From d62ea54aa32d5d4595fb53e19cb788c9670feadd Mon Sep 17 00:00:00 2001 From: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com> Date: Sun, 25 Oct 2026 09:51:50 +0000 Subject: [PATCH] feat(gallery): add Saluki abliterated variants Add Kujira's mixed IQ2 Saluki fine-tune for llama.cpp with plain and MTP decoding. Pin the weights and document the text-only entries. Assisted-by: Codex:gpt-6 Signed-off-by: Ettore Di Giacinto (cherry picked from commit 24f23a66aa9c2c186d21e367db90f8a8d724f428, mudler/LocalAI#12688) Assisted-by: Claude:claude-sonnet-5-5 Signed-off-by: Ettore Di Giacinto --- gallery/index.yaml | 87 ++++++++++++++++++++++++++++++++ gallery/models/index.yaml | 103 ++++++++++++++++++++++++++++++++++++++ 2 files changed, 190 insertions(+) diff --git a/gallery/index.yaml b/gallery/index.yaml index e2e7091c2..0ace1d9b6 100644 --- a/gallery/index.yaml +++ b/gallery/index.yaml @@ -3,6 +3,93 @@ # written by scripts/models/cmd/modelcatalog and stays only so releases before # the cutover keep working. CI fails when it differs from the generated copy: # go run ./scripts/models/cmd/modelcatalog --library gallery/models/index.yaml --out gallery/index.yaml +- url: github:mudler/LocalAI/gallery/virtual.yaml@master + name: underdog-saluki-27b-1.0-abliterated + description: | + Kujira's abliterated adaptation of Underdog Saluki 27B 1.0 uses mixed + IQ2 weights with an embedded MTP head. This text-only llama.cpp build + uses ordinary decoding; the MTP variant enables speculative decoding. + license: apache-2.0 + urls: + - https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0 + - https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF + tags: + - llm + - gguf + - cpu + - gpu + - reasoning + - function-calling + - uncensored + files: + - filename: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + sha256: b9b231748c94e13ecc0f841942798eb36b27eaea60014fea896218bc86742bbc + uri: https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF/resolve/c1d826d73831efcb819d9d64c625f0fab56ae8fc/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + overrides: + backend: llama-cpp + context_size: 32768 + function: + automatic_tool_parsing_fallback: true + grammar: + disable: true + known_usecases: + - chat + options: + - use_jinja:true + parameters: + model: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + presence_penalty: 1.1 + temperature: 1.0 + top_k: 20 + top_p: 0.95 + template: + use_tokenizer_template: true + variants: + - model: underdog-saluki-27b-1.0-abliterated-mtp +- url: github:mudler/LocalAI/gallery/virtual.yaml@master + name: underdog-saluki-27b-1.0-abliterated-mtp + description: | + Kujira's abliterated adaptation of Underdog Saluki 27B 1.0 uses mixed + IQ2 weights with an embedded MTP head. This text-only llama.cpp build + enables multi-token prediction with up to two draft tokens. + license: apache-2.0 + urls: + - https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0 + - https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF + tags: + - llm + - gguf + - cpu + - gpu + - reasoning + - function-calling + - uncensored + - mtp + files: + - filename: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + sha256: b9b231748c94e13ecc0f841942798eb36b27eaea60014fea896218bc86742bbc + uri: https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF/resolve/c1d826d73831efcb819d9d64c625f0fab56ae8fc/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + overrides: + backend: llama-cpp + context_size: 32768 + function: + automatic_tool_parsing_fallback: true + grammar: + disable: true + known_usecases: + - chat + options: + - use_jinja:true + - spec_type:draft-mtp + - spec_n_max:2 + parameters: + model: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + presence_penalty: 1.1 + temperature: 1.0 + top_k: 20 + top_p: 0.95 + template: + use_tokenizer_template: true - url: github:mudler/LocalAI/gallery/virtual.yaml@master name: huihui-qwen3.8-flash-next-abliterated-v2-q4 description: | diff --git a/gallery/models/index.yaml b/gallery/models/index.yaml index 73351aeb3..7403d2958 100644 --- a/gallery/models/index.yaml +++ b/gallery/models/index.yaml @@ -4,6 +4,109 @@ # changes and never edit it by hand. CI fails when it is stale. A build from # master reads it as the default model source; a tagged build reads the gallery # until the release cutover. +- kind: model + id: underdog-saluki-27b-1.0-abliterated + name: underdog-saluki-27b-1.0-abliterated + version: 0.0.0-gallery.cdab61f388f4 + description: | + Kujira's abliterated adaptation of Underdog Saluki 27B 1.0 uses mixed + IQ2 weights with an embedded MTP head. This text-only llama.cpp build + uses ordinary decoding; the MTP variant enables speculative decoding. + license: apache-2.0 + urls: + - https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0 + - https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF + tags: + - llm + - gguf + - cpu + - gpu + - reasoning + - function-calling + - uncensored + channel: gallery + policy: + requireSignature: false + schemaVersion: 1 + model: + url: github:mudler/LocalAI/gallery/virtual.yaml@master + files: + - filename: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + sha256: b9b231748c94e13ecc0f841942798eb36b27eaea60014fea896218bc86742bbc + uri: https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF/resolve/c1d826d73831efcb819d9d64c625f0fab56ae8fc/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + overrides: + backend: llama-cpp + context_size: 32768 + function: + automatic_tool_parsing_fallback: true + grammar: + disable: true + known_usecases: + - chat + options: + - use_jinja:true + parameters: + model: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + presence_penalty: 1.1 + temperature: 1.0 + top_k: 20 + top_p: 0.95 + template: + use_tokenizer_template: true + variants: + - model: underdog-saluki-27b-1.0-abliterated-mtp +- kind: model + id: underdog-saluki-27b-1.0-abliterated-mtp + name: underdog-saluki-27b-1.0-abliterated-mtp + version: 0.0.0-gallery.774b2b47f0dc + description: | + Kujira's abliterated adaptation of Underdog Saluki 27B 1.0 uses mixed + IQ2 weights with an embedded MTP head. This text-only llama.cpp build + enables multi-token prediction with up to two draft tokens. + license: apache-2.0 + urls: + - https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0 + - https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF + tags: + - llm + - gguf + - cpu + - gpu + - reasoning + - function-calling + - uncensored + - mtp + channel: gallery + policy: + requireSignature: false + schemaVersion: 1 + model: + url: github:mudler/LocalAI/gallery/virtual.yaml@master + files: + - filename: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + sha256: b9b231748c94e13ecc0f841942798eb36b27eaea60014fea896218bc86742bbc + uri: https://huggingface.co/Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF/resolve/c1d826d73831efcb819d9d64c625f0fab56ae8fc/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + overrides: + backend: llama-cpp + context_size: 32768 + function: + automatic_tool_parsing_fallback: true + grammar: + disable: true + known_usecases: + - chat + options: + - use_jinja:true + - spec_type:draft-mtp + - spec_n_max:2 + parameters: + model: llama-cpp/models/underdog-saluki-27b-1.0-abliterated/Underdog-Saluki-27B-1.0-IQ2-mix-MTP-abliterated.gguf + presence_penalty: 1.1 + temperature: 1.0 + top_k: 20 + top_p: 0.95 + template: + use_tokenizer_template: true - kind: model id: huihui-qwen3.8-flash-next-abliterated-v2-q4 name: huihui-qwen3.8-flash-next-abliterated-v2-q4