From 6cb8f95b5b7ec67b1ad530ef093597fa816edbba Mon Sep 17 00:00:00 2001
From: Concedo <39025047+LostRuins@users.noreply.github.com>
Date: Sun, 3 Aug 2025 16:21:57 +0800
Subject: [PATCH] tool calling params have been ported over to KAI api and can
be used, same syntax as OAI endpoint
---
kcpp_sdui.embd | 30 +++++++++++++++---------------
klite.embd | 4 ++--
koboldcpp.py | 29 +++++++++++++++++++++++++++--
3 files changed, 44 insertions(+), 19 deletions(-)
diff --git a/kcpp_sdui.embd b/kcpp_sdui.embd
index 111386c55..94c03812b 100644
--- a/kcpp_sdui.embd
+++ b/kcpp_sdui.embd
@@ -5,18 +5,18 @@
Stable UI for KoboldCpp
-
-
diff --git a/klite.embd b/klite.embd
index e4e90fb24..200cfcbf5 100644
--- a/klite.embd
+++ b/klite.embd
@@ -3435,7 +3435,7 @@ Current version indicated by LITEVER below.
max_length: (localflag?512:256),
auto_ctxlen: true,
auto_genamt: true,
- rep_pen: 1.07,
+ rep_pen: 1.06,
rep_pen_range: 360,
rep_pen_slope: 0.7,
temperature: 0.75,
@@ -3490,7 +3490,7 @@ Current version indicated by LITEVER below.
rep_pen_slope: defaultsettings.rep_pen_slope,
sampler_order: defaultsettings.sampler_order
},
- {"preset":"Simple Logical","description":"A very predictable preset with low randomness.","temp":0.3,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":100,"top_p":0.6,"min_p":0.0,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.02,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,0,1,3,4,2,5]},{"preset":"Simple Balanced","description":"A good balanced preset with medium randomness.","temp":0.75,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":100,"top_p":0.92,"min_p":0.0,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.07,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,0,1,3,4,2,5]},{"preset":"Simple Creative","description":"A wild and unpredictable preset with higher randomness.","temp":1.0,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":100,"top_p":0.98,"min_p":0.0,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.15,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,0,1,3,4,2,5]},{"preset":"Basic Min-P","description":"A good default for Min-P, only works on backends with min-p.","temp":1.25,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":0,"top_p":1,"min_p":0.1,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.03,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,5,0,1,3,4,2]},{"preset":"Basic Top-nsigma","description":"A good default for Top-nsigma, only works on backends with Top-nsigma.","temp":1,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":1.0,"top_k":0,"top_p":1,"min_p":0.01,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,5,0,1,3,4,2]},{"preset":"Basic DynaTemp","description":"A good default for DynaTemp, only works on backends with it.","temp":1.25,"dynatemp_range":0.75,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":0,"top_p":1,"min_p":0.05,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.03,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,5,0,1,3,4,2]},{"preset":"Basic SmoothSample","description":"A good default for Smooth Sampling, only works on backends with it.","temp":1.0,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.25,"nsigma":0.0,"top_k":0,"top_p":1,"min_p":0.05,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.03,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,5,0,1,3,4,2]},{"preset":"Basic SillyTavern","description":"Similar to default preset used in SillyTavern.","temp":0.75,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":40,"top_p":0.6,"min_p":0,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1.0,"rep_pen":1.18,"rep_pen_range":1024,"rep_pen_slope":0.8,"sampler_order":[6,0,1,3,4,2,5]},{"preset":"Immortal","description":"Modernized version of the Godlike preset, designed for creative and longer story co-writing use.","temp":0.7,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":1.75,"top_k":0,"top_p":1.0,"min_p":0.0,"presence_penalty":0.0,"top_a":0.75,"typical":0.19,"tfs":0.97,"rep_pen":1.1,"rep_pen_range":1024,"rep_pen_slope":0.7,"sampler_order":[6,5,4,3,2,1,0]},{"preset":"Neutral (Disabled)","description":"Sets all samplers neutralized, allowing you to customize your own.","temp":1.0,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":200,"top_p":1.0,"min_p":0.0,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.0,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,0,1,3,4,2,5]},{"preset":"CoherentCreativity (Legacy)","description":"Legacy preset. A good balance between coherence, creativity, and quality of prose.","rep_pen":1.2,"rep_pen_range":360,"rep_pen_slope":0,"sampler_order":[6,5,0,2,3,1,4],"temp":0.5,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"tfs":0.99,"top_a":0,"top_k":0,"top_p":1,"min_p":0.0,"presence_penalty":0.0,"typical":1},{"preset":"Godlike (Legacy)","description":"Legacy preset. Makes AI give a descriptive and sensual output.","temp":0.7,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":0,"top_p":0.5,"min_p":0.0,"presence_penalty":0.0,"top_a":0.75,"typical":0.19,"tfs":0.97,"rep_pen":1.1,"rep_pen_range":1024,"rep_pen_slope":0.7,"sampler_order":[6,5,4,3,2,1,0]},{"preset":"LiminalDrift (Legacy)","description":"Legacy preset. Sometimes surreal situations arise based on information already present in the story.","temp":0.66,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":0,"top_p":1,"min_p":0.0,"presence_penalty":0.0,"top_a":0.96,"typical":0.6,"tfs":1,"rep_pen":1.1,"rep_pen_range":1024,"rep_pen_slope":0.7,"sampler_order":[6,4,5,1,0,2,3]}
+ {"preset":"Simple Logical","description":"A very predictable preset with low randomness.","temp":0.3,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":100,"top_p":0.6,"min_p":0.0,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.02,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,0,1,3,4,2,5]},{"preset":"Simple Balanced","description":"A good balanced preset with medium randomness.","temp":0.75,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":100,"top_p":0.92,"min_p":0.0,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.07,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,0,1,3,4,2,5]},{"preset":"Simple Creative","description":"A wild and unpredictable preset with higher randomness.","temp":1.0,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":100,"top_p":0.98,"min_p":0.0,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.12,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,0,1,3,4,2,5]},{"preset":"Basic Min-P","description":"A good default for Min-P, only works on backends with min-p.","temp":1.25,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":0,"top_p":1,"min_p":0.1,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.03,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,5,0,1,3,4,2]},{"preset":"Basic Top-nsigma","description":"A good default for Top-nsigma, only works on backends with Top-nsigma.","temp":1,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":1.0,"top_k":0,"top_p":1,"min_p":0.01,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,5,0,1,3,4,2]},{"preset":"Basic DynaTemp","description":"A good default for DynaTemp, only works on backends with it.","temp":1.25,"dynatemp_range":0.75,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":0,"top_p":1,"min_p":0.05,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.03,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,5,0,1,3,4,2]},{"preset":"Basic SmoothSample","description":"A good default for Smooth Sampling, only works on backends with it.","temp":1.0,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.25,"nsigma":0.0,"top_k":0,"top_p":1,"min_p":0.05,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.03,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,5,0,1,3,4,2]},{"preset":"Basic SillyTavern","description":"Similar to default preset used in SillyTavern.","temp":0.75,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":40,"top_p":0.6,"min_p":0,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1.0,"rep_pen":1.18,"rep_pen_range":1024,"rep_pen_slope":0.8,"sampler_order":[6,0,1,3,4,2,5]},{"preset":"Immortal","description":"Modernized version of the Godlike preset, designed for creative and longer story co-writing use.","temp":0.7,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":1.75,"top_k":0,"top_p":1.0,"min_p":0.0,"presence_penalty":0.0,"top_a":0.75,"typical":0.19,"tfs":0.97,"rep_pen":1.1,"rep_pen_range":1024,"rep_pen_slope":0.7,"sampler_order":[6,5,4,3,2,1,0]},{"preset":"Neutral (Disabled)","description":"Sets all samplers neutralized, allowing you to customize your own.","temp":1.0,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":200,"top_p":1.0,"min_p":0.0,"presence_penalty":0.0,"top_a":0,"typical":1,"tfs":1,"rep_pen":1.0,"rep_pen_range":360,"rep_pen_slope":0.7,"sampler_order":[6,0,1,3,4,2,5]},{"preset":"CoherentCreativity (Legacy)","description":"Legacy preset. A good balance between coherence, creativity, and quality of prose.","rep_pen":1.2,"rep_pen_range":360,"rep_pen_slope":0,"sampler_order":[6,5,0,2,3,1,4],"temp":0.5,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"tfs":0.99,"top_a":0,"top_k":0,"top_p":1,"min_p":0.0,"presence_penalty":0.0,"typical":1},{"preset":"Godlike (Legacy)","description":"Legacy preset. Makes AI give a descriptive and sensual output.","temp":0.7,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":0,"top_p":0.5,"min_p":0.0,"presence_penalty":0.0,"top_a":0.75,"typical":0.19,"tfs":0.97,"rep_pen":1.1,"rep_pen_range":1024,"rep_pen_slope":0.7,"sampler_order":[6,5,4,3,2,1,0]},{"preset":"LiminalDrift (Legacy)","description":"Legacy preset. Sometimes surreal situations arise based on information already present in the story.","temp":0.66,"dynatemp_range":0.0,"dynatemp_exponent":1.0,"smoothing_factor":0.0,"nsigma":0.0,"top_k":0,"top_p":1,"min_p":0.0,"presence_penalty":0.0,"top_a":0.96,"typical":0.6,"tfs":1,"rep_pen":1.1,"rep_pen_range":1024,"rep_pen_slope":0.7,"sampler_order":[6,4,5,1,0,2,3]}
];
const instructpresets = [
diff --git a/koboldcpp.py b/koboldcpp.py
index 72b64f97e..b1efa84c2 100644
--- a/koboldcpp.py
+++ b/koboldcpp.py
@@ -2249,6 +2249,8 @@ def determine_tool_json_to_use(genparams, curr_ctx, assistant_message_start, is_
chosen_tool = genparams.get('tool_choice', "auto")
# first handle auto mode, determine whether a tool is needed
used_tool_json = None
+ if not curr_ctx:
+ return None
if tools_array and len(tools_array) > 0 and chosen_tool is not None and chosen_tool!="none":
tools_string = json.dumps(tools_array, indent=0)
should_use_tools = True
@@ -2313,7 +2315,6 @@ def determine_tool_json_to_use(genparams, curr_ctx, assistant_message_start, is_
return used_tool_json
-
def transform_genparams(genparams, api_format):
global chatcompl_adapter, maxctx
@@ -2363,7 +2364,31 @@ ws ::= | " " | "\n" [ \t]{0,20}
genparams["max_length"] = int(genparams.get('max', args.defaultgenamt))
elif api_format==2:
- pass
+ #tool calls only possible if forced, or if ending with assistant tag
+ adapter_obj = {} if chatcompl_adapter is None else chatcompl_adapter
+ assistant_message_start = adapter_obj.get("assistant_start", "\n### Response:\n")
+ used_tool_json = determine_tool_json_to_use(genparams, genparams.get('prompt', ""), assistant_message_start, True)
+ if used_tool_json and not genparams.get('grammar', ""):
+ toolparamjson = None
+ toolname = None
+ # Set temperature lower automatically if function calling, cannot exceed 0.5
+ genparams["temperature"] = (1.0 if genparams.get("temperature", 0.5) > 1.0 else genparams.get("temperature", 0.5))
+ genparams["using_openai_tools"] = True
+ # Set grammar to llamacpp example grammar to force json response (see https://github.com/ggerganov/llama.cpp/blob/master/grammars/json_arr.gbnf)
+ genparams["grammar"] = jsongrammar
+ try:
+ toolname = used_tool_json.get('function').get('name')
+ toolparamjson = used_tool_json.get('function').get('parameters')
+ bettergrammarjson = {"type":"array","items":{"type":"object","properties":{"id":{"type":"string","enum":["call_001"]},"type":{"type":"string","enum":["function"]},"function":{"type":"object","properties":{"name":{"type":"string"},"arguments":{}},"required":["name","arguments"],"additionalProperties":False}},"required":["id","type","function"],"additionalProperties":False}}
+ bettergrammarjson["items"]["properties"]["function"]["properties"]["arguments"] = toolparamjson
+ decoded = convert_json_to_gbnf(bettergrammarjson)
+ if decoded:
+ genparams["grammar"] = decoded
+ except Exception:
+ pass
+ tool_json_formatting_instruction = f"\nPlease use the provided schema to fill the parameters to create a function call for {toolname}, in the following format: " + json.dumps([{"id": "call_001", "type": "function", "function": {"name": f"{toolname}", "arguments": {"first property key": "first property value", "second property key": "second property value"}}}], indent=0)
+ genparams["prompt"] += f"\n\nJSON Schema:\n{used_tool_json}\n\n{tool_json_formatting_instruction}{assistant_message_start}"
+
elif api_format==3 or api_format==4 or api_format==7:
default_adapter = {} if chatcompl_adapter is None else chatcompl_adapter