 'deepseek', CompletionMode.STREAMED, None, "^<think>First, I [\\s\\S]*?</think>To find the sum of[\\s\\S]*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
@pytest.mark.parametrize("stream", [CompletionMode.NORMAL, CompletionMode.STREAMED])
@pytest.mark.parametrize("n_predict,reasoning_format,expect_reasoning_content,expect_content,hf_repo,template_override", [
(128, 'deepseek', None, "^The sum of 102 and 7 is 109[\\s\\S]*", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(128, None, None, "^The sum of 102 and 7 is 109[\\s\\S]*", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(1024, 'deepseek', "I need to calculate the sum of 102 and 7[\\s\\S]*", "To find the sum of[\\s\\S]*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, 'deepseek', "First, I [\\s\\S]*", "To find the sum of[\\s\\S]*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
# (1024, 'none', CompletionMode.NORMAL, None, "^(<think>\\s*)?I need[\\s\\S]*?</think>\\s*To find[\\s\\S]*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
# (128, 'deepseek', None, "^Okay, let me figure out the sum of 102 and 7[\\s\\S]*", "bartowski/Qwen_QwQ-32B-GGUF:Q4_K_M", None),
])

11
tools/server/tests/utils.py
View file

@ -308,10 +308,12 @@ class ServerProcess:
stream = data.get('stream', False)
if stream:
content: list[str] = []
reasoning_content: list[str] = []
tool_calls: list[dict] = []
finish_reason: Optional[str] = None
content_parts = 0
reasoning_content_parts = 0
tool_call_parts = 0
arguments_parts = 0
@ -322,6 +324,10 @@ class ServerProcess:
assert len(choice['delta']['content']) > 0, f'Expected non empty content delta!'
content.append(choice['delta']['content'])
content_parts += 1
if choice['delta'].get('reasoning_content') is not None:
assert len(choice['delta']['reasoning_content']) > 0, f'Expected non empty reasoning_content delta!'
reasoning_content.append(choice['delta']['reasoning_content'])
reasoning_content_parts += 1
if choice['delta'].get('finish_reason') is not None:
finish_reason = choice['delta']['finish_reason']
for tc in choice['delta'].get('tool_calls', []):
@ -349,8 +355,10 @@ class ServerProcess:
tool_call['function']['name'] = tool_call['function'].get('name', '') + fct['name']
if fct.get('arguments') is not None:
tool_call['function']['arguments'] += fct['arguments']
arguments_parts += 1
tool_call_parts += 1
print(f'Streamed response had {content_parts} content parts, {tool_call_parts} tool call parts incl. {arguments_parts} arguments parts')
print(f'Streamed response had {content_parts} content parts, {reasoning_content_parts} reasoning_content parts, {tool_call_parts} tool call parts incl. {arguments_parts} arguments parts')
result = dict(
choices=[
dict(
@ -359,6 +367,7 @@ class ServerProcess:
message=dict(
role='assistant',
content=''.join(content) if content else None,
reasoning_content=''.join(reasoning_content) if reasoning_content else None,
tool_calls=tool_calls if tool_calls else None,
),
)