haotian/unsloth
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

1.修改微调模型\n2.保存模型时添加限制使用GPU显存参数,以确保不会爆显存

Browse Source
This commit is contained in:
haotian 2025-07-11 18:02:58 +08:00
parent 4e914bb7b3
commit f318eedac5
1 changed files with 8 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
003加载自己的数据集微调.py
View File

@ -25,7 +25,9 @@ dataset = dataset.map(to_chatml)
# 加载预训练模型 # 加载预训练模型
model, tokenizer = FastLanguageModel.from_pretrained( model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", # model_name = "unsloth/Qwen3-1.7B-unsloth-bnb-4bit",
model_name = "unsloth/Qwen3-8B-unsloth-bnb-4bit",
# model_name = "unsloth/Qwen3-4B-unsloth-bnb-4bit",
max_seq_length = 2048, # Context length - can be longer, but uses more memory max_seq_length = 2048, # Context length - can be longer, but uses more memory
load_in_4bit = True, # 4bit uses much less memory , 启用QLoRA load_in_4bit = True, # 4bit uses much less memory , 启用QLoRA
load_in_8bit = False, # A bit more accurate, uses 2x memory load_in_8bit = False, # A bit more accurate, uses 2x memory
@ -93,8 +95,11 @@ _ = model.generate(
streamer = TextStreamer(tokenizer, skip_prompt = True), streamer = TextStreamer(tokenizer, skip_prompt = True),
) )
# model.cpu()
model.save_pretrained_gguf( model.save_pretrained_gguf(
"ollama_model", "Qwen3-8B",
tokenizer, tokenizer,
# quantization_method="q4_k_m" # 或 "q8_0" # 量化模式--默认 q8_0, 可选f16, "q4_k_m", "q8_0", "q5_k_m", # quantization_method="q4_k_m", # 或 "q8_0" # 量化模式--默认 q8_0, 可选f16, "q4_k_m", "q8_0", "q5_k_m",
maximum_memory_usage=0.7 # 限制使用 GPU 显存为总容量的 50%
) )