Marcel Hofer
ec7cd7f84d
- Added Qwen3.8-27B-ABLITERATED-GGUF model to the model catalog with detailed specifications and context length. - Improved duration widget behavior to ensure it remains hidden for non-video presets and preserves last valid values. - Updated documentation to clarify duration handling and widget visibility. - Expanded test coverage for model catalog labels and Qwen3.8 model presence in both text and vision catalogs. - Refactored model access functions to better manage installed models and local paths. - Enhanced tooltips and recommendations for model selection, emphasizing installed and local models. - Fixed issues with dynamic duration-widget serialization and visibility in the frontend.
276 lines
9.7 KiB
JSON
276 lines
9.7 KiB
JSON
{
|
|
"version": 1,
|
|
"rules": [
|
|
{
|
|
"id": "qwen38_blackfrost_27b",
|
|
"priority": 110,
|
|
"match_any": [
|
|
"Qwen3.8-27B-ABLITERATED"
|
|
],
|
|
"family_label": "Qwen3.8 27B Blackfrost abliterated",
|
|
"sources": [
|
|
"Blackfrost-AI Qwen3.8 27B Abliterated GGUF model card",
|
|
"Qwen3.8 27B official model card"
|
|
],
|
|
"support_notes": {
|
|
"hf": "This catalog entry is GGUF-only. Use a GGUF node for this repository.",
|
|
"gguf": "Hard toggle via enable_thinking in the Qwen3.8 chat template. Vision requires the matching mmproj file and a current llama.cpp build.",
|
|
"soft_switch": "Qwen3.8 uses chat-template thinking controls; no /think or /no_think soft switch is documented."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 262144
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 0.7,
|
|
"top_p": 0.8,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 262144
|
|
},
|
|
"notes": [
|
|
"Thinking is enabled by default; disable it for direct prompt rewrites and keep it enabled for deeper scene planning.",
|
|
"Q4_K_M is the model author's balanced default; use the compact Q8 projector for image or video input.",
|
|
"The model supports 262K context architecturally, but large llama.cpp contexts consume substantial KV-cache memory; lower ctx for ordinary prompt work."
|
|
]
|
|
},
|
|
{
|
|
"id": "qwen36_hauhau_multimodal",
|
|
"priority": 100,
|
|
"match_any": [
|
|
"Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive"
|
|
],
|
|
"family_label": "Qwen3.6 35B A3B HauhauCS aggressive",
|
|
"sources": [
|
|
"HauhauCS model card",
|
|
"Qwen3.5 official model card"
|
|
],
|
|
"support_notes": {
|
|
"hf": "Not in the HF lineup here. Use the GGUF nodes for this model.",
|
|
"gguf": "Hard toggle via enable_thinking. Pair the vision node with the matching mmproj file.",
|
|
"soft_switch": "No official /think or /no_think soft switch for this family."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 131072
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 0.7,
|
|
"top_p": 0.8,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 131072
|
|
},
|
|
"notes": [
|
|
"For creative prompt generation, use the general-task sampler: thinking ON for richer scene planning, thinking OFF for shorter direct rewrites.",
|
|
"Q4_K_P is the best quality-size starting point; move to Q6_K_P or Q8_K_P when VRAM allows.",
|
|
"Keep a large context when expanding long image/video prompt sequences so character and scene continuity are not lost."
|
|
]
|
|
},
|
|
{
|
|
"id": "qwen35_family",
|
|
"priority": 80,
|
|
"match_any": [
|
|
"Qwen3.5",
|
|
"qwen3.5"
|
|
],
|
|
"family_label": "Qwen3.5 family",
|
|
"sources": [
|
|
"Qwen3.5 official model card"
|
|
],
|
|
"support_notes": {
|
|
"hf": "Hard toggle via enable_thinking when the backend honors the chat template.",
|
|
"gguf": "Hard toggle via enable_thinking in the Qwen chat template.",
|
|
"soft_switch": "No official /think or /no_think soft switch for Qwen3.5."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 131072
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 0.7,
|
|
"top_p": 0.8,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 131072
|
|
},
|
|
"notes": [
|
|
"For creative prompt generation, Qwen's general-task settings are the right fit: thinking ON for richer planning, non-thinking for concise prompt rewrites.",
|
|
"Use the larger output budget for multi-shot scene descriptions, storyboard continuity, or detailed style transfer prompts.",
|
|
"If generations loop or over-elaborate, prefer a shorter custom prompt and lower max_tokens before changing the sampler."
|
|
]
|
|
},
|
|
{
|
|
"id": "qwen3_family",
|
|
"priority": 70,
|
|
"match_any": [
|
|
"Qwen3VL-",
|
|
"qwen3vl-",
|
|
"Qwen3-VL",
|
|
"Qwen3-",
|
|
"qwen3-"
|
|
],
|
|
"exclude_any": [
|
|
"Qwen3.5",
|
|
"qwen3.5",
|
|
"Qwen3.6",
|
|
"qwen3.6"
|
|
],
|
|
"family_label": "Qwen3 family",
|
|
"sources": [
|
|
"Qwen3 official model card",
|
|
"Qwen Transformers docs"
|
|
],
|
|
"support_notes": {
|
|
"hf": "Hard toggle via enable_thinking, with /think and /no_think available as soft switches.",
|
|
"gguf": "Hard toggle via enable_thinking. /think and /no_think are available only on Qwen3 chat-template builds.",
|
|
"soft_switch": "Qwen3 officially supports /think and /no_think when thinking is enabled."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 0.6,
|
|
"top_p": 0.95,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 32768
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 0.7,
|
|
"top_p": 0.8,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 32768
|
|
},
|
|
"notes": [
|
|
"For creative prompt drafting, use thinking ON when you want structured scene planning and thinking OFF when you want a fast final prompt.",
|
|
"Provider recommends 32K output budget for most tasks and only enabling YaRN when you really need longer than the native 32K context.",
|
|
"Avoid greedy sampling for creative generation; the recommended sampler gives more varied but still coherent prompt language."
|
|
]
|
|
},
|
|
{
|
|
"id": "gemma4_12b_unified",
|
|
"priority": 65,
|
|
"match_any": [
|
|
"Gemma-4-12B",
|
|
"gemma-4-12b"
|
|
],
|
|
"family_label": "Gemma 4 12B Unified",
|
|
"sources": [
|
|
"Google Gemma 4 model overview",
|
|
"Unsloth Gemma 4 12B GGUF model card"
|
|
],
|
|
"support_notes": {
|
|
"hf": "Gemma 4 12B is a unified multimodal model. HF support depends on current Transformers processor support for image/audio/video inputs.",
|
|
"gguf": "Use Gemma4ChatHandler in a recent multimodal llama-cpp-python build. Pair the model with mmproj-F16.gguf.",
|
|
"audio": "Audio input is intended for Gemma 4 E2B, E4B, and 12B only. Use a recent Gemma4ChatHandler build; 16 kHz mono WAV-style audio is safest.",
|
|
"soft_switch": "Gemma uses configurable thinking controls instead of Qwen-style /think or /no_think soft switches."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 64,
|
|
"max_tokens": 1024,
|
|
"ctx_min": 262144,
|
|
"image_max_tokens": 4096
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 64,
|
|
"max_tokens": 1024,
|
|
"ctx_min": 262144,
|
|
"image_max_tokens": 4096
|
|
},
|
|
"notes": [
|
|
"For creative prompt generation, Gemma 4's default temperature/top_p is intentionally open enough for varied visual language.",
|
|
"For audio analysis, 16 kHz mono WAV is the safest input shape.",
|
|
"Prefer Q4_K_M or UD-Q4_K_XL first; move up to Q6/Q8/BF16 when VRAM allows."
|
|
]
|
|
},
|
|
{
|
|
"id": "gemma4_large_family",
|
|
"priority": 60,
|
|
"match_any": [
|
|
"Gemma-4-26B",
|
|
"gemma-4-26B",
|
|
"Gemma-4-31B",
|
|
"gemma-4-31B"
|
|
],
|
|
"family_label": "Gemma 4 large family",
|
|
"sources": [
|
|
"Gemma 4 official model card"
|
|
],
|
|
"support_notes": {
|
|
"hf": "Gemma 4 supports configurable thinking, but current repo behavior can vary by backend.",
|
|
"gguf": "Use the settings manually and verify behavior in RAW_TRACE because GGUF support varies by backend.",
|
|
"soft_switch": "Gemma uses chat-template thinking controls instead of /think or /no_think."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 64,
|
|
"max_tokens": 1024,
|
|
"ctx_min": 262144
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 64,
|
|
"max_tokens": 1024,
|
|
"ctx_min": 262144
|
|
},
|
|
"notes": [
|
|
"For creative prompt generation, keep Gemma 4 near its official sampler defaults before tuning style or length.",
|
|
"Official context target is 256K on 26B A4B and 31B variants.",
|
|
"If the backend does not expose top_k, keep temperature and top_p at the official defaults."
|
|
]
|
|
},
|
|
{
|
|
"id": "gemma4_small_family",
|
|
"priority": 50,
|
|
"match_any": [
|
|
"Gemma-4-E2B",
|
|
"gemma-4-E2B",
|
|
"Gemma-4-E4B",
|
|
"gemma-4-E4B"
|
|
],
|
|
"family_label": "Gemma 4 small family",
|
|
"sources": [
|
|
"Gemma 4 official model card"
|
|
],
|
|
"support_notes": {
|
|
"hf": "Gemma 4 supports configurable thinking, but current repo behavior can vary by backend.",
|
|
"gguf": "Use the settings manually and verify behavior in RAW_TRACE because GGUF support varies by backend.",
|
|
"soft_switch": "Gemma uses chat-template thinking controls instead of /think or /no_think."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 64,
|
|
"max_tokens": 1024,
|
|
"ctx_min": 131072
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 64,
|
|
"max_tokens": 1024,
|
|
"ctx_min": 131072
|
|
},
|
|
"notes": [
|
|
"For creative prompt generation, keep Gemma 4 near its official sampler defaults before tuning style or length.",
|
|
"Official context target is 128K on E2B and E4B variants.",
|
|
"Gemma recommends higher visual token budgets only for OCR or fine detail tasks; lower budgets are faster for captioning or video summaries."
|
|
]
|
|
}
|
|
]
|
|
}
|