ML Engineering: GPU, LLM serving, kompressiya, distributed training

Kurs haqida

Kurs sanoat miqyosidagi modellarni o'qitish va deploy qilish uchun zarur bo'lgan tizim muhandisligini qamrab oladi. GPU arxitekturasi, CUDA-yadrolarini optimallashtirish, distributed training texnikalari (data/model/pipeline parallelism), model kompressiyasi va yuqori unumdor LLM inference tizimlari o'rganiladi. Kurs Amaliyot 3 jamoasining muvaffaqiyatini bevosita ta'minlaydi va samarali arxitekturalar bo'yicha research-mavzular uchun kirish hisoblanadi.

Nimalarni o'rganasiz

✓CUDA-yadrolarini profillash va GPU hisoblashlaridagi bottleneck'larni bartaraf etish

✓1B dan ortiq parametrli model uchun FSDP yoki tensor parallelism yordamida distributed training amalga oshirish

✓PTQ kvantizatsiyasini qo'llash va vazifada sifat degradatsiyasini o'lchash

✓Maqsadli latency'ga erishish uchun continuous batching va KV-cache bilan LLM serving tizimini sozlash

Asosiy mavzular

•GPU arxitekturasi: SIMT, warp, shared memory, memory coalescing

•CUDA programming: kernels, streams, profillash (Nsight)

•Aralash aniqlik (mixed precision): FP16, BF16, FP8, loss scaling

•Distributed training: DDP, FSDP, tensor/pipeline parallelism

•Transformerlarni optimallashtirish: FlashAttention, KV-cache, continuous batching

•Kvantizatsiya: post-training (PTQ), quantization-aware training (QAT)

•Pruning: structured/unstructured, magnitude va gradient-based

•Knowledge distillation: response-based, feature-based

•LLM serving: vLLM, TensorRT-LLM, speculative decoding

Ushbu tavsif avtomatik tarzda yaratilgan va hali o'qituvchi tomonidan tekshirilmagan — bu muhokama uchun qoralama.

Kurs haqida

Nimalarni o'rganasiz

Asosiy mavzular

Fikr-mulohaza