From 75f06d3a4a372c56c9c2f6545b63900bd83b857b Mon Sep 17 00:00:00 2001 From: Will Burford Date: Tue, 14 Jul 2026 15:12:08 -0400 Subject: [PATCH 1/4] Add reasoning budget message localization --- en/config.json | 2 ++ 1 file changed, 2 insertions(+) diff --git a/en/config.json b/en/config.json index 90027cbc..51aa4072 100644 --- a/en/config.json +++ b/en/config.json @@ -134,6 +134,8 @@ "llm.load.useUnifiedKvCache/subTitle": "Controls whether concurrent predictions share a single KV cache, saving memory. Disabling this ensures each prediction can utilize the full context length, at the cost of using more memory", "llm.load.llama.contextCheckpoints/title": "Context Checkpoints", "llm.load.llama.contextCheckpoints/subTitle": "Maximum number of context checkpoints to keep per slot. Lower values reduce RAM usage for SWA/recurrent models. 0 disables checkpoints.", + "llm.load.llama.reasoningBudgetMessage/title": "Reasoning Budget Message", + "llm.load.llama.reasoningBudgetMessage/subTitle": "Text emitted before the inferred end-of-reasoning sequence when a reasoning budget is exhausted. Leave empty to emit no message. Changing this requires reloading the model.", "load.gpuStrictVramCap/title": "Limit Model Offload to Dedicated GPU Memory", "load.gpuStrictVramCap.customSubTitleOff": "OFF: Allow model weights to offload to shared memory if dedicated GPU memory is full", "load.gpuStrictVramCap.customSubTitleOn": "ON: The system will limit offload of model weights to dedicated GPU memory and RAM only. Context may still use shared memory", From c234141bf145b673f4b96390594925ad7a336c35 Mon Sep 17 00:00:00 2001 From: Will Burford Date: Tue, 14 Jul 2026 16:10:52 -0400 Subject: [PATCH 2/4] Add reasoning budget localization --- en/config.json | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/en/config.json b/en/config.json index 51aa4072..dd028cfb 100644 --- a/en/config.json +++ b/en/config.json @@ -59,6 +59,10 @@ "llm.prediction.maxPredictedTokens/info": "Control the max length of the chatbot's response. Turn on to set a limit on the max length of a response, or turn off to let the chatbot decide when to stop.", "llm.prediction.maxPredictedTokens/inputLabel": "Maximum response length (tokens)", "llm.prediction.maxPredictedTokens/wordEstimate": "About {{maxWords}} words", + "llm.prediction.reasoning.budgetTokens/title": "Reasoning Budget", + "llm.prediction.reasoning.budgetTokens/subTitle": "Optionally limit the number of tokens generated inside the reasoning section", + "llm.prediction.reasoning.budgetTokens/info": "When enabled, llama-server ends the current reasoning section after this many generated reasoning tokens and continues with the visible answer. Requires a compatible reasoning chat template.", + "llm.prediction.reasoning.budgetTokens/inputLabel": "Reasoning budget (tokens)", "llm.prediction.repeatPenalty/title": "Repeat Penalty", "llm.prediction.repeatPenalty/subTitle": "How much to discourage repeating the same token", "llm.prediction.repeatPenalty/info": "From llama.cpp help docs: \"Helps prevent the model from generating repetitive or monotonous text.\n\nA higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient.\" • The default value is <{{dynamicValue}}>", @@ -565,6 +569,7 @@ "llamaKvCacheQuantizationWarning": "KV Cache Quantization is an experimental feature that may cause issues with some models. Flash Attention must be enabled for V cache quantization. If you encounter problems, reset to the default \"F16\".", "seedUncheckedHint": "Random Seed", + "reasoningBudgetUncheckedHint": "Unrestricted", "ropeFrequencyBaseUncheckedHint": "Auto", "ropeFrequencyScaleUncheckedHint": "Auto", From c14222e68ed3738f0fac2b6ac2c514ee0ac0e2d3 Mon Sep 17 00:00:00 2001 From: Will Burford Date: Tue, 14 Jul 2026 16:28:43 -0400 Subject: [PATCH 3/4] Align reasoning budget help text --- en/config.json | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/en/config.json b/en/config.json index dd028cfb..47549c11 100644 --- a/en/config.json +++ b/en/config.json @@ -60,9 +60,8 @@ "llm.prediction.maxPredictedTokens/inputLabel": "Maximum response length (tokens)", "llm.prediction.maxPredictedTokens/wordEstimate": "About {{maxWords}} words", "llm.prediction.reasoning.budgetTokens/title": "Reasoning Budget", - "llm.prediction.reasoning.budgetTokens/subTitle": "Optionally limit the number of tokens generated inside the reasoning section", - "llm.prediction.reasoning.budgetTokens/info": "When enabled, llama-server ends the current reasoning section after this many generated reasoning tokens and continues with the visible answer. Requires a compatible reasoning chat template.", - "llm.prediction.reasoning.budgetTokens/inputLabel": "Reasoning budget (tokens)", + "llm.prediction.reasoning.budgetTokens/subTitle": "Token budget for thinking: off for unrestricted, 0 for immediate end, N > 0 for token budget", + "llm.prediction.reasoning.budgetTokens/info": "From llama.cpp help docs:\n\nToken budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1).", "llm.prediction.repeatPenalty/title": "Repeat Penalty", "llm.prediction.repeatPenalty/subTitle": "How much to discourage repeating the same token", "llm.prediction.repeatPenalty/info": "From llama.cpp help docs: \"Helps prevent the model from generating repetitive or monotonous text.\n\nA higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient.\" • The default value is <{{dynamicValue}}>", @@ -139,7 +138,7 @@ "llm.load.llama.contextCheckpoints/title": "Context Checkpoints", "llm.load.llama.contextCheckpoints/subTitle": "Maximum number of context checkpoints to keep per slot. Lower values reduce RAM usage for SWA/recurrent models. 0 disables checkpoints.", "llm.load.llama.reasoningBudgetMessage/title": "Reasoning Budget Message", - "llm.load.llama.reasoningBudgetMessage/subTitle": "Text emitted before the inferred end-of-reasoning sequence when a reasoning budget is exhausted. Leave empty to emit no message. Changing this requires reloading the model.", + "llm.load.llama.reasoningBudgetMessage/subTitle": "Message injected before the end-of-thinking tag when reasoning budget is exhausted (default: none)", "load.gpuStrictVramCap/title": "Limit Model Offload to Dedicated GPU Memory", "load.gpuStrictVramCap.customSubTitleOff": "OFF: Allow model weights to offload to shared memory if dedicated GPU memory is full", "load.gpuStrictVramCap.customSubTitleOn": "ON: The system will limit offload of model weights to dedicated GPU memory and RAM only. Context may still use shared memory", From cb86dded110598cb647cbc628376f8ac1925c2f5 Mon Sep 17 00:00:00 2001 From: Will Burford Date: Tue, 14 Jul 2026 16:30:56 -0400 Subject: [PATCH 4/4] Remove unused reasoning budget info text --- en/config.json | 1 - 1 file changed, 1 deletion(-) diff --git a/en/config.json b/en/config.json index 47549c11..4cd5fe6e 100644 --- a/en/config.json +++ b/en/config.json @@ -61,7 +61,6 @@ "llm.prediction.maxPredictedTokens/wordEstimate": "About {{maxWords}} words", "llm.prediction.reasoning.budgetTokens/title": "Reasoning Budget", "llm.prediction.reasoning.budgetTokens/subTitle": "Token budget for thinking: off for unrestricted, 0 for immediate end, N > 0 for token budget", - "llm.prediction.reasoning.budgetTokens/info": "From llama.cpp help docs:\n\nToken budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1).", "llm.prediction.repeatPenalty/title": "Repeat Penalty", "llm.prediction.repeatPenalty/subTitle": "How much to discourage repeating the same token", "llm.prediction.repeatPenalty/info": "From llama.cpp help docs: \"Helps prevent the model from generating repetitive or monotonous text.\n\nA higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient.\" • The default value is <{{dynamicValue}}>",