Back to Courses
(Ken Cen出品)Generative AI第23部 訓練大語言模型必修 - LoRA & Quantization
Generative AIBeginner

(Ken Cen出品)Generative AI第23部 訓練大語言模型必修 - LoRA & Quantization

ProviderUdemy
Duration6 total hours
PricePaid
#AI#Udemy#Generative AI

About this Course

想讓您的 AI 模型在邊緣設備上跑得更快、更小、更智能?加入我們的課程,帶您從原理到實操,徹底掌握:參數高效、成本低LoRA 僅透過低秩矩陣 AAA 和 BBB 的更新,凍結原始權重,使得微調參數量相比傳統方法大幅縮減,無需重訓整個模型,省時、省算力、省錢。PyTorch 直擊實戰從 torch.nn.utils.parametrize 開始,手把手教你如何在線性層中注入 LoRA 參數化邏輯;以 MNIST 分類示範,讓你快速上手微調流程。細節解密、數據驗證不僅帶你凍結原始權重、訓練與測試,還深入探討如何對比 LoRA 啟用/禁用時的性能差異,並驗證原始權重是否真正保持不變。參數效率深度分析計算並對比原始模型參數量與 LoRA 額外引入的參數量,用數據說話,讓你更加清晰地掌握技術優勢與取捨。完整數學推導與程式實現從 SVD(奇異值分解)基礎概念,到低秩近似矩陣生成,再到 W+BA⋅scaleW + B A\cdot\text{scale}W+BA⋅scale 的數學邏輯,理論與實作一氣呵成。對稱 vs. 非對稱:深度對比兩大量化策略,掌握 S(Scale)與 Z(Zero‐Point)的配置技巧,讓零點映射更“零失真”;PTQ 實戰:手把手演示「訓練後量化」流程:插入觀察器、校準統計、S/Z 計算,一鍵生成 INT8 模型,輕鬆部署;QAT 精髓:揭開「量化感知訓練」原理,在訓練圖中插入“假量化”模塊,讓模型在訓練時就學會“自我修正”,大幅減少量化誤差;

Ratings & Reviews Separation

Compare student feedback directly from the course provider with ratings submitted by the AIWay community.

Udemy RatingsOfficial Provider
4.9
Provider rating
Verified student count: 6 reviews. This data is refreshed regularly via platform sync endpoints.
AIWay CommunityAIWay Platform
0.0
Community rating
We have compiled 0 community student reviews. Feedback is written by verified AIWay accounts who completed this course.

Community Rating Distribution

5
0
4
0
3
0
2
0
1
0

Share Your Feedback

Please log in to submit a review.

Log In / Sign Up

AIWay User Reviews (0)

Loading community reviews...