Repository navigation
依第一次真的 GPU 實測(RTX 4060 Laptop 8 GB、Windows 11,10 項中 9 項通過)做的修正。
修正
eval gpu:VRAM 峰值改成「最高減最低讀數」;量不到 VRAM 變化時只提出警告,不再判為失敗(實際裝置以引擎記錄為準)。這次唯一的失敗就是這個誤判。eval gpu的 Whisper 檢查改用真的英文語音(LibriSpeech 片段,需要pyarrow),要求 WER ≤ 0.5;embedding 檢查改用一對一的問答,要求 recall@1 ≥ 0.8。- llm-finetune 0.1.2、llm-preference 0.1.1:最低 VRAM 宣告從 8 GB 改成 6 GB(0.5B QLoRA 實測 1.4–1.6 GB),6 GB 的顯示卡也能用;gradient checkpointing 改用
use_reentrant=False,不再出現警告。 - text-classification 0.1.1:記錄 loss 改用
loss.item()。
實測結果:https://honguan.github.io/autoforge/GPU/
升級並重新驗證
pip install -U autoforge-trainer pyarrow
autoforge eval gpu
完整變更紀錄:https://github.com/Honguan/autoforge/blob/main/CHANGELOG.md