NanoGPT: Add reasoning effort control Closes #4999

372db63cd537e7188dff7580b00f3293562fc6eb

Cohee <18619528+Cohee1207@users.noreply.github.com>

Parents: 618918535eefecfcbdd0566a75f62093881644b7
3 files changed, +14 -1Showing whitespace changes
public/index.html+1 -1
@@ -2110,7 +2110,7 @@
2110 </strong>2110 </strong>
2111 </div>2111 </div>
2112 </div>2112 </div>
2113 <div class="flex-container flexFlowColumn wide100p textAlignCenter marginTop10" data-source="openai,custom,claude,xai,makersuite,vertexai,aimlapi,openrouter,pollinations,perplexity,cometapi,electronhub,azure_openai,chutes">2113 <div class="flex-container flexFlowColumn wide100p textAlignCenter marginTop10" data-source="openai,custom,claude,xai,makersuite,vertexai,aimlapi,openrouter,pollinations,perplexity,cometapi,electronhub,azure_openai,chutes,nanogpt">
2114 <div class="flex-container oneline-dropdown" title="Constrains effort on reasoning for reasoning models.&#10;Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response." data-i18n="[title]Constrains effort on reasoning for reasoning models.">2114 <div class="flex-container oneline-dropdown" title="Constrains effort on reasoning for reasoning models.&#10;Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response." data-i18n="[title]Constrains effort on reasoning for reasoning models.">
2115 <label for="openai_reasoning_effort">2115 <label for="openai_reasoning_effort">
2116 <span data-i18n="Reasoning Effort">Reasoning Effort</span>2116 <span data-i18n="Reasoning Effort">Reasoning Effort</span>
src/constants.js+8 -0
@@ -482,6 +482,14 @@ export const OPENAI_REASONING_EFFORT_MAP = {
482 min: 'minimal',482 min: 'minimal',
483};483};
484484
485export const NANOGPT_REASONING_EFFORT_MAP = {
486 min: 'none',
487 low: 'minimal',
488 medium: 'low',
489 high: 'medium',
490 max: 'high',
491};
492
485export const LOG_LEVELS = {493export const LOG_LEVELS = {
486 DEBUG: 0,494 DEBUG: 0,
487 INFO: 1,495 INFO: 1,
src/endpoints/backends/chat-completions.js+5 -0
@@ -9,6 +9,7 @@ import {
9 AZURE_OPENAI_KEYS,9 AZURE_OPENAI_KEYS,
10 CHAT_COMPLETION_SOURCES,10 CHAT_COMPLETION_SOURCES,
11 GEMINI_SAFETY,11 GEMINI_SAFETY,
12 NANOGPT_REASONING_EFFORT_MAP,
12 OPENAI_REASONING_EFFORT_MAP,13 OPENAI_REASONING_EFFORT_MAP,
13 OPENAI_REASONING_EFFORT_MODELS,14 OPENAI_REASONING_EFFORT_MODELS,
14 OPENAI_VERBOSITY_MODELS,15 OPENAI_VERBOSITY_MODELS,
@@ -2228,6 +2229,10 @@ router.post('/generate', async function (request, response) {
2228 if (request.body.repetition_penalty !== undefined) {2229 if (request.body.repetition_penalty !== undefined) {
2229 bodyParams['repetition_penalty'] = request.body.repetition_penalty;2230 bodyParams['repetition_penalty'] = request.body.repetition_penalty;
2230 }2231 }
2232 if (request.body.reasoning_effort) {
2233 const effort = NANOGPT_REASONING_EFFORT_MAP[request.body.reasoning_effort];
2234 bodyParams['reasoning'] = { effort: effort };
2235 }
22312236
2232 const isClaude = /^claude-/.test(request.body.model);2237 const isClaude = /^claude-/.test(request.body.model);
2233 if (enableSystemPromptCache && isClaude) {2238 if (enableSystemPromptCache && isClaude) {