From 5492bc887e00726beedb1b908495325012642705 Mon Sep 17 00:00:00 2001 From: Warren B Date: Thu, 3 Sep 2026 02:16:21 +0100 Subject: [PATCH] webui: stop overriding engine and package defaults from model_params.json Every default in this file is sent on every request, so a value here is not a fallback the engine can improve on -- it is an override. Six of them fought the model. - vibevoice num_inference_steps was 10 while the shipped GGUF carries ddpm_num_inference_steps 20 (vibevoice/assets.cpp:112), so the UI silently halved diffusion quality. The default is omitted rather than corrected: the value is read per package, and any literal would override it again. - ace_step shift was 3.0 against an engine default of 1.0 (include/engine/models/ace_step/types.h:42, applied in ace_step/request_parser.cpp:231). - firered-audio-asr max_new_tokens was 512 against an engine default of 300, and sending it at all defeated the model's own bump to 1024 when enable_thinking is set (firered_audio/session.cpp:162). Omitted. - personaplex text_top_k follows top_k when unset (personaplex/request.cpp:111); pinning it to 250 prevented that. Omitted. - vevo2 top_k is read from the checkpoint's own generation config (vevo2/session.cpp:898); the literal 20 overrode it. Omitted. - fireredtts3 voice design carried the instruct path's guidance_scale of 2.0 instead of its own tuned 1.2 (fireredtts3/session.cpp:159, 170, 179). Omitting a default is how this file says "unset": the option parser skips empty values, so the control renders blank and sends nothing until the user types something. Validation: python3 -m json.tool webui/configs/model_params.json python3 tools/check_loader_catalog_sync.py # in sync --- webui/configs/model_params.json | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/webui/configs/model_params.json b/webui/configs/model_params.json index 7b3d71bbc..d098f459f 100644 --- a/webui/configs/model_params.json +++ b/webui/configs/model_params.json @@ -21,7 +21,7 @@ ], "vibevoice": [ - {"name": "num_inference_steps", "type": "number", "label": "num_inference_steps", "default": 10, "minimum": 1, "step": 1, "precision": 0, "info": "扩散步数（官方默认 10），越大越慢越稳"}, + {"name": "num_inference_steps", "type": "number", "label": "num_inference_steps", "minimum": 1, "step": 1, "precision": 0, "info": "留空=用模型自带的 ddpm_num_inference_steps（随包，7B 为 20）", "info_en": "Blank uses the packaged model's own ddpm_num_inference_steps (20 in the shipped 7B). Setting a value overrides it for every request."}, {"name": "guidance_scale", "type": "slider", "label": "guidance_scale", "default": 1.3, "minimum": 0.0, "maximum": 5.0, "step": 0.1, "info": "CFG 引导强度"}, {"name": "max_length_times", "type": "number", "label": "max_length_times", "default": 2.0, "minimum": 0.1, "step": 0.1, "info": "最大输出长度倍数"}, {"name": "temperature", "type": "slider", "label": "temperature", "default": 1.0, "minimum": 0.05, "maximum": 2.0, "step": 0.05}, @@ -136,7 +136,7 @@ "ace_step": [ {"name": "route", "type": "choice", "label": "route（操作类型）", "default": "text2music", "choices": ["text2music", "complete", "lego", "extract", "cover", "cover-nofsq", "repaint", "remix"], "info": "cover/remix=换词翻唱，非 text2music 需上传源音频；详见 webui/README.md"}, {"name": "num_inference_steps", "type": "number", "label": "num_inference_steps", "default": 8, "minimum": 1, "maximum": 20, "step": 1, "precision": 0, "info": "扩散步数（turbo 上限 20）；remix 路由不填时默认 16，其他路由默认 8"}, - {"name": "shift", "type": "slider", "label": "shift（时间步弯曲）", "default": 3.0, "minimum": 1.0, "maximum": 5.0, "step": 0.5, "info": "原版 turbo 默认 3.0；1.0 会明显劣化 remix 换词咬字"}, + {"name": "shift", "type": "slider", "label": "shift（时间步弯曲）", "label_en": "shift (timestep curve)", "default": 1.0, "minimum": 1.0, "maximum": 5.0, "step": 0.5, "info": "官方默认 1.0；extract 路由由模型自动改用 3.0", "info_en": "Model default is 1.0. The extract route sets 3.0 itself, so leaving this at 1.0 no longer overrides it."}, {"name": "guidance_scale", "type": "slider", "label": "guidance_scale", "default": 1.0, "minimum": 0.0, "maximum": 5.0, "step": 0.1}, {"name": "audio_cover_strength", "type": "slider", "label": "【cover】audio_cover_strength", "default": 1.0, "minimum": 0.0, "maximum": 1.0, "step": 0.05, "info": "1=贴近原曲，0=自由发挥；建议 0.5"}, {"name": "cover_noise_strength", "type": "slider", "label": "【cover】cover_noise_strength", "default": 0.0, "minimum": 0.0, "maximum": 1.0, "step": 0.05, "info": "保旋律强度；推荐 0.1~0.25"}, @@ -209,7 +209,7 @@ {"name": "temperature", "type": "slider", "label": "temperature", "default": 0.8, "minimum": 0.0, "maximum": 2.0, "step": 0.05}, {"name": "text_temperature", "type": "slider", "label": "text_temperature", "default": 0.8, "minimum": 0.0, "maximum": 2.0, "step": 0.05}, {"name": "top_k", "type": "number", "label": "top_k", "default": 250, "minimum": 0, "step": 1, "precision": 0}, - {"name": "text_top_k", "type": "number", "label": "text_top_k", "default": 250, "minimum": 0, "step": 1, "precision": 0}, + {"name": "text_top_k", "type": "number", "label": "text_top_k", "label_en": "text_top_k (blank follows top_k)", "minimum": 0, "step": 1, "precision": 0, "info_en": "Blank makes the text head follow top_k."}, {"name": "do_sample", "type": "bool", "label": "do_sample", "default": true} ], @@ -218,7 +218,7 @@ {"name": "num_inference_steps", "type": "number", "label": "num_inference_steps", "default": 32, "minimum": 1, "step": 1, "precision": 0, "info": "流匹配步数"}, {"name": "use_pitch_shift", "type": "choice", "label": "use_pitch_shift（自动音高对齐）", "default": "", "choices": ["", "true", "false"], "info": "留空=按路线默认"}, {"name": "temperature", "type": "slider", "label": "temperature（AR 路线用）", "default": 0.7, "minimum": 0.0, "maximum": 2.0, "step": 0.05, "info": "默认取自模型 generation_config.json"}, - {"name": "top_k", "type": "number", "label": "top_k（AR 路线用）", "default": 20, "minimum": 0, "step": 1, "precision": 0, "info": "默认取自模型 generation_config.json"}, + {"name": "top_k", "type": "number", "label": "top_k（AR 路线用）", "label_en": "top_k (AR paths)", "minimum": 0, "step": 1, "precision": 0, "info": "留空=取自模型 generation_config.json（回退 25）", "info_en": "Blank uses the checkpoint's generation_config.json (C++ fallback 25)."}, {"name": "top_p", "type": "slider", "label": "top_p（AR 路线用）", "default": 0.8, "minimum": 0.0, "maximum": 1.0, "step": 0.01} ], @@ -308,7 +308,7 @@ {"name": "template_name", "type": "choice", "label": "template_name", "default": "voice_design", "choices": ["voice_design"]}, {"name": "language", "type": "choice", "label": "language", "default": "Chinese", "choices": ["Chinese", "English", "Cantonese", "Japanese", "Korean", "Spanish", "French", "Russian", "Arabic", "Turkish", "Indonesian", "Portuguese", "Italian", "Dutch", "Vietnamese", "German", "Ukrainian", "Thai", "Polish", "Romanian", "Greek", "Czech", "Finnish", "Hindi", "ZH_Anhui", "ZH_Fujian", "ZH_Gansu", "ZH_Guizhou", "ZH_Hebei", "ZH_Henan", "ZH_Hubei", "ZH_Hunan", "ZH_Jiangxi", "ZH_Liaoning", "ZH_Minnan", "ZH_Ningxia", "ZH_Shaanxi", "ZH_Shandong", "ZH_Shanghai", "ZH_Shanxi", "ZH_Tianjin", "ZH_Wenzhou", "ZH_Wu", "ZH_Yunnan"]}, {"name": "num_inference_steps", "type": "number", "label": "num_inference_steps", "default": 10, "minimum": 1, "step": 1, "precision": 0}, - {"name": "guidance_scale", "type": "slider", "label": "guidance_scale", "default": 2.0, "minimum": 0.0, "maximum": 10.0, "step": 0.1}, + {"name": "guidance_scale", "type": "slider", "label": "guidance_scale", "default": 1.2, "minimum": 0.0, "maximum": 10.0, "step": 0.1, "info_en": "FireRedTTS3 voice design is tuned for 1.2; 2.0 is the instruct_tts value."}, {"name": "stop_threshold", "type": "slider", "label": "stop_threshold", "default": 0.5, "minimum": 0.0, "maximum": 1.0, "step": 0.05} ], @@ -354,7 +354,7 @@ {"name": "template_name", "type": "choice", "label": "template_name", "default": "asr", "choices": ["asr", "understand"]}, {"name": "language", "type": "choice", "label": "language", "default": "zh", "choices": ["zh", "en"]}, {"name": "enable_thinking", "type": "bool", "label": "enable_thinking", "default": false}, - {"name": "max_new_tokens", "type": "number", "label": "max_new_tokens", "default": 512, "minimum": 1, "step": 1, "precision": 0}, + {"name": "max_new_tokens", "type": "number", "label": "max_new_tokens", "label_en": "max_new_tokens", "minimum": 1, "step": 1, "precision": 0, "info_en": "Blank uses 300, or 1024 when enable_thinking is on. Setting a value here defeats the automatic thinking-mode bump."}, {"name": "top_k", "type": "number", "label": "top_k", "default": 20, "minimum": 0, "step": 1, "precision": 0}, {"name": "top_p", "type": "slider", "label": "top_p", "default": 0.8, "minimum": 0.0, "maximum": 1.0, "step": 0.01}, {"name": "temperature", "type": "slider", "label": "temperature", "default": 0.7, "minimum": 0.0, "maximum": 2.0, "step": 0.05}