logo
17
4
WeChat Login

微调课程 - Lesson 4 v2 量化

网络相关七剑下天山

uv tool install huggingface-hub[cli] HF_ENDPOINT=https://hf-mirror.com hf download unslothai/1 HF_ENDPOINT=https://hf-mirror.com hf download unslothai/other HF_ENDPOINT=https://hf-mirror.com hf download unslothai/repeat HF_ENDPOINT=https://hf-mirror.com hf download unslothai/vram-40 HF_ENDPOINT=https://hf-mirror.com hf download unsloth/qwen2.5-0.5b-instruct HF_ENDPOINT=https://hf-mirror.com hf download unsloth/qwen2.5-0.5b-instruct-bnb-4bit HF_ENDPOINT=https://hf-mirror.com hf download unsloth/qwen2.5-0.5b-instruct-unsloth-bnb-4bit

目录结构

1. 1-calculate-vram-and-flops - 显存和算力计算

计算模型的显存需求和算力消耗,帮助选择合适的硬件配置。

  • 1-calculate-vram.txt: 包含显存计算器链接和GPU/NPU算力表格,用于估算模型训练和推理所需的显存
  • 2-calculate-flops/: 算力计算模块

2. 2-qat - 量化感知训练 (Quantization Aware Training)

使用量化感知训练技术对模型进行微调

3. 3-llamacpp - Llama.cpp

使用Llama.cpp框架进行GGUF格式转换

4. 4-llamafactory - LlamaFactory 无代码微调

使用LlamaFactory进行无代码模型微调

注意事项

  1. 在执行脚本前请执行 source /etc/network_turbo
  2. 确保在AutoDL平台上运行

About

No description, topics, or website provided.
Language
Python77.6%
Shell22.4%