LLM inference in C/C++
C++ ggml
convert: add @ModelBase.example (#27208)
Xuan-Son Nguyenfa88ae9368f7
diff --git a/conversion/afmoe.py b/conversion/afmoe.pyindex 5e66a51da6..844925dca7 100644--- a/conversion/afmoe.py+++ b/conversion/afmoe.py@@ -13,6 +13,7 @@ from .llama import LlamaModel@ModelBase.register("AfmoeForCausalLM")[email protected]("arcee-ai/Trinity-Large-Thinking")class AfmoeModel(LlamaModel):model_arch = gguf.MODEL_ARCH.AFMOEdiff --git a/conversion/arctic.py b/conversion/arctic.pyindex 775cacaab9..843e24a7b9 100644--- a/conversion/arctic.py+++ b/conversion/arctic.py@@ -16,6 +16,7 @@ from .llama import LlamaModel@ModelBase.register("ArcticForCausalLM")[email protected]("Snowflake/snowflake-arctic-instruct")class ArcticModel(TextModel):model_arch = gguf.MODEL_ARCH.ARCTICdiff --git a/conversion/baichuan.py b/conversion/baichuan.pyindex 4cf34057cd..769bdd5678 100644--- a/conversion/baichuan.py+++ b/conversion/baichuan.py@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("BaichuanForCausalLM", "BaiChuanForCausalLM")[email protected]("baichuan-inc/Baichuan2-7B-Chat", "baichuan-inc/Baichuan-7B")class BaichuanModel(TextModel):model_arch = gguf.MODEL_ARCH.BAICHUANdiff --git a/conversion/bailingmoe.py b/conversion/bailingmoe.pyindex 2c6425cb64..351be1df17 100644--- a/conversion/bailingmoe.py+++ b/conversion/bailingmoe.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("BailingMoeForCausalLM")[email protected]("inclusionAI/Ling-lite")class BailingMoeModel(TextModel):model_arch = gguf.MODEL_ARCH.BAILINGMOE@@ -108,6 +109,7 @@ class BailingMoeModel(TextModel):@ModelBase.register("BailingMoeV2ForCausalLM")[email protected]("inclusionAI/Ling-mini-2.0")class BailingMoeV2Model(TextModel):model_arch = gguf.MODEL_ARCH.BAILINGMOE2@@ -189,6 +191,7 @@ class BailingMoeV2Model(TextModel):@ModelBase.register("SarvamMoEForCausalLM", "modeling_sarvam_moe.SarvamMoEForCausalLM")[email protected]("sarvamai/sarvam-30b")class SarvamMoEModel(BailingMoeV2Model):model_arch = gguf.MODEL_ARCH.BAILINGMOE2# Sarvam-MoE shares the BailingMoeV2 architecture; only differences:diff --git a/conversion/bailingmoe3.py b/conversion/bailingmoe3.pyindex 3b0501bb2a..20bba23e51 100644--- a/conversion/bailingmoe3.py+++ b/conversion/bailingmoe3.py@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("BailingMoeV3ForCausalLM")[email protected]("inclusionAI/Ling-3.0-tiny", "inclusionAI/Ling-3.0-flash")class BailingMoeV3Model(TextModel):model_arch = gguf.MODEL_ARCH.BAILINGMOE3supports_mtp_export = Truediff --git a/conversion/base.py b/conversion/base.pyindex 4f4fbca347..56547ace00 100644--- a/conversion/base.py+++ b/conversion/base.py@@ -1149,6 +1149,14 @@ class ModelBase:return modelclsreturn func+ @classmethod+ def example(cls, *hf_repos: str) -> Callable[[AnyModel], AnyModel]:+ del hf_repos # unused++ def func(modelcls: AnyModel) -> AnyModel:+ return modelcls+ return func+@classmethoddef print_registered_models(cls):for model_type, model_classes in cls._model_classes.items():diff --git a/conversion/bert.py b/conversion/bert.pyindex 0d25d0d62d..8ea6c42dc6 100644--- a/conversion/bert.py+++ b/conversion/bert.py@@ -15,6 +15,7 @@ from .base import ModelBase, SentencePieceTokenTypes, TextModel, gguf, logger@ModelBase.register("BertModel", "BertForMaskedLM", "CamembertModel", "BertForSequenceClassification")[email protected]("BAAI/bge-small-en-v1.5", "dangvantuan/sentence-camembert-base")class BertModel(TextModel):model_arch = gguf.MODEL_ARCH.BERT@@ -240,6 +241,7 @@ class BertModel(TextModel):@ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification")[email protected]("distilbert/distilbert-base-uncased")class DistilBertModel(BertModel):model_arch = gguf.MODEL_ARCH.BERT@@ -263,6 +265,7 @@ class DistilBertModel(BertModel):@ModelBase.register("RobertaModel", "RobertaForSequenceClassification")[email protected]("sentence-transformers/stsb-roberta-base")class RobertaModel(BertModel):model_arch = gguf.MODEL_ARCH.BERT@@ -312,6 +315,7 @@ class RobertaModel(BertModel):@ModelBase.register("NomicBertModel")[email protected]("nomic-ai/nomic-embed-text-v1.5")class NomicBertModel(BertModel):model_arch = gguf.MODEL_ARCH.BERT@@ -400,6 +404,7 @@ class NomicBertModel(BertModel):@ModelBase.register("NeoBERT", "NeoBERTLMHead", "NeoBERTForSequenceClassification")[email protected]("chandar-lab/NeoBERT")class NeoBert(BertModel):model_arch = gguf.MODEL_ARCH.NEO_BERT@@ -431,6 +436,7 @@ class NeoBert(BertModel):@ModelBase.register("EuroBertModel", "JinaEmbeddingsV5Model")[email protected]("hf-tiny-v2/tiny-random-EuroBertModel", "jinaai/jina-embeddings-v5-text-nano")class EuroBertModel(TextModel):model_arch = gguf.MODEL_ARCH.EUROBERT@@ -459,6 +465,7 @@ class EuroBertModel(TextModel):@ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")[email protected]("BAAI/bge-m3")class XLMRobertaModel(BertModel):model_arch = gguf.MODEL_ARCH.BERT_lora_files = {}@@ -561,6 +568,7 @@ class XLMRobertaModel(BertModel):@ModelBase.register("JinaBertModel", "JinaBertForMaskedLM")[email protected]("jinaai/jina-embeddings-v2-base-en")class JinaBertV2Model(BertModel):model_arch = gguf.MODEL_ARCH.JINA_BERT_V2@@ -588,6 +596,7 @@ class JinaBertV2Model(BertModel):@ModelBase.register("ModernBertModel", "ModernBertForMaskedLM", "ModernBertForSequenceClassification")[email protected]("answerdotai/ModernBERT-base")class ModernBertModel(BertModel):model_arch = gguf.MODEL_ARCH.MODERN_BERTdiff --git a/conversion/bitnet.py b/conversion/bitnet.pyindex 0c2baee876..82bcadaf9a 100644--- a/conversion/bitnet.py+++ b/conversion/bitnet.py@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("BitnetForCausalLM", "BitNetForCausalLM")[email protected]("microsoft/bitnet-b1.58-2B-4T")class BitnetModel(TextModel):model_arch = gguf.MODEL_ARCH.BITNETdiff --git a/conversion/bloom.py b/conversion/bloom.pyindex d98edf6d50..9654cd4a0f 100644--- a/conversion/bloom.py+++ b/conversion/bloom.py@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("BloomForCausalLM", "BloomModel")[email protected]("bigscience/bloom-560m")class BloomModel(TextModel):model_arch = gguf.MODEL_ARCH.BLOOMdiff --git a/conversion/chameleon.py b/conversion/chameleon.pyindex a996bfa53c..8f2065df66 100644--- a/conversion/chameleon.py+++ b/conversion/chameleon.py@@ -12,6 +12,8 @@ from .llama import LlamaModel@ModelBase.register("ChameleonForConditionalGeneration")@ModelBase.register("ChameleonForCausalLM") # obsolete+# [TAG_HF_EXAMPLE_GATED] facebook/chameleon-7b is gated+# [TAG_HF_EXAMPLE_MISSING]class ChameleonModel(TextModel):model_arch = gguf.MODEL_ARCH.CHAMELEONdiff --git a/conversion/chatglm.py b/conversion/chatglm.pyindex d638550387..9b902dae30 100644--- a/conversion/chatglm.py+++ b/conversion/chatglm.py@@ -9,6 +9,7 @@ from .base import ModelBase, SentencePieceTokenTypes, TextModel, gguf@ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")[email protected]("THUDM/chatglm3-6b", "zai-org/glm-4-9b-chat-hf")class ChatGLMModel(TextModel):model_arch = gguf.MODEL_ARCH.CHATGLMdiff --git a/conversion/codeshell.py b/conversion/codeshell.pyindex 8bfc3178d4..1c7f1129b5 100644--- a/conversion/codeshell.py+++ b/conversion/codeshell.py@@ -4,6 +4,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("CodeShellForCausalLM")[email protected]("WisdomShell/CodeShell-7B")class CodeShellModel(TextModel):model_arch = gguf.MODEL_ARCH.CODESHELLdiff --git a/conversion/cogvlm.py b/conversion/cogvlm.pyindex d92df55d46..13c314441b 100644--- a/conversion/cogvlm.py+++ b/conversion/cogvlm.py@@ -11,6 +11,7 @@ from .llama import LlamaModel@ModelBase.register("CogVLMForCausalLM")[email protected]("THUDM/cogvlm2-llama3-chat-19B", "THUDM/cogvlm-chat-hf")class CogVLMVisionModel(MmprojModel):def set_gguf_parameters(self):@@ -29,5 +30,6 @@ class CogVLMVisionModel(MmprojModel):@ModelBase.register("CogVLMForCausalLM")[email protected]("THUDM/cogvlm2-llama3-chat-19B", "THUDM/cogvlm-chat-hf")class CogVLMModel(LlamaModel):model_arch = gguf.MODEL_ARCH.COGVLMdiff --git a/conversion/command_r.py b/conversion/command_r.pyindex 118565c669..971f93ebdf 100644--- a/conversion/command_r.py+++ b/conversion/command_r.py@@ -12,6 +12,8 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("CohereForCausalLM")+# [TAG_HF_EXAMPLE_GATED] CohereLabs/c4ai-command-r-v01 is gated+# [TAG_HF_EXAMPLE_MISSING]class CommandR2Model(TextModel):model_arch = gguf.MODEL_ARCH.COMMAND_R@@ -30,6 +32,8 @@ class CommandR2Model(TextModel):@ModelBase.register("Cohere2ForCausalLM")+# [TAG_HF_EXAMPLE_GATED] CohereLabs/c4ai-command-r7b-12-2024 is gated[email protected]("hf-tiny-v2/tiny-random-Cohere2ForCausalLM")class Cohere2Model(TextModel):model_arch = gguf.MODEL_ARCH.COHERE2@@ -59,6 +63,7 @@ class Cohere2Model(TextModel):@ModelBase.register("Cohere2MoeForCausalLM")[email protected]("CohereLabs/North-Mini-Code-1.0")class Cohere2MoeModel(TextModel):model_arch = gguf.MODEL_ARCH.COHERE2MOE_n_main_layers: int | None = Nonediff --git a/conversion/dbrx.py b/conversion/dbrx.pyindex 207ebcb893..d37ce83e78 100644--- a/conversion/dbrx.py+++ b/conversion/dbrx.py@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("DbrxForCausalLM")[email protected]("alpindale/dbrx-instruct")class DbrxModel(TextModel):model_arch = gguf.MODEL_ARCH.DBRXdiff --git a/conversion/deci.py b/conversion/deci.pyindex be446eefa6..2ccaa92a98 100644--- a/conversion/deci.py+++ b/conversion/deci.py@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("DeciLMForCausalLM")[email protected]("nvidia/Llama-3_1-Nemotron-51B-Instruct", "Deci/DeciLM-7B")class DeciModel(TextModel):model_arch = gguf.MODEL_ARCH.DECIdiff --git a/conversion/deepseek.py b/conversion/deepseek.pyindex dbfa96c2ee..225f8645d8 100644--- a/conversion/deepseek.py+++ b/conversion/deepseek.py@@ -18,6 +18,7 @@ from .qwen import QwenModel@ModelBase.register("DeepseekOCRForCausalLM")[email protected]("deepseek-ai/DeepSeek-OCR")class DeepseekOCRVisionModel(MmprojModel):# HF dynamic_preprocess() max_num, which differs per modelpreproc_max_tiles = 9@@ -100,11 +101,13 @@ class DeepseekOCRVisionModel(MmprojModel):@ModelBase.register("UnlimitedOCRForCausalLM")[email protected]("baidu/Unlimited-OCR")class UnlimitedOCRVisionModel(DeepseekOCRVisionModel):preproc_max_tiles = 32@ModelBase.register("DeepseekOCR2ForCausalLM")[email protected]("deepseek-ai/DeepSeek-OCR-2")class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):preproc_max_tiles = 6@@ -134,6 +137,7 @@ class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):@ModelBase.register("DeepseekForCausalLM")[email protected]("deepseek-ai/deepseek-moe-16b-chat")class DeepseekModel(TextModel):model_arch = gguf.MODEL_ARCH.DEEPSEEK@@ -228,6 +232,7 @@ class DeepseekModel(TextModel):"YoutuForCausalLM","YoutuVLForConditionalGeneration",)[email protected]("deepseek-ai/DeepSeek-V2-Lite", "deepseek-ai/DeepSeek-V3")class DeepseekV2Model(TextModel):model_arch = gguf.MODEL_ARCH.DEEPSEEK2@@ -457,6 +462,7 @@ class DeepseekV2Model(TextModel):@ModelBase.register("DeepseekV32ForCausalLM")[email protected]("deepseek-ai/DeepSeek-V3.2-Exp")class DeepseekV32Model(DeepseekV2Model):model_arch = gguf.MODEL_ARCH.DEEPSEEK32skip_mtp = False@@ -517,6 +523,7 @@ class DeepseekV32Model(DeepseekV2Model):@ModelBase.register("DeepseekV4ForCausalLM")[email protected]("deepseek-ai/DeepSeek-V4-Flash-Base")class DeepseekV4Model(TextModel):model_arch = gguf.MODEL_ARCH.DEEPSEEK4supports_mtp_export = True@@ -911,6 +918,7 @@ class DeepseekV4Model(TextModel):@ModelBase.register("DeepseekV4DSparkModel")[email protected]("deepseek-ai/DeepSeek-V4-Flash-DSpark")class DeepseekV4DSparkModel(DeepseekV4Model):model_arch = gguf.MODEL_ARCH.DFLASHdiff --git a/conversion/dots1.py b/conversion/dots1.pyindex 7ac299a6e6..ffa3b6db44 100644--- a/conversion/dots1.py+++ b/conversion/dots1.py@@ -11,6 +11,7 @@ from .qwen import Qwen2MoeModel@ModelBase.register("Dots1ForCausalLM")[email protected]("rednote-hilab/dots.llm1.inst")class Dots1Model(Qwen2MoeModel):model_arch = gguf.MODEL_ARCH.DOTS1diff --git a/conversion/dotsocr.py b/conversion/dotsocr.pyindex f87f62abde..ace6aa9a13 100644--- a/conversion/dotsocr.py+++ b/conversion/dotsocr.py@@ -9,6 +9,7 @@ from .base import MmprojModel, ModelBase, gguf@ModelBase.register("DotsOCRForCausalLM")[email protected]("rednote-hilab/dots.ocr")class DotsOCRVisionModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)diff --git a/conversion/dream.py b/conversion/dream.pyindex 459e8d46af..14f25404d6 100644--- a/conversion/dream.py+++ b/conversion/dream.py@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("DreamModel")[email protected]("Dream-org/Dream-v0-Instruct-7B")class DreamModel(TextModel):model_arch = gguf.MODEL_ARCH.DREAMdiff --git a/conversion/ernie.py b/conversion/ernie.pyindex aa8a3bc8ee..3c4226a259 100644--- a/conversion/ernie.py+++ b/conversion/ernie.py@@ -15,6 +15,7 @@ from .base import MmprojModel, ModelBase, TextModel, gguf@ModelBase.register("Ernie4_5_ForCausalLM", "Ernie4_5ForCausalLM")[email protected]("baidu/ERNIE-4.5-0.3B-PT")class Ernie4_5Model(TextModel):model_arch = gguf.MODEL_ARCH.ERNIE4_5@@ -73,6 +74,7 @@ class Ernie4_5Model(TextModel):@ModelBase.register("Ernie4_5_MoeForCausalLM")[email protected]("baidu/ERNIE-4.5-21B-A3B-PT")class Ernie4_5MoeModel(Ernie4_5Model):model_arch = gguf.MODEL_ARCH.ERNIE4_5_MOE_experts: list[dict[str, Tensor]] | None = None@@ -156,11 +158,13 @@ class Ernie4_5MoeModel(Ernie4_5Model):@ModelBase.register("PaddleOCRVLForConditionalGeneration")[email protected]("PaddlePaddle/PaddleOCR-VL")class PaddleOCRModel(Ernie4_5Model):model_arch = gguf.MODEL_ARCH.PADDLEOCR@ModelBase.register("PaddleOCRVisionModel")[email protected]("PaddlePaddle/PaddleOCR-VL")class PaddleOCRVisionModel(MmprojModel):# PaddleOCR-VL uses a modified version of Siglipmin_pixels: int = 0diff --git a/conversion/exaone.py b/conversion/exaone.pyindex 1cd2244dbc..0919d2ffaf 100644--- a/conversion/exaone.py+++ b/conversion/exaone.py@@ -15,6 +15,7 @@ from .qwenvl import Qwen2VLVisionModel@ModelBase.register("ExaoneForCausalLM")[email protected]("LGAI-EXAONE/EXAONE-3.5-2.4B-Instruct")class ExaoneModel(TextModel):model_arch = gguf.MODEL_ARCH.EXAONE@@ -60,6 +61,7 @@ class ExaoneModel(TextModel):@ModelBase.register("Exaone4ForCausalLM")[email protected]("LGAI-EXAONE/EXAONE-4.0-32B")class Exaone4Model(TextModel):model_arch = gguf.MODEL_ARCH.EXAONE4@@ -126,6 +128,7 @@ class Exaone4Model(TextModel):# note: transformers >= 5.1 renamed the class to "ExaoneMoeForCausalLM" (lowercase 'e'),# so accept both spellings - LG AI have updated the configs of already-released models@ModelBase.register("ExaoneMoEForCausalLM", "ExaoneMoeForCausalLM")[email protected]("LGAI-EXAONE/K-EXAONE-236B-A23B")class ExaoneMoEModel(Exaone4Model):model_arch = gguf.MODEL_ARCH.EXAONE_MOE@@ -214,6 +217,7 @@ class ExaoneMoEModel(Exaone4Model):@ModelBase.register("Exaone4_5_ForConditionalGeneration")[email protected]("LGAI-EXAONE/EXAONE-4.5-33B")class Exaone4_5_TextModel(Exaone4Model):"""Text tower of EXAONE 4.5; Tensors match EXAONE4"""@@ -267,6 +271,7 @@ class Exaone4_5_TextModel(Exaone4Model):@ModelBase.register("Exaone4_5_ForConditionalGeneration")[email protected]("LGAI-EXAONE/EXAONE-4.5-33B")class Exaone4_5VisionModel(Qwen2VLVisionModel):"""Vision tower for EXAONE 4.5; Qwen2-VL-style ViT (GQA) + patch merger"""diff --git a/conversion/falcon.py b/conversion/falcon.pyindex 085fd4cd33..2c55511a09 100644--- a/conversion/falcon.py+++ b/conversion/falcon.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("FalconForCausalLM", "RWForCausalLM")[email protected]("tiiuae/falcon-7b")class FalconModel(TextModel):model_arch = gguf.MODEL_ARCH.FALCONdiff --git a/conversion/falcon_h1.py b/conversion/falcon_h1.pyindex a8bc880b2c..6686f7001c 100644--- a/conversion/falcon_h1.py+++ b/conversion/falcon_h1.py@@ -12,6 +12,7 @@ from .mamba import Mamba2Model@ModelBase.register("FalconH1ForCausalLM")[email protected]("tiiuae/Falcon-H1-0.5B-Base")class FalconH1Model(Mamba2Model):model_arch = gguf.MODEL_ARCH.FALCON_H1diff --git a/conversion/gemma.py b/conversion/gemma.pyindex f15a10a38b..6b4d7d1715 100644--- a/conversion/gemma.py+++ b/conversion/gemma.py@@ -14,6 +14,8 @@ from .base import MmprojModel, ModelBase, TextModel, gguf, logger@ModelBase.register("GemmaForCausalLM")+# [TAG_HF_EXAMPLE_GATED] google/gemma-2b is gated[email protected]("trl-internal-testing/tiny-GemmaForCausalLM")class GemmaModel(TextModel):model_arch = gguf.MODEL_ARCH.GEMMA@@ -68,6 +70,8 @@ class GemmaModel(TextModel):@ModelBase.register("Gemma2ForCausalLM")+# [TAG_HF_EXAMPLE_GATED] google/gemma-2-9b-it is gated[email protected]("trl-internal-testing/tiny-Gemma2ForCausalLM")class Gemma2Model(TextModel):model_arch = gguf.MODEL_ARCH.GEMMA2@@ -118,6 +122,8 @@ class Gemma2Model(TextModel):@ModelBase.register("Gemma3ForCausalLM", "Gemma3ForConditionalGeneration")+# [TAG_HF_EXAMPLE_GATED] google/gemma-3-4b-it is gated[email protected]("trl-internal-testing/tiny-Gemma3ForConditionalGeneration", "hf-tiny-v2/tiny-random-Gemma3ForCausalLM")class Gemma3Model(TextModel):model_arch = gguf.MODEL_ARCH.GEMMA3@@ -174,6 +180,8 @@ class Gemma3Model(TextModel):@ModelBase.register("Gemma3TextModel")+# [TAG_HF_EXAMPLE_GATED] google/embeddinggemma-300m is gated[email protected]("hf-tiny-v2/tiny-random-Gemma3TextModel")class EmbeddingGemma(Gemma3Model):model_arch = gguf.MODEL_ARCH.GEMMA_EMBEDDINGmodule_paths = []@@ -248,6 +256,8 @@ class EmbeddingGemma(Gemma3Model):@ModelBase.register("Gemma3ForConditionalGeneration")+# [TAG_HF_EXAMPLE_GATED] google/gemma-3-4b-it is gated[email protected]("trl-internal-testing/tiny-Gemma3ForConditionalGeneration")class Gemma3VisionModel(MmprojModel):def set_gguf_parameters(self):super().set_gguf_parameters()@@ -352,6 +362,8 @@ class ConformerAudioModel(MmprojModel):@ModelBase.register("Gemma3nForConditionalGeneration")+# [TAG_HF_EXAMPLE_GATED] google/gemma-3n-E2B-it is gated[email protected]("hf-tiny-v2/tiny-random-Gemma3nForConditionalGeneration")class Gemma3nVisionAudioModel(ConformerAudioModel):has_audio_encoder = Truehas_vision_encoder = True@@ -471,6 +483,8 @@ class Gemma3nVisionAudioModel(ConformerAudioModel):@ModelBase.register("Gemma3nForCausalLM", "Gemma3nForConditionalGeneration")+# [TAG_HF_EXAMPLE_GATED] google/gemma-3n-E2B-it is gated[email protected]("hf-tiny-v2/tiny-random-Gemma3nForConditionalGeneration")class Gemma3NModel(Gemma3Model):model_arch = gguf.MODEL_ARCH.GEMMA3N@@ -615,6 +629,7 @@ class Gemma3NModel(Gemma3Model):@ModelBase.register("Gemma4ForConditionalGeneration", "Gemma4ForCausalLM")[email protected]("google/gemma-4-31B-it", "google/gemma-4-26B-A4B-it", "google/gemma-4-E2B-it")class Gemma4Model(Gemma3Model):model_arch = gguf.MODEL_ARCH.GEMMA4@@ -795,6 +810,7 @@ class Gemma4Model(Gemma3Model):@ModelBase.register("Gemma4UnifiedForConditionalGeneration")[email protected]("hf-tiny-v2/tiny-random-Gemma4UnifiedForConditionalGeneration")class Gemma4UnifiedModel(Gemma4Model):model_arch = gguf.MODEL_ARCH.GEMMA4@@ -815,6 +831,7 @@ class Gemma4UnifiedModel(Gemma4Model):@ModelBase.register("Gemma4AssistantForCausalLM", "Gemma4UnifiedAssistantForCausalLM")[email protected]("google/gemma-4-31B-it-assistant", "google/gemma-4-26B-A4B-it-assistant", "google/gemma-4-E2B-it-assistant")class Gemma4AssistantModel(Gemma4Model):model_arch = gguf.MODEL_ARCH.GEMMA4_ASSISTANT@@ -835,6 +852,7 @@ class Gemma4AssistantModel(Gemma4Model):@ModelBase.register("Gemma4ForConditionalGeneration")[email protected]("google/gemma-4-31B-it", "google/gemma-4-26B-A4B-it", "google/gemma-4-E2B-it")class Gemma4VisionAudioModel(MmprojModel):has_audio_encoder = Truehas_vision_encoder = True@@ -913,6 +931,7 @@ class Gemma4VisionAudioModel(MmprojModel):@ModelBase.register("Gemma4UnifiedForConditionalGeneration")[email protected]("hf-tiny-v2/tiny-random-Gemma4UnifiedForConditionalGeneration")class Gemma4UnifiedVisionAudioModel(Gemma4VisionAudioModel):has_audio_encoder = Truehas_vision_encoder = Truediff --git a/conversion/glm.py b/conversion/glm.pyindex e28f54574e..abd7f279f9 100644--- a/conversion/glm.py+++ b/conversion/glm.py@@ -15,6 +15,7 @@ from .deepseek import DeepseekV2Model@ModelBase.register("Glm4ForCausalLM", "Glm4vForConditionalGeneration")[email protected]("zai-org/GLM-4-9B-0414")class Glm4Model(TextModel):model_arch = gguf.MODEL_ARCH.GLM4use_mrope = False@@ -86,6 +87,7 @@ class Glm4Model(TextModel):@ModelBase.register("GlmOcrForConditionalGeneration")[email protected]("zai-org/GLM-OCR")class GlmOCRModel(Glm4Model):model_arch = gguf.MODEL_ARCH.GLM4use_mrope = False@@ -107,6 +109,7 @@ class GlmOCRModel(Glm4Model):@ModelBase.register("Glm4MoeForCausalLM", "Glm4vMoeForConditionalGeneration")[email protected]("zai-org/GLM-4.5-Air")class Glm4MoeModel(TextModel):model_arch = gguf.MODEL_ARCH.GLM4_MOE@@ -204,6 +207,7 @@ class Glm4MoeModel(TextModel):@ModelBase.register("Glm4MoeLiteForCausalLM")[email protected]("zai-org/GLM-4.7-Flash")class Glm4MoeLiteModel(DeepseekV2Model):model_arch = gguf.MODEL_ARCH.DEEPSEEK2skip_mtp = False@@ -272,6 +276,7 @@ class Glm4MoeLiteModel(DeepseekV2Model):@ModelBase.register("GlmMoeDsaForCausalLM")[email protected]("zai-org/GLM-5.2")class GlmMoeDsaModel(DeepseekV2Model):model_arch = gguf.MODEL_ARCH.GLM_DSAskip_mtp = False@@ -340,6 +345,7 @@ class GlmMoeDsaModel(DeepseekV2Model):@ModelBase.register("SolarOpenForCausalLM")[email protected]("upstage/Solar-Open-100B")class SolarOpenModel(Glm4MoeModel):model_arch = gguf.MODEL_ARCH.GLM4_MOEdiff --git a/conversion/gpt2.py b/conversion/gpt2.pyindex 1cf06ae8b5..06dff9e4c7 100644--- a/conversion/gpt2.py+++ b/conversion/gpt2.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("GPT2LMHeadModel")[email protected]("openai-community/gpt2")class GPT2Model(TextModel):model_arch = gguf.MODEL_ARCH.GPT2@@ -38,6 +39,7 @@ class GPT2Model(TextModel):@ModelBase.register("RuGPT3XLForCausalLM")[email protected]("evilfreelancer/ruGPT3XL")class RuGPT3XLModel(TextModel):model_arch = gguf.MODEL_ARCH.GPT2diff --git a/conversion/gpt_oss.py b/conversion/gpt_oss.pyindex d2c70c0bba..7542ec0ea8 100644--- a/conversion/gpt_oss.py+++ b/conversion/gpt_oss.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("GptOssForCausalLM")[email protected]("openai/gpt-oss-20b")class GptOssModel(TextModel):model_arch = gguf.MODEL_ARCH.GPT_OSSdiff --git a/conversion/gptneox.py b/conversion/gptneox.pyindex 6a42b12b15..0b0e91c4f5 100644--- a/conversion/gptneox.py+++ b/conversion/gptneox.py@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("GPTNeoXForCausalLM")[email protected]("EleutherAI/pythia-70m")class GPTNeoXModel(TextModel):model_arch = gguf.MODEL_ARCH.GPTNEOXdiff --git a/conversion/granite.py b/conversion/granite.pyindex 956342e6d6..5f1e3e8472 100644--- a/conversion/granite.py+++ b/conversion/granite.py@@ -15,6 +15,7 @@ from .mamba import Mamba2Model@ModelBase.register("GraniteForCausalLM")[email protected]("ibm-granite/granite-3.3-2b-instruct")class GraniteModel(LlamaModel):"""Conversion for IBM's GraniteForCausalLM"""model_arch = gguf.MODEL_ARCH.GRANITE@@ -74,6 +75,7 @@ class GraniteModel(LlamaModel):@ModelBase.register("GraniteMoeForCausalLM", "GraniteMoeSharedForCausalLM")[email protected]("ibm-granite/granite-3.1-3b-a800m-instruct")class GraniteMoeModel(GraniteModel):"""Conversion for IBM's GraniteMoeForCausalLM"""model_arch = gguf.MODEL_ARCH.GRANITE_MOE@@ -124,6 +126,7 @@ class GraniteMoeModel(GraniteModel):@ModelBase.register("GraniteSwitchForCausalLM")[email protected]("ibm-granite/granite-switch-4.1-3b-preview")class GraniteSwitchModel(GraniteMoeModel):"""Dense, all-attention Granite with N per-token embedded LoRA adapters, stackedover the adapter dim with a zero adapter at slot 0 (N = num_adapters + 1)."""@@ -284,6 +287,7 @@ class GraniteSwitchModel(GraniteMoeModel):@ModelBase.register("GraniteMoeHybridForCausalLM", "BambaForCausalLM")[email protected]("ibm-granite/granite-4.0-h-tiny", "ibm-ai-platform/Bamba-9B-v2")class GraniteHybridModel(Mamba2Model, GraniteMoeModel):"""GraniteHybrid is a hybrid SSM + Attention model that uses Mamba2 SSMlayers and optionally uses MoE w/ a shared expert"""@@ -426,6 +430,7 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):@ModelBase.register("GraniteSpeechForConditionalGeneration")[email protected]("ibm-granite/granite-speech-3.3-2b", "ibm-granite/granite-4.0-1b-speech")class GraniteSpeechMmprojModel(MmprojModel):has_vision_encoder = Falsehas_audio_encoder = True@@ -509,6 +514,7 @@ class GraniteSpeechMmprojModel(MmprojModel):@ModelBase.register("GraniteSpeechPlusForConditionalGeneration")[email protected]("ibm-granite/granite-speech-4.1-2b-plus")class GraniteSpeechPlusMmprojModel(GraniteSpeechMmprojModel):"""Conversion for GraniteSpeechPlus - extends GraniteSpeech with feature layer concatenation"""has_vision_encoder = False@@ -537,6 +543,7 @@ class GraniteSpeechPlusMmprojModel(GraniteSpeechMmprojModel):@ModelBase.register("Granite4VisionForConditionalGeneration")[email protected]("ibm-granite/granite-4.0-3b-vision")class Granite4VisionMmprojModel(MmprojModel):has_vision_encoder = Truehas_audio_encoder = Falsediff --git a/conversion/grok.py b/conversion/grok.pyindex 9098e514a3..b966361d29 100644--- a/conversion/grok.py+++ b/conversion/grok.py@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("GrokForCausalLM", "Grok1ForCausalLM")[email protected]("keyfan/grok-1-hf")class GrokModel(TextModel):model_arch = gguf.MODEL_ARCH.GROKdiff --git a/conversion/grovemoe.py b/conversion/grovemoe.pyindex a8be931cb9..f418f18ac4 100644--- a/conversion/grovemoe.py+++ b/conversion/grovemoe.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("GroveMoeForCausalLM", "modeling_grove_moe.GroveMoeForCausalLM")[email protected]("inclusionAI/GroveMoE-Inst")class GroveMoeModel(TextModel):model_arch = gguf.MODEL_ARCH.GROVEMOEdiff --git a/conversion/hunyuan.py b/conversion/hunyuan.pyindex f5ac8a4fb7..ee1a106545 100644--- a/conversion/hunyuan.py+++ b/conversion/hunyuan.py@@ -17,6 +17,7 @@ from .qwen import QwenModel@ModelBase.register("HunYuanMoEV1ForCausalLM")[email protected]("tencent/Hunyuan-A13B-Instruct")class HunYuanMoEModel(TextModel):model_arch = gguf.MODEL_ARCH.HUNYUAN_MOE@@ -154,6 +155,7 @@ class HunYuanMoEModel(TextModel):@ModelBase.register("HunYuanDenseV1ForCausalLM")[email protected]("tencent/Hunyuan-4B-Instruct")class HunYuanModel(TextModel):model_arch = gguf.MODEL_ARCH.HUNYUAN_DENSE@@ -290,6 +292,7 @@ class HunYuanModel(TextModel):@ModelBase.register("HunYuanVLForConditionalGeneration")[email protected]("tencent/HunyuanOCR")class HunyuanVLVisionModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)@@ -333,6 +336,7 @@ class HunyuanVLVisionModel(MmprojModel):@ModelBase.register("HunYuanVLForConditionalGeneration")[email protected]("tencent/HunyuanOCR")class HunyuanVLTextModel(HunYuanModel):model_arch = gguf.MODEL_ARCH.HUNYUAN_VL@@ -365,6 +369,7 @@ class HunyuanVLTextModel(HunYuanModel):@ModelBase.register("HYV3ForCausalLM")[email protected]("tencent/Hy3")class HYV3Model(TextModel):model_arch = gguf.MODEL_ARCH.HY_V3supports_mtp_export = Truediff --git a/conversion/internlm.py b/conversion/internlm.pyindex 7e11aca3ce..df2668474f 100644--- a/conversion/internlm.py+++ b/conversion/internlm.py@@ -14,6 +14,7 @@ from .llama import LlamaModel@ModelBase.register("InternLM2ForCausalLM")[email protected]("internlm/internlm2-chat-7b")class InternLM2Model(TextModel):model_arch = gguf.MODEL_ARCH.INTERNLM2@@ -170,6 +171,7 @@ class InternLM2Model(TextModel):@ModelBase.register("InternLM3ForCausalLM")[email protected]("internlm/internlm3-8b-instruct")class InternLM3Model(TextModel):model_arch = gguf.MODEL_ARCH.LLAMAdiff --git a/conversion/internvl.py b/conversion/internvl.pyindex 9a2a1e43df..799e23f5f5 100644--- a/conversion/internvl.py+++ b/conversion/internvl.py@@ -9,6 +9,7 @@ from .base import MmprojModel, ModelBase, gguf@ModelBase.register("InternVisionModel")[email protected]("OpenGVLab/InternVL3-2B", "OpenGVLab/InternVL2_5-1B")class InternVisionModel(MmprojModel):min_dynamic_tiles: int = 0diff --git a/conversion/jais.py b/conversion/jais.pyindex 00add4c77f..f3f96c3efd 100644--- a/conversion/jais.py+++ b/conversion/jais.py@@ -11,6 +11,8 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("Jais2ForCausalLM")+# [TAG_HF_EXAMPLE_GATED] inceptionai/Jais-2-8B-Chat is gated+# [TAG_HF_EXAMPLE_MISSING]class Jais2Model(TextModel):model_arch = gguf.MODEL_ARCH.JAIS2@@ -22,6 +24,7 @@ class Jais2Model(TextModel):@ModelBase.register("JAISLMHeadModel")[email protected]("inceptionai/jais-family-590m")class JaisModel(TextModel):model_arch = gguf.MODEL_ARCH.JAISdiff --git a/conversion/jamba.py b/conversion/jamba.pyindex da712ba501..a2e642cb01 100644--- a/conversion/jamba.py+++ b/conversion/jamba.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("JambaForCausalLM")[email protected]("ai21labs/Jamba-v0.1")class JambaModel(TextModel):model_arch = gguf.MODEL_ARCH.JAMBAdiff --git a/conversion/januspro.py b/conversion/januspro.pyindex b49691205c..0f71ab3cd6 100644--- a/conversion/januspro.py+++ b/conversion/januspro.py@@ -11,6 +11,7 @@ from .llama import LlamaModel@ModelBase.register("JanusForConditionalGeneration")[email protected]("deepseek-community/Janus-Pro-1B")class JanusProModel(LlamaModel):model_arch = gguf.MODEL_ARCH.LLAMA # reuse Llama arch@@ -34,6 +35,7 @@ class JanusProModel(LlamaModel):@ModelBase.register("JanusForConditionalGeneration")[email protected]("deepseek-community/Janus-Pro-1B")class JanusProVisionModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)diff --git a/conversion/kimi_k3.py b/conversion/kimi_k3.pyindex a5005e15bd..d15d1d64bf 100644--- a/conversion/kimi_k3.py+++ b/conversion/kimi_k3.py@@ -16,6 +16,7 @@ from .kimi_linear import KimiLinearModel@ModelBase.register("KimiK3ForConditionalGeneration")[email protected]("moonshotai/Kimi-K3")class KimiK3Model(TextModel):"""Kimi-K3 text model (KimiLinearForCausalLM under a `language_model.` prefix).diff --git a/conversion/kimi_linear.py b/conversion/kimi_linear.pyindex f2e6cda83c..697ab1b4a9 100644--- a/conversion/kimi_linear.py+++ b/conversion/kimi_linear.py@@ -13,6 +13,7 @@ from .qwen import QwenModel@ModelBase.register("KimiLinearModel", "KimiLinearForCausalLM")[email protected]("moonshotai/Kimi-Linear-48B-A3B-Instruct")class KimiLinearModel(TextModel):"""Kimi-Linear model with hybrid MLA+KDA architecture"""model_arch = gguf.MODEL_ARCH.KIMI_LINEARdiff --git a/conversion/kimivl.py b/conversion/kimivl.pyindex 5ff3c39ca9..ae60abf309 100644--- a/conversion/kimivl.py+++ b/conversion/kimivl.py@@ -11,6 +11,7 @@ from .base import MmprojModel, ModelBase, gguf@ModelBase.register("KimiVLForConditionalGeneration")[email protected]("moonshotai/Kimi-VL-A3B-Instruct")class KimiVLModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)@@ -52,6 +53,7 @@ class KimiVLModel(MmprojModel):@ModelBase.register("KimiK25ForConditionalGeneration")[email protected]("moonshotai/Kimi-K2.5")class KimiK25Model(MmprojModel):"""Kimi-K2.5 with MoonViT3d vision encoder"""@@ -155,6 +157,7 @@ class KimiK25Model(MmprojModel):@ModelBase.register("Glm5vForConditionalGeneration")+# [TAG_HF_EXAMPLE_MISSING]class Glm5vModel(KimiK25Model):"""GLM-5.2-Vision MoonViT3d encoder and projectordiff --git a/conversion/laguna.py b/conversion/laguna.pyindex a90f355ca9..29e0b3d6b3 100644--- a/conversion/laguna.py+++ b/conversion/laguna.py@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("LagunaForCausalLM")[email protected]("poolside/Laguna-XS.2", "poolside/Laguna-S-2.1")class LagunaModel(TextModel):model_arch = gguf.MODEL_ARCH.LAGUNA_experts: list[dict] | None = Nonediff --git a/conversion/lfm2.py b/conversion/lfm2.pyindex 70ce45658b..984f444806 100644--- a/conversion/lfm2.py+++ b/conversion/lfm2.py@@ -13,6 +13,7 @@ from .gemma import ConformerAudioModel@ModelBase.register("Lfm2ForCausalLM", "LFM2ForCausalLM")[email protected]("LiquidAI/LFM2-1.2B", "LiquidAI/LFM2.5-350M")class LFM2Model(TextModel):model_arch = gguf.MODEL_ARCH.LFM2@@ -65,6 +66,7 @@ class LFM2Model(TextModel):@ModelBase.register("Lfm2Model", "Lfm2BidirectionalModel")[email protected]("LiquidAI/LFM2.5-ColBERT-350M", "LiquidAI/LFM2.5-Embedding-350M")class LFM2ColBertModel(LFM2Model):model_arch = gguf.MODEL_ARCH.LFM2dense_tensor_name = "dense_2"@@ -93,6 +95,7 @@ class LFM2ColBertModel(LFM2Model):@ModelBase.register("Lfm2MoeForCausalLM")[email protected]("LiquidAI/LFM2-8B-A1B")class LFM2MoeModel(TextModel):model_arch = gguf.MODEL_ARCH.LFM2MOE@@ -166,6 +169,7 @@ class LFM2MoeModel(TextModel):@ModelBase.register("Lfm2VlForConditionalGeneration")[email protected]("LiquidAI/LFM2-VL-450M")class LFM2VLModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)@@ -200,6 +204,7 @@ class LFM2VLModel(MmprojModel):@ModelBase.register("Lfm2AudioForConditionalGeneration")[email protected]("LiquidAI/LFM2.5-Audio-1.5B", "LiquidAI/LFM2-Audio-1.5B")class LFM2AudioModel(ConformerAudioModel):has_vision_encoder = Falsehas_audio_encoder = True@@ -238,6 +243,7 @@ class LFM2AudioModel(ConformerAudioModel):@ModelBase.register("Lfm25AudioTokenizer")[email protected]("LiquidAI/LFM2.5-Audio-1.5B")class LFM25AudioTokenizer(LFM2Model):model_arch = gguf.MODEL_ARCH.LFM2diff --git a/conversion/lighton_ocr.py b/conversion/lighton_ocr.pyindex ead3200ac1..8686fe5c91 100644--- a/conversion/lighton_ocr.py+++ b/conversion/lighton_ocr.py@@ -11,6 +11,7 @@ from .llava import LlavaVisionModel@ModelBase.register("LightOnOCRForConditionalGeneration")[email protected]("lightonai/LightOnOCR-1B-1025")class LightOnOCRVisionModel(LlavaVisionModel):is_mistral_format = Falseuse_break_tok = Falsediff --git a/conversion/llada.py b/conversion/llada.pyindex 98dc9de95b..c03607191a 100644--- a/conversion/llada.py+++ b/conversion/llada.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("LLaDAModelLM")[email protected]("GSAI-ML/LLaDA-8B-Instruct")class LLaDAModel(TextModel):model_arch = gguf.MODEL_ARCH.LLADAundo_permute = True@@ -114,6 +115,7 @@ class LLaDAModel(TextModel):@ModelBase.register("LLaDAMoEModel", "LLaDAMoEModelLM")[email protected]("inclusionAI/LLaDA-MoE-7B-A1B-Instruct")class LLaDAMoEModel(TextModel):model_arch = gguf.MODEL_ARCH.LLADA_MOEdiff --git a/conversion/llama.py b/conversion/llama.pyindex 1aced49c54..41d8c23092 100644--- a/conversion/llama.py+++ b/conversion/llama.py@@ -28,6 +28,8 @@ from .base import ModelBase, TextModel, gguf, logger"Eagle3DraftModel","IQuestCoderForCausalLM","LlamaModel")+# [TAG_HF_EXAMPLE_GATED] meta-llama/Llama-3.2-1B-Instruct is gated[email protected]("unsloth/Llama-3.2-1B-Instruct", "mistralai/Mistral-7B-Instruct-v0.3", "mistralai/Mixtral-8x7B-Instruct-v0.1")class LlamaModel(TextModel):model_arch = gguf.MODEL_ARCH.LLAMAundo_permute = True@@ -359,6 +361,7 @@ class LlamaModel(TextModel):@ModelBase.register("ArceeForCausalLM")[email protected]("arcee-ai/AFM-4.5B")class ArceeModel(LlamaModel):model_arch = gguf.MODEL_ARCH.ARCEE@@ -371,6 +374,8 @@ class ArceeModel(LlamaModel):"Llama4ForConditionalGeneration","Llama4ForCausalLM",)+# [TAG_HF_EXAMPLE_GATED] meta-llama/Llama-4-Scout-17B-16E-Instruct is gated[email protected]("unsloth/Llama-4-Scout-17B-16E-Instruct")class Llama4Model(LlamaModel):model_arch = gguf.MODEL_ARCH.LLAMA4undo_permute = False@@ -412,16 +417,19 @@ class Llama4Model(LlamaModel):@ModelBase.register("LlamaBidirectionalModel")[email protected]("nvidia/llama-embed-nemotron-8b")class LlamaEmbedNemotronModel(LlamaModel):model_arch = gguf.MODEL_ARCH.LLAMA_EMBED@ModelBase.register("SmolLM3ForCausalLM")[email protected]("HuggingFaceTB/SmolLM3-3B")class SmolLM3Model(LlamaModel):model_arch = gguf.MODEL_ARCH.SMOLLM3@ModelBase.register("ApertusForCausalLM")[email protected]("swiss-ai/Apertus-8B-Instruct-2509")class ApertusModel(LlamaModel):model_arch = gguf.MODEL_ARCH.APERTUSundo_permute = Falsediff --git a/conversion/llama4.py b/conversion/llama4.pyindex f84c762961..280e309dd5 100644--- a/conversion/llama4.py+++ b/conversion/llama4.py@@ -9,6 +9,8 @@ from .base import MmprojModel, ModelBase, gguf@ModelBase.register("Llama4ForConditionalGeneration")+# [TAG_HF_EXAMPLE_GATED] meta-llama/Llama-4-Scout-17B-16E-Instruct is gated[email protected]("unsloth/Llama-4-Scout-17B-16E-Instruct")class Llama4VisionModel(MmprojModel):def set_gguf_parameters(self):super().set_gguf_parameters()diff --git a/conversion/llava.py b/conversion/llava.pyindex 31d6e2ad80..98a004f986 100644--- a/conversion/llava.py+++ b/conversion/llava.py@@ -16,6 +16,7 @@ from .llama import LlamaModel"LlavaForConditionalGeneration", # pixtral"Mistral3ForConditionalGeneration", # mistral small 3.1)[email protected]("mistral-community/pixtral-12b", "mistralai/Mistral-Small-3.1-24B-Instruct-2503")class LlavaVisionModel(MmprojModel):img_break_tok_id = -1use_break_tok = Truediff --git a/conversion/maincoder.py b/conversion/maincoder.pyindex 18b625b08f..2e291b8a96 100644--- a/conversion/maincoder.py+++ b/conversion/maincoder.py@@ -4,6 +4,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("MaincoderForCausalLM")[email protected]("Maincode/Maincoder-1B")class MaincoderModel(TextModel):model_arch = gguf.MODEL_ARCH.MAINCODERdiff --git a/conversion/mamba.py b/conversion/mamba.pyindex 43d559ffb0..8a2a463752 100644--- a/conversion/mamba.py+++ b/conversion/mamba.py@@ -14,6 +14,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")[email protected]("state-spaces/mamba-130m-hf", "tiiuae/falcon-mamba-7b")class MambaModel(TextModel):model_arch = gguf.MODEL_ARCH.MAMBA@@ -100,6 +101,7 @@ class MambaModel(TextModel):@ModelBase.register("Mamba2ForCausalLM")[email protected]("mistralai/Mamba-Codestral-7B-v0.1")class Mamba2Model(TextModel):model_arch = gguf.MODEL_ARCH.MAMBA2diff --git a/conversion/mellum.py b/conversion/mellum.pyindex 79bc6755cc..1e50f92aea 100644--- a/conversion/mellum.py+++ b/conversion/mellum.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("MellumForCausalLM")[email protected]("JetBrains/Mellum2-12B-A2.5B-Base")class MellumModel(TextModel):model_arch = gguf.MODEL_ARCH.MELLUMdiff --git a/conversion/mimo.py b/conversion/mimo.pyindex ca2ed28ad3..15dbeb7e75 100644--- a/conversion/mimo.py+++ b/conversion/mimo.py@@ -14,6 +14,7 @@ from .base import MmprojModel, ModelBase, TextModel, gguf@ModelBase.register("MiMoV2FlashForCausalLM", "MiMoV2ForCausalLM")[email protected]("XiaomiMiMo/MiMo-V2.5")class MimoV2Model(TextModel):model_arch = gguf.MODEL_ARCH.MIMO2@@ -230,6 +231,7 @@ class MimoV2Model(TextModel):@ModelBase.register("MiMoV2ForCausalLM")[email protected]("XiaomiMiMo/MiMo-V2.5")class MiMoV2VisionAudioModel(MmprojModel):has_audio_encoder = Truediff --git a/conversion/minicpm.py b/conversion/minicpm.pyindex bf3fa81421..678d7bec18 100644--- a/conversion/minicpm.py+++ b/conversion/minicpm.py@@ -14,6 +14,7 @@ from .qwen import Qwen3_5TextModel@ModelBase.register("MiniCPMForCausalLM")[email protected]("openbmb/MiniCPM-2B-sft-bf16")class MiniCPMModel(TextModel):model_arch = gguf.MODEL_ARCH.MINICPM@@ -61,6 +62,7 @@ class MiniCPMModel(TextModel):@ModelBase.register("MiniCPM3ForCausalLM")[email protected]("openbmb/MiniCPM3-4B")class MiniCPM3Model(TextModel):model_arch = gguf.MODEL_ARCH.MINICPM3@@ -117,6 +119,7 @@ class MiniCPM3Model(TextModel):# the LM (text mode) and once as the mmproj (vision mode), mirroring the Qwen3-VL setup.@ModelBase.register("MiniCPMV4_6ForConditionalGeneration")[email protected]("openbmb/MiniCPM-V-4_6")class MiniCPMV4_6TextModel(Qwen3_5TextModel):model_arch = gguf.MODEL_ARCH.QWEN35@@ -134,6 +137,7 @@ class MiniCPMV4_6TextModel(Qwen3_5TextModel):@ModelBase.register("MiniCPMV4_6ForConditionalGeneration")[email protected]("openbmb/MiniCPM-V-4_6")class MiniCPMV4_6VisionModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)diff --git a/conversion/minimax.py b/conversion/minimax.pyindex d7a00bac95..53a9ff60f8 100644--- a/conversion/minimax.py+++ b/conversion/minimax.py@@ -12,6 +12,7 @@ from .base import ModelBase, TextModel, MmprojModel, gguf, logger@ModelBase.register("MiniMaxText01ForCausalLM")@ModelBase.register("MiniMaxM1ForCausalLM")[email protected]("MiniMaxAI/MiniMax-Text-01", "MiniMaxAI/MiniMax-M1-40k")class MiniMaxText01Model(TextModel):model_arch = gguf.MODEL_ARCH.MINIMAX01@@ -119,6 +120,7 @@ class MiniMaxText01Model(TextModel):@ModelBase.register("MiniMaxM2ForCausalLM")[email protected]("MiniMaxAI/MiniMax-M2")class MiniMaxM2Model(TextModel):model_arch = gguf.MODEL_ARCH.MINIMAXM2_experts_cache: dict[int, dict[str, Tensor]] = {}@@ -163,6 +165,7 @@ class MiniMaxM2Model(TextModel):@ModelBase.register("MiniMaxM3SparseForCausalLM", "MiniMaxM3SparseForConditionalGeneration")[email protected]("MiniMaxAI/MiniMax-M3")class MiniMaxM3Model(MiniMaxM2Model):model_arch = gguf.MODEL_ARCH.MINIMAXM3@@ -203,6 +206,7 @@ class MiniMaxM3Model(MiniMaxM2Model):@ModelBase.register("MiniMaxM3SparseForConditionalGeneration", "MiniMaxM3VLForConditionalGeneration")[email protected]("MiniMaxAI/MiniMax-M3")class MiniMaxM3VisionModel(MmprojModel):@classmethoddef filter_tensors(cls, item):diff --git a/conversion/mistral3.py b/conversion/mistral3.pyindex af9438ae70..fee039b353 100644--- a/conversion/mistral3.py+++ b/conversion/mistral3.py@@ -15,6 +15,7 @@ from .llama import LlamaModel"Mistral3ForConditionalGeneration","Ministral3ForCausalLM",)[email protected]("mistralai/Mistral-Small-3.1-24B-Instruct-2503", "hf-tiny-v2/tiny-random-Ministral3ForCausalLM")class Mistral3Model(TextModel):class Ministral3Model(LlamaModel):model_arch = gguf.MODEL_ARCH.MISTRAL3diff --git a/conversion/mpt.py b/conversion/mpt.pyindex 9557ab7fa6..d5d849ff35 100644--- a/conversion/mpt.py+++ b/conversion/mpt.py@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("MPTForCausalLM")[email protected]("anas-awadalla/mpt-7b")class MPTModel(TextModel):model_arch = gguf.MODEL_ARCH.MPTdiff --git a/conversion/muse_glimmer.py b/conversion/muse_glimmer.pyindex cc588e8321..b205f70a0e 100644--- a/conversion/muse_glimmer.py+++ b/conversion/muse_glimmer.py@@ -24,6 +24,7 @@ def _unpermute_for_rope(tensor: "Tensor", n_heads: int) -> "Tensor":@ModelBase.register("MuseGlimmerForConditionalGeneration")[email protected]("meta-models/Muse-Glimmer-30B")class MuseGlimmerModel(TextModel):model_arch = gguf.MODEL_ARCH.MUSE_GLIMMER@@ -78,6 +79,7 @@ class MuseGlimmerModel(TextModel):@ModelBase.register("MuseGlimmerForConditionalGeneration")[email protected]("meta-models/Muse-Glimmer-30B")class MuseGlimmerVisionModel(MmprojModel):def get_vision_config(self) -> dict[str, Any] | None:c = self.global_config.get("vision_config")@@ -131,6 +133,7 @@ class MuseGlimmerVisionModel(MmprojModel):@ModelBase.register("MuseGlimmerAssistantModel")[email protected]("meta-models/Muse-Glimmer-30B-assistant")class MuseGlimmerAssistantModel(TextModel):model_arch = gguf.MODEL_ARCH.DFLASHdiff --git a/conversion/nanbeige.py b/conversion/nanbeige.pyindex f1fc425b3a..a5b269a7a2 100644--- a/conversion/nanbeige.py+++ b/conversion/nanbeige.py@@ -5,6 +5,7 @@ from .llama import LlamaModel@ModelBase.register("NanbeigeForCausalLM")[email protected]("Nanbeige/Nanbeige4.2-3B")class NanbeigeModel(LlamaModel):model_arch = gguf.MODEL_ARCH.NANBEIGEundo_permute = Truediff --git a/conversion/nemotron.py b/conversion/nemotron.pyindex c46cec1438..3e37c7b469 100644--- a/conversion/nemotron.py+++ b/conversion/nemotron.py@@ -16,6 +16,7 @@ from .granite import GraniteHybridModel"NemotronH_Nano_VL_V2","RADIOModel",)[email protected]("nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16")class NemotronNanoV2VLModel(MmprojModel):# ViT-Huge architecture parameters for RADIO v2.5-h_vit_hidden_size = 1280@@ -151,6 +152,7 @@ class NemotronNanoV2VLModel(MmprojModel):@ModelBase.register("NemotronForCausalLM")[email protected]("nvidia/Minitron-4B-Base")class NemotronModel(TextModel):model_arch = gguf.MODEL_ARCH.NEMOTRON@@ -193,6 +195,7 @@ class NemotronModel(TextModel):@ModelBase.register("NemotronHForCausalLM")[email protected]("nvidia/Nemotron-H-8B-Base-8K")class NemotronHModel(GraniteHybridModel):"""Hybrid mamba2/attention model from NVIDIA"""model_arch = gguf.MODEL_ARCH.NEMOTRON_Hdiff --git a/conversion/olmo.py b/conversion/olmo.pyindex 1664c30e40..e6faa19758 100644--- a/conversion/olmo.py+++ b/conversion/olmo.py@@ -14,6 +14,7 @@ from .llama import LlamaModel@ModelBase.register("OlmoForCausalLM")@ModelBase.register("OLMoForCausalLM")[email protected]("allenai/OLMo-1.7-7B-hf")class OlmoModel(TextModel):model_arch = gguf.MODEL_ARCH.OLMO@@ -39,12 +40,14 @@ class OlmoModel(TextModel):@ModelBase.register("SeedOssForCausalLM")[email protected]("ByteDance-Seed/Seed-OSS-36B-Instruct")class SeedOssModel(TextModel):model_arch = gguf.MODEL_ARCH.SEED_OSS@ModelBase.register("Olmo2ForCausalLM")@ModelBase.register("Olmo3ForCausalLM")[email protected]("allenai/OLMo-2-1124-7B-Instruct", "allenai/Olmo-3-7B-Instruct")class Olmo2Model(TextModel):model_arch = gguf.MODEL_ARCH.OLMO2@@ -67,6 +70,7 @@ class Olmo2Model(TextModel):@ModelBase.register("OlmoeForCausalLM")[email protected]("allenai/OLMoE-1B-7B-0924")class OlmoeModel(TextModel):model_arch = gguf.MODEL_ARCH.OLMOEdiff --git a/conversion/openelm.py b/conversion/openelm.pyindex ecc746dc34..8863378bbf 100644--- a/conversion/openelm.py+++ b/conversion/openelm.py@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("OpenELMForCausalLM")[email protected]("apple/OpenELM-270M")class OpenELMModel(TextModel):model_arch = gguf.MODEL_ARCH.OPENELMdiff --git a/conversion/orion.py b/conversion/orion.pyindex 8dfceeed1f..3e4c633c18 100644--- a/conversion/orion.py+++ b/conversion/orion.py@@ -4,6 +4,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("OrionForCausalLM")[email protected]("OrionStarAI/Orion-14B-Base")class OrionModel(TextModel):model_arch = gguf.MODEL_ARCH.ORIONdiff --git a/conversion/pangu.py b/conversion/pangu.pyindex 42016ba028..74c76532b5 100644--- a/conversion/pangu.py+++ b/conversion/pangu.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("PanguEmbeddedForCausalLM")[email protected]("FreedomIntelligence/openPangu-Embedded-7B-V1.1")class PanguEmbeddedModel(TextModel):model_arch = gguf.MODEL_ARCH.PANGU_EMBEDdiff --git a/conversion/phi.py b/conversion/phi.pyindex df4bfe809a..7d2532067b 100644--- a/conversion/phi.py+++ b/conversion/phi.py@@ -14,6 +14,7 @@ from .base import MmprojModel, ModelBase, SentencePieceTokenTypes, TextModel, gg@ModelBase.register("PhiForCausalLM")[email protected]("microsoft/phi-2")class Phi2Model(TextModel):model_arch = gguf.MODEL_ARCH.PHI2@@ -36,6 +37,7 @@ class Phi2Model(TextModel):@ModelBase.register("Phi3ForCausalLM", "Phi4ForCausalLMV")[email protected]("microsoft/Phi-3-mini-4k-instruct")class Phi3MiniModel(TextModel):model_arch = gguf.MODEL_ARCH.PHI3@@ -210,6 +212,7 @@ class Phi3MiniModel(TextModel):@ModelBase.register("Phi4ForCausalLMV")+# [TAG_HF_EXAMPLE_MISSING]class Phi4VisionMmprojModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)@@ -336,6 +339,7 @@ class Phi4VisionMmprojModel(MmprojModel):@ModelBase.register("PhiMoEForCausalLM")[email protected]("microsoft/Phi-3.5-MoE-instruct")class PhiMoeModel(Phi3MiniModel):model_arch = gguf.MODEL_ARCH.PHIMOEdiff --git a/conversion/plamo.py b/conversion/plamo.pyindex c4bcbdf06b..31c6455aaf 100644--- a/conversion/plamo.py+++ b/conversion/plamo.py@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("PlamoForCausalLM")[email protected]("pfnet/plamo-13b")class PlamoModel(TextModel):model_arch = gguf.MODEL_ARCH.PLAMO@@ -58,6 +59,7 @@ class PlamoModel(TextModel):@ModelBase.register("Plamo2ForCausalLM", "PLaMo2ForCausalLM")[email protected]("pfnet/plamo-2-1b")class Plamo2Model(TextModel):model_arch = gguf.MODEL_ARCH.PLAMO2@@ -147,6 +149,8 @@ class Plamo2Model(TextModel):@ModelBase.register("Plamo3ForCausalLM", "PLaMo3ForCausalLM")+# [TAG_HF_EXAMPLE_GATED] pfnet/plamo-3-nict-2b-base is gated[email protected]("midorin-Linux/plamo-3-12b-self-merged-base")class Plamo3Model(TextModel):model_arch = gguf.MODEL_ARCH.PLAMO3diff --git a/conversion/plm.py b/conversion/plm.pyindex 3fde487085..bca0147e63 100644--- a/conversion/plm.py+++ b/conversion/plm.py@@ -4,6 +4,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("PLMForCausalLM")[email protected]("PLM-Team/PLM-1.8B-Instruct")class PLMModel(TextModel):model_arch = gguf.MODEL_ARCH.PLMdiff --git a/conversion/pockettts.py b/conversion/pockettts.pyindex 62ecb5acde..1c99e58cfc 100644--- a/conversion/pockettts.py+++ b/conversion/pockettts.py@@ -77,6 +77,7 @@ def _load_hparams(dir_model: Path) -> dict[str, Any]:@ModelBase.register("PocketTTSModel")+# [TAG_HF_EXAMPLE_MISSING] model is gated, and the checkpoint requires cd to subdir, not supported hereclass PocketTTSModel(TextModel):model_arch = gguf.MODEL_ARCH.POCKETTTS@@ -174,6 +175,7 @@ class PocketTTSModel(TextModel):@ModelBase.register("PocketTTSModel")+# [TAG_HF_EXAMPLE_MISSING] model is gated, and the checkpoint requires cd to subdir, not supported hereclass PocketTTSMmprojModel(MmprojModel):has_audio_encoder = Truehas_vision_encoder = Falsediff --git a/conversion/qwen.py b/conversion/qwen.pyindex ead435455d..7a1c11317b 100644--- a/conversion/qwen.py+++ b/conversion/qwen.py@@ -13,6 +13,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("QWenLMHeadModel")[email protected]("Qwen/Qwen-7B")class QwenModel(TextModel):model_arch = gguf.MODEL_ARCH.QWEN@@ -51,6 +52,7 @@ class QwenModel(TextModel):"AudioFlamingo3ForConditionalGeneration","DotsOCRForCausalLM",)[email protected]("Qwen/Qwen2.5-7B-Instruct")class Qwen2Model(TextModel):model_arch = gguf.MODEL_ARCH.QWEN2@@ -71,6 +73,7 @@ class Qwen2Model(TextModel):@ModelBase.register("Qwen2MoeForCausalLM")[email protected]("Qwen/Qwen1.5-MoE-A2.7B")class Qwen2MoeModel(TextModel):model_arch = gguf.MODEL_ARCH.QWEN2MOE@@ -153,6 +156,7 @@ class Qwen2MoeModel(TextModel):@ModelBase.register("Qwen3ForCausalLM", "Qwen3Model")[email protected]("Qwen/Qwen3-8B")class Qwen3Model(Qwen2Model):model_arch = gguf.MODEL_ARCH.QWEN3@@ -251,6 +255,7 @@ class Qwen3Model(Qwen2Model):@ModelBase.register("Qwen3MoeForCausalLM")[email protected]("Qwen/Qwen3-30B-A3B")class Qwen3MoeModel(Qwen2MoeModel):model_arch = gguf.MODEL_ARCH.QWEN3MOE@@ -362,6 +367,7 @@ class _QwenMtpMixin:@ModelBase.register("Qwen3NextForCausalLM")[email protected]("Qwen/Qwen3-Next-80B-A3B-Instruct")class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel):model_arch = gguf.MODEL_ARCH.QWEN3NEXT@@ -421,6 +427,7 @@ class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel):@ModelBase.register("RND1")[email protected]("radicalnumerics/RND1-Base-0910")class RND1Model(Qwen2MoeModel):model_arch = gguf.MODEL_ARCH.RND1@@ -620,16 +627,19 @@ class _Qwen35MRopeMixin:@ModelBase.register("Qwen3_5ForConditionalGeneration", "Qwen3_5ForCausalLM")[email protected]("Qwen/Qwen3.5-9B")class Qwen3_5TextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):model_arch = gguf.MODEL_ARCH.QWEN35@ModelBase.register("Qwen3_5MoeForConditionalGeneration", "Qwen3_5MoeForCausalLM")[email protected]("Qwen/Qwen3.5-35B-A3B")class Qwen3_5MoeTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):model_arch = gguf.MODEL_ARCH.QWEN35MOE@ModelBase.register("DFlashDraftModel")[email protected]("z-lab/Qwen3.5-9B-DFlash")class DFlashModel(Qwen3Model):model_arch = gguf.MODEL_ARCH.DFLASH@@ -699,6 +709,7 @@ class DFlashModel(Qwen3Model):@ModelBase.register("Qwen3DSparkModel")[email protected]("satgeze/Qwen3.6-27B-DSpark")class DSparkModel(DFlashModel):# DSpark = DFlash + a semi-autoregressive Markov headmodel_arch = gguf.MODEL_ARCH.DFLASHdiff --git a/conversion/qwen3tts.py b/conversion/qwen3tts.pyindex d21a505951..1f6b9a1b0e 100644--- a/conversion/qwen3tts.py+++ b/conversion/qwen3tts.py@@ -37,6 +37,7 @@ _ACT2FN = {@ModelBase.register("Qwen3TTSForConditionalGeneration")[email protected]("Qwen/Qwen3-TTS-12Hz-1.7B-Base")class Qwen3TTSTalkerModel(TextModel):model_arch = gguf.MODEL_ARCH.QWEN3TTS@@ -185,6 +186,7 @@ class Qwen3TTSTalkerModel(TextModel):@ModelBase.register("Qwen3TTSForConditionalGeneration")[email protected]("Qwen/Qwen3-TTS-12Hz-1.7B-Base")class Qwen3TTSSpeakerEncoderModel(MmprojModel):has_vision_encoder = Falsehas_audio_encoder = Truediff --git a/conversion/qwen3vl.py b/conversion/qwen3vl.pyindex 9f11757697..4fec708c9f 100644--- a/conversion/qwen3vl.py+++ b/conversion/qwen3vl.py@@ -14,6 +14,7 @@ from .qwenvl import Qwen25AudioModel@ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration", "Qwen3_5ForConditionalGeneration", "Qwen3_5MoeForConditionalGeneration")[email protected]("Qwen/Qwen3-VL-4B-Instruct", "Qwen/Qwen3-VL-30B-A3B-Instruct", "Qwen/Qwen3.5-9B", "Qwen/Qwen3.5-35B-A3B")class Qwen3VLVisionModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)@@ -144,6 +145,7 @@ class Qwen3VLVisionModel(MmprojModel):@ModelBase.register("Qwen3OmniMoeForConditionalGeneration")[email protected]("Qwen/Qwen3-Omni-30B-A3B-Instruct")class Qwen3OmniMmprojModel(Qwen3VLVisionModel, Qwen25AudioModel):has_audio_encoder = Truehas_vision_encoder = True@@ -217,12 +219,14 @@ class Qwen3OmniMmprojModel(Qwen3VLVisionModel, Qwen25AudioModel):@ModelBase.register("Qwen3ASRForConditionalGeneration")[email protected]("Qwen/Qwen3-ASR-0.6B-hf")class Qwen3ASRMmprojModel(Qwen3OmniMmprojModel):has_audio_encoder = Truehas_vision_encoder = False@ModelBase.register("Glm4vForConditionalGeneration", "Glm4vMoeForConditionalGeneration", "GlmOcrForConditionalGeneration")[email protected]("zai-org/GLM-4.1V-9B-Thinking", "zai-org/GLM-4.5V")class Glm4VVisionModel(Qwen3VLVisionModel):def set_gguf_parameters(self):MmprojModel.set_gguf_parameters(self) # skip Qwen3VLVisionModel parameters@@ -246,6 +250,7 @@ class Glm4VVisionModel(Qwen3VLVisionModel):@ModelBase.register("Qwen3VLForConditionalGeneration")[email protected]("Qwen/Qwen3-VL-4B-Instruct")class Qwen3VLTextModel(Qwen3Model):model_arch = gguf.MODEL_ARCH.QWEN3VL@@ -268,6 +273,7 @@ class Qwen3VLTextModel(Qwen3Model):@ModelBase.register("Qwen3VLMoeForConditionalGeneration")[email protected]("Qwen/Qwen3-VL-30B-A3B-Instruct")class Qwen3VLMoeTextModel(Qwen3MoeModel):model_arch = gguf.MODEL_ARCH.QWEN3VLMOE@@ -317,6 +323,7 @@ class Qwen3VLMoeTextModel(Qwen3MoeModel):@ModelBase.register("Qwen3OmniMoeForConditionalGeneration")[email protected]("Qwen/Qwen3-Omni-30B-A3B-Instruct")class Qwen3OmniMoeTextModel(Qwen3VLMoeTextModel):model_arch = gguf.MODEL_ARCH.QWEN3VLMOE@@ -338,6 +345,7 @@ class Qwen3OmniMoeTextModel(Qwen3VLMoeTextModel):@ModelBase.register("Qwen3ASRForConditionalGeneration")[email protected]("Qwen/Qwen3-ASR-0.6B-hf")class Qwen3ASRTextModel(Qwen3VLTextModel):model_arch = gguf.MODEL_ARCH.QWEN3VLdiff --git a/conversion/qwenvl.py b/conversion/qwenvl.pyindex 202a47961b..579a86a99f 100644--- a/conversion/qwenvl.py+++ b/conversion/qwenvl.py@@ -17,6 +17,7 @@ from .base import MmprojModel, ModelBase, TextModel, gguf"Qwen2_5_VLForConditionalGeneration","Qwen2_5OmniModel",)[email protected]("Qwen/Qwen2-VL-2B-Instruct", "Qwen/Qwen2.5-VL-3B-Instruct")class Qwen2VLModel(TextModel):model_arch = gguf.MODEL_ARCH.QWEN2VL@@ -40,6 +41,7 @@ class Qwen2VLModel(TextModel):@ModelBase.register("Qwen2VLModel", "Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")[email protected]("Qwen/Qwen2-VL-2B-Instruct", "Qwen/Qwen2.5-VL-3B-Instruct")class Qwen2VLVisionModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)@@ -161,6 +163,7 @@ class Qwen25AudioModel(MmprojModel):@ModelBase.register("Qwen2_5OmniModel")[email protected]("Qwen/Qwen2.5-Omni-3B")class Qwen25OmniModel(Qwen2VLVisionModel, Qwen25AudioModel):has_audio_encoder = Truehas_vision_encoder = Truediff --git a/conversion/refact.py b/conversion/refact.pyindex 1170cddeb2..d6361512f7 100644--- a/conversion/refact.py+++ b/conversion/refact.py@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("GPTRefactForCausalLM")[email protected]("smallcloudai/Refact-1_6-base")class RefactModel(TextModel):model_arch = gguf.MODEL_ARCH.REFACTdiff --git a/conversion/rwkv.py b/conversion/rwkv.pyindex 2de0aa5346..e6fa84264e 100644--- a/conversion/rwkv.py+++ b/conversion/rwkv.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("Rwkv6ForCausalLM")[email protected]("RWKV/v6-Finch-1B6-HF")class Rwkv6Model(TextModel):model_arch = gguf.MODEL_ARCH.RWKV6@@ -83,6 +84,7 @@ class Rwkv6Model(TextModel):@ModelBase.register("RWKV6Qwen2ForCausalLM")[email protected]("recursal/QRWKV6-32B-Instruct-Preview-v0.1")class RWKV6Qwen2Model(Rwkv6Model):model_arch = gguf.MODEL_ARCH.RWKV6QWEN2@@ -136,6 +138,7 @@ class RWKV6Qwen2Model(Rwkv6Model):@ModelBase.register("Rwkv7ForCausalLM", "RWKV7ForCausalLM")[email protected]("fla-hub/rwkv7-1.5B-world")class Rwkv7Model(TextModel):model_arch = gguf.MODEL_ARCH.RWKV7@@ -261,6 +264,7 @@ class Rwkv7Model(TextModel):@ModelBase.register("RwkvHybridForCausalLM")[email protected]("RWKV-Red-Team/ARWKV-7B-Preview-0.1")class ARwkv7Model(Rwkv7Model):model_arch = gguf.MODEL_ARCH.ARWKV7diff --git a/conversion/sarashina2.py b/conversion/sarashina2.pyindex 05448db812..fdb3e78da6 100644--- a/conversion/sarashina2.py+++ b/conversion/sarashina2.py@@ -12,6 +12,7 @@ from .qwenvl import Qwen2VLVisionModel@ModelBase.register("Sarashina2VisionForCausalLM")[email protected]("sbintuitions/sarashina2.2-vision-3b")class Sarashina2VLTextModel(LlamaModel):model_arch = gguf.MODEL_ARCH.LLAMA@@ -26,6 +27,7 @@ class Sarashina2VLTextModel(LlamaModel):@ModelBase.register("Sarashina2VisionForCausalLM")[email protected]("sbintuitions/sarashina2.2-vision-3b")class Sarashina2VLVisionModel(Qwen2VLVisionModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)diff --git a/conversion/smallthinker.py b/conversion/smallthinker.pyindex 1b0f79aa3e..73d07b51a3 100644--- a/conversion/smallthinker.py+++ b/conversion/smallthinker.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("SmallThinkerForCausalLM")[email protected]("PowerInfer/SmallThinker-4BA0.6B-Instruct")class SmallThinkerModel(TextModel):model_arch = gguf.MODEL_ARCH.SMALLTHINKERdiff --git a/conversion/smolvlm.py b/conversion/smolvlm.pyindex 30e9dca329..0cccb8f6f9 100644--- a/conversion/smolvlm.py+++ b/conversion/smolvlm.py@@ -9,6 +9,7 @@ from .base import MmprojModel, ModelBase, gguf@ModelBase.register("Idefics3ForConditionalGeneration", "SmolVLMForConditionalGeneration")[email protected]("HuggingFaceTB/SmolVLM-Instruct", "HuggingFaceM4/Idefics3-8B-Llama3")class SmolVLMModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)diff --git a/conversion/stablelm.py b/conversion/stablelm.pyindex 6e16378a03..ac3a1ca9e1 100644--- a/conversion/stablelm.py+++ b/conversion/stablelm.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM")[email protected]("stabilityai/stablelm-2-1_6b")class StableLMModel(TextModel):model_arch = gguf.MODEL_ARCH.STABLELMdiff --git a/conversion/starcoder.py b/conversion/starcoder.pyindex 0b4ffd8470..4a726ac36a 100644--- a/conversion/starcoder.py+++ b/conversion/starcoder.py@@ -4,6 +4,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("GPTBigCodeForCausalLM")[email protected]("bigcode/gpt_bigcode-santacoder")class StarCoderModel(TextModel):model_arch = gguf.MODEL_ARCH.STARCODER@@ -19,5 +20,6 @@ class StarCoderModel(TextModel):@ModelBase.register("Starcoder2ForCausalLM")[email protected]("bigcode/starcoder2-3b")class StarCoder2Model(TextModel):model_arch = gguf.MODEL_ARCH.STARCODER2diff --git a/conversion/step3.py b/conversion/step3.pyindex f7cdc997e5..93eb3134e0 100644--- a/conversion/step3.py+++ b/conversion/step3.py@@ -16,6 +16,7 @@ from .qwen import Qwen3Model@ModelBase.register("StepVLForConditionalGeneration", "Step3p7ForConditionalGeneration")[email protected]("stepfun-ai/Step3-VL-10B", "stepfun-ai/Step-3.7-Flash")class Step3VLVisionModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)@@ -91,11 +92,13 @@ class Step3VLVisionModel(MmprojModel):@ModelBase.register("StepVLForConditionalGeneration")[email protected]("stepfun-ai/Step3-VL-10B")class Step3VLTextModel(Qwen3Model):model_arch = gguf.MODEL_ARCH.QWEN3@ModelBase.register("Step3p5ForCausalLM", "Step3p7ForConditionalGeneration")[email protected]("stepfun-ai/Step-3.7-Flash")class Step35Model(TextModel):model_arch = gguf.MODEL_ARCH.STEP35supports_mtp_export = Truediff --git a/conversion/t5.py b/conversion/t5.pyindex 73dcfd1a2c..3466ce49da 100644--- a/conversion/t5.py+++ b/conversion/t5.py@@ -16,6 +16,7 @@ from .base import ModelBase, SentencePieceTokenTypes, TextModel, gguf, logger@ModelBase.register("MT5ForConditionalGeneration")@ModelBase.register("UMT5ForConditionalGeneration")@ModelBase.register("UMT5Model")[email protected]("google-t5/t5-small", "google/flan-t5-small", "google/umt5-small")class T5Model(TextModel):model_arch = gguf.MODEL_ARCH.T5@@ -153,6 +154,7 @@ class T5Model(TextModel):@ModelBase.register("T5EncoderModel")[email protected]("sentence-transformers/sentence-t5-base")class T5EncoderModel(TextModel):model_arch = gguf.MODEL_ARCH.T5ENCODERdiff --git a/conversion/talkie.py b/conversion/talkie.pyindex a970b32d3b..31445243de 100644--- a/conversion/talkie.py+++ b/conversion/talkie.py@@ -11,6 +11,7 @@ from .base import LazyTorchTensor, ModelBase, TextModel, gguf@ModelBase.register("TalkieForCausalLM")[email protected]("lewtun/talkie-1930-13b-it-hf")class TalkieModel(TextModel):model_arch = gguf.MODEL_ARCH.TALKIEdiff --git a/conversion/ultravox.py b/conversion/ultravox.pyindex 347188733a..62819e574d 100644--- a/conversion/ultravox.py+++ b/conversion/ultravox.py@@ -9,6 +9,7 @@ from .base import MmprojModel, ModelBase, TextModel, gguf@ModelBase.register("UltravoxModel")[email protected]("fixie-ai/ultravox-v0_5-llama-3_2-1b")class UltravoxModel(TextModel):model_arch = gguf.MODEL_ARCH.LLAMA # dummy@@ -18,6 +19,7 @@ class UltravoxModel(TextModel):@ModelBase.register("GlmasrModel")[email protected]("zai-org/GLM-ASR-Nano-2512")class GlmASRWhisperEncoderModel(MmprojModel):has_vision_encoder = Falsehas_audio_encoder = True@@ -82,6 +84,7 @@ class GlmASRWhisperEncoderModel(MmprojModel):@ModelBase.register("Qwen2AudioForConditionalGeneration")[email protected]("Qwen/Qwen2-Audio-7B-Instruct")class WhisperEncoderModel(MmprojModel):has_vision_encoder = False # no vision encoderhas_audio_encoder = True@@ -123,6 +126,7 @@ class WhisperEncoderModel(MmprojModel):@ModelBase.register("UltravoxModel")[email protected]("fixie-ai/ultravox-v0_5-llama-3_2-1b")class UltravoxWhisperEncoderModel(WhisperEncoderModel):has_vision_encoder = False # no vision encoderhas_audio_encoder = True@@ -134,6 +138,7 @@ class UltravoxWhisperEncoderModel(WhisperEncoderModel):@ModelBase.register("MERaLiON2ForConditionalGeneration")[email protected]("MERaLiON/MERaLiON-2-3B")class MERaLiONWhisperEncoderModel(WhisperEncoderModel):has_vision_encoder = Falsehas_audio_encoder = True@@ -180,6 +185,7 @@ class MERaLiONWhisperEncoderModel(WhisperEncoderModel):@ModelBase.register("VoxtralForConditionalGeneration")[email protected]("mistralai/Voxtral-Mini-3B-2507")class VoxtralWhisperEncoderModel(WhisperEncoderModel):has_vision_encoder = False # no vision encoderhas_audio_encoder = True@@ -191,6 +197,7 @@ class VoxtralWhisperEncoderModel(WhisperEncoderModel):@ModelBase.register("AudioFlamingo3ForConditionalGeneration")[email protected]("nvidia/audio-flamingo-3-hf")class AudioFlamingo3WhisperEncoderModel(WhisperEncoderModel):def set_gguf_parameters(self):super().set_gguf_parameters()diff --git a/conversion/wavtokenizer.py b/conversion/wavtokenizer.pyindex 7d25447be8..c9a4b505da 100644--- a/conversion/wavtokenizer.py+++ b/conversion/wavtokenizer.py@@ -9,6 +9,7 @@ from .base import ModelBase, TextModel, gguf, logger@ModelBase.register("WavTokenizerDec")[email protected]("novateur/WavTokenizer-large-speech-75token")class WavTokenizerDecModel(TextModel):model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DECdiff --git a/conversion/xverse.py b/conversion/xverse.pyindex fa8a31a133..aa3b338802 100644--- a/conversion/xverse.py+++ b/conversion/xverse.py@@ -11,6 +11,7 @@ from .base import ModelBase, TextModel, gguf@ModelBase.register("XverseForCausalLM")[email protected]("xverse/XVERSE-7B")class XverseModel(TextModel):model_arch = gguf.MODEL_ARCH.XVERSEdiff --git a/conversion/youtuvl.py b/conversion/youtuvl.pyindex cabc44445f..e972610772 100644--- a/conversion/youtuvl.py+++ b/conversion/youtuvl.py@@ -9,6 +9,7 @@ from .base import MmprojModel, ModelBase, gguf, logger@ModelBase.register("YoutuVLForConditionalGeneration")[email protected]("tencent/Youtu-VL-4B-Instruct")class YoutuVLVisionModel(MmprojModel):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)diff --git a/docs/development/HOWTO-add-model.md b/docs/development/HOWTO-add-model.mdindex 270e6b7356..fcc87f1651 100644--- a/docs/development/HOWTO-add-model.md+++ b/docs/development/HOWTO-add-model.md@@ -29,6 +29,7 @@ The required steps to implement for an HF model are:```python@ModelBase.register("MyModelForCausalLM")[email protected]("user/model")class MyModel(TextModel):model_arch = gguf.MODEL_ARCH.MYMODEL```@@ -37,10 +38,13 @@ or```python@ModelBase.register("MyModelForConditionalGeneration")[email protected]("user/model")class MyModel(MmprojModel):model_arch = gguf.MODEL_ARCH.MYMODEL```+The `example` should point to a valid Hugging Face model that will be used for testing. You can add multiple models if necessary. Prefer a non-gated model, or tiny random weights if no such model exists.+2. Define the layout of the GGUF tensors in [constants.py](/gguf-py/gguf/constants.py)Add an enum entry in `MODEL_ARCH`, the model human friendly name in `MODEL_ARCH_NAMES` and the GGUF tensor names in `MODEL_TENSORS`.