Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
docs(save): correct save_method values in docstrings (16bit/4bit -> merged_16bit/merged_4bit) (#10394)
陈志谦100a5cf75164
diff --git a/unsloth/save.py b/unsloth/save.pyindex 082b5a793e4..d24e9452acb 100644--- a/unsloth/save.py+++ b/unsloth/save.py@@ -2346,8 +2346,8 @@ def unsloth_save_pretrained_merged(converted to float16 with as few overhead as possible.Choose for `save_method` to be either:- 1. `16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.- 2. `4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.+ 1. `merged_16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.+ 2. `merged_4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.3. `lora`: Save the LoRA adapter itself, with no merging: `adapter_config.json`plus `adapter_model.safetensors`, and no base-model weights at all (the adapteris written as `adapter_model.bin` instead when `safe_serialization = False`).@@ -2485,8 +2485,8 @@ def unsloth_push_to_hub_merged(converted to float16 with as few overhead as possible.Choose for `save_method` to be either:- 1. `16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.- 2. `4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.+ 1. `merged_16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.+ 2. `merged_4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.3. `lora`: Save the LoRA adapter itself, with no merging: `adapter_config.json`plus `adapter_model.safetensors`, and no base-model weights at all (the adapteris written as `adapter_model.bin` instead when `safe_serialization = False`).@@ -5766,8 +5766,8 @@ def unsloth_generic_save_pretrained_merged(converted to float16 with as few overhead as possible.Choose for `save_method` to be either:- 1. `16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.- 2. `4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.+ 1. `merged_16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.+ 2. `merged_4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.3. `lora`: Save the LoRA adapter itself, with no merging: `adapter_config.json`plus `adapter_model.safetensors`, and no base-model weights at all (the adapteris written as `adapter_model.bin` instead when `safe_serialization = False`).@@ -5905,8 +5905,8 @@ def unsloth_generic_push_to_hub_merged(converted to float16 with as few overhead as possible.Choose for `save_method` to be either:- 1. `16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.- 2. `4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.+ 1. `merged_16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.+ 2. `merged_4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.3. `lora`: Save the LoRA adapter itself, with no merging: `adapter_config.json`plus `adapter_model.safetensors`, and no base-model weights at all (the adapteris written as `adapter_model.bin` instead when `safe_serialization = False`).