Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-18 16:55:14 +02:00
Code Issues Packages Projects Releases Wiki Activity

reasoning output parsing improvements

Browse Source
This commit is contained in:
Concedo
2026-03-29 23:35:24 +08:00
parent 4a09f3805b
commit 4fc3c28f1a
2 changed files with 57 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files
embd_res/klite.embd
+3 -2
View File
File diff suppressed because one or more lines are too long
koboldcpp.py
+54 -24
View File
@@ -3464,7 +3464,7 @@ ws ::= | " " | "\n" [ \t]{0,20}
if jinja_output:
messages_string = jinja_output
if jinja_output.rstrip().endswith("<think>"): #the prompt template already forced a start think.
genparams["jinja_already_started_thinking"] = True
genparams["already_started_thinking"] = True
if jinjatools and len(jinjatools)>0:
genparams["using_openai_tools"] = True
# handle media
@@ -3581,6 +3581,8 @@ ws ::= | " " | "\n" [ \t]{0,20}
if (latest_turn_was_assistant and continue_assistant_turn): #allow continue a prefill, chop off end
messages_string = messages_string[:-(len(assistant_message_gen)+len(assistant_message_end))]
genparams["prompt"] = messages_string
if messages_string.rstrip().endswith("<think>"): #the prompt template already forced a start think.
genparams["already_started_thinking"] = True
if len(images_added)>0:
genparams["images"] = images_added
if len(audio_added)>0:
@@ -4253,7 +4255,7 @@ class KcppServerRequestHandler(http.server.SimpleHTTPRequestHandler):
return
encap_in_thinking = False
if genparams.get('jinja_already_started_thinking', False):
if genparams.get('already_started_thinking', False):
encap_in_thinking = True
encap_first_loop = True
thinkpairs = [{"start":"<|channel|>analysis<|message|>","end":"<|start|>assistant<|channel|>final<|message|>"},
@@ -4314,32 +4316,60 @@ class KcppServerRequestHandler(http.server.SimpleHTTPRequestHandler):
if tokenStr!="" or streamDone:
need_split_final_msg = True if (currfinishreason is not None and streamDone and tokenStr!="") else False
# hack for lcppui reasoning_content for thinking models
delta = {'role':'assistant','content':tokenStr}
# Hack for lcppui reasoning_content for thinking models
delta = {'role': 'assistant'}
if genparams.get('encapsulate_thinking', True):
for pair in thinkpairs:
if encap_first_loop and not encap_in_thinking and genparams.get("prompt","").endswith(pair["start"]):
encap_in_thinking = True
delta = {'role':'assistant','reasoning_content':tokenStr}
thinkpairs = [pair] #remove all others
break
elif not encap_in_thinking and (pair["start"] in tokenStr):
encap_in_thinking = True
out1, out2 = tokenStr.split(pair["start"], 1)
delta = {'role':'assistant','reasoning_content':out2}
thinkpairs = [pair] #remove all others
break
elif encap_in_thinking and pair["end"] in tokenStr:
if encap_in_thinking:
# We are already inside a thinking block. thinkpairs has already been reduced to [pair], so we just check the active one.
active_pair = thinkpairs[0]
if active_pair["end"] in tokenStr:
encap_in_thinking = False
out1, out2 = tokenStr.split(pair["end"], 1)
delta = {'role':'assistant','reasoning_content':out1,'content':out2}
thinkpairs = [pair] #remove all others
break
elif encap_in_thinking:
delta = {'role':'assistant','reasoning_content':tokenStr}
out1, out2 = tokenStr.split(active_pair["end"], 1)
if out1:
delta['reasoning_content'] = out1
if out2:
delta['content'] = out2
else:
delta = {'role':'assistant','content':tokenStr}
# Still thinking
delta['reasoning_content'] = tokenStr
else:
# Not thinking. Let's see if a start tag appears in this chunk.
matched_start = False
for pair in thinkpairs:
# Condition A: The prompt ended exactly with the start tag
if encap_first_loop and genparams.get("prompt", "").endswith(pair["start"]):
encap_in_thinking = True
thinkpairs = [pair] # lock in this pair
delta['reasoning_content'] = tokenStr
matched_start = True
break
# Condition B: The start tag is inside this chunk
elif pair["start"] in tokenStr:
encap_in_thinking = True
thinkpairs = [pair] # lock in this pair
out1, out2 = tokenStr.split(pair["start"], 1)
# Preserve text that came BEFORE the start tag
if out1:
delta['content'] = out1
if out2:
delta['reasoning_content'] = out2
# Edge Case: The end tag is ALSO in this exact same chunk (in out2)
if pair["end"] in out2:
encap_in_thinking = False
out2_think, out2_content = out2.split(pair["end"], 1)
# Overwrite reasoning with the exact thinking part
delta['reasoning_content'] = out2_think
# Append anything after the end tag to the content part
if out2_content:
delta['content'] = delta.get('content', '') + out2_content
matched_start = True
break
# Condition C: No start tag found, just normal text
if not matched_start:
delta['content'] = tokenStr
encap_first_loop = False
else:
delta['content'] = tokenStr
if need_split_final_msg: #we need to send one message without the finish reason, then send a finish reason with no msg to follow standards
if api_format == 4: # if oai chat, set format to expected openai streaming response