Memuat…

Course Forum Asisten Sumber Cara belajar RAG Peta modul

Masuk Mulai →

socrates

Semua tentang AI, satu jalur belajar. Gratis, bersumber terbuka, dan setiap jawaban punya rujukan.

Navigasi

Course
Forum
Asisten
Sumber
Cara belajar
RAG
Peta modul

Dibangun dengan

Next.js · Tailwind ·
Lenis + GSAP · SQLite + BM25
(RAG hybrid: cosine + BM25, RRF)

© 2026 socratesgratis · terbuka · sumber selalu dicantumkan

LLM Infrastructure — socrates · socrates

LLM Infrastructure

Lima babak sistem. Nol sampai cluster.

I0–I4: dari kernel komputasi, vLLM, hingga cluster serving multi-GPU skala produksi.

Modul: 5
Lesson: 24
Project: 5
Durasi: ±7.5 minggu

Progress Lintasan Inti0/24 lesson

Fondasi Hardware & Komputasi GPU untuk LLM

Arsitektur GPU, VRAM, memory bandwidth, Tensor Cores, presisi angka (FP16/BF16/FP8)

Buka bab lengkap →

Anatomi GPU: SM, Warp, Tensor Cores & VRAM Hierarchy Bottleneck Utama: Memory Bandwidth vs Compute FLOPS Presisi Angka: Dari FP32, FP16, BF16 hingga FP8 & FP4 Project: GPU Memory Bandwidth & FLOPS Profiler

1.5 mgg0/4 selesai

GPU Kernels: Dari CUDA ke Triton & FlashAttention

Kernel vs Engine, penulisan kernel CUDA/Triton, fused operations, FlashAttention-1/2/3

Buka bab lengkap →

Kernel GPU vs Inference Engine: Memahami Batas & Peran Fused Operations: Mengeliminasi Pembacaan Memori Berulang FlashAttention: Fast and Memory-Efficient Exact Attention OpenAI Triton: Menulis Kernel GPU Tanpa C++/CUDA Murni Project: Menulis Fused Softmax Kernel dengan Triton

1.5 mgg0/5 selesai

Inference Engines: vLLM, PagedAttention & Continuous Batching

Arsitektur vLLM, PagedAttention, continuous batching, chunked prefill, disaggregated serving

Buka bab lengkap →

Anatomi vLLM: Mesin Inferensi LLM Skala Produksi PagedAttention: Manajemen Memori KV Cache Tanpa Fragmentasi Continuous Batching (Iteration-Level) vs Static Batching Disaggregated Serving: Memisahkan Prefill & Decode Node Project: Benchmark vLLM Server & Concurrency Stress Test

1.5 mgg0/5 selesai

Distributed Serving & Parallelism Strategies

Tensor Parallelism, Pipeline Parallelism, Expert Parallelism (MoE), Interconnect (NVLink, PCIe)

Buka bab lengkap →

Kapan dan Mengapa Butuh Distributed GPU Serving?Tensor Parallelism (TP): Megatron-LM Style Pipeline Parallelism (PP): Pembagian Layer Lintas Device Mixture-of-Experts (MoE) & Expert Parallelism (EP)Project: Multi-GPU Cluster Simulation & TP Profiling

1.5 mgg0/5 selesai

Advanced Optimization: Quantization, Speculative Decoding & Production LLMOps

Kuantisasi (AWQ, GPTQ, FP8), Speculative Decoding, RadixAttention, Observability Cluster

Buka bab lengkap →

Teknik Kuantisasi Modern: AWQ, GPTQ, SmoothQuant & FP8 Speculative Decoding: Mempercepat Generasi Tanpa Mengubah Model RadixAttention & Structured Caching (SGLang Pattern)Kapasitas & Sizing Cluster Produksi Capstone: High-Throughput LLM Production Cluster

1.5 mgg0/5 selesai

Mode lesson

TEXT: — baca terstruktur
VIDEO: — playlist kurasi
ARTICLE: — paper & medium
PROJECT: — langsung bangun

PROJECT = learning-by-doing: kerjakan, commit, dan tandai selesai.

Mulai dari nol — tanpa daftar, tanpa paywall

Mulai: I0 Fondasi Hardware & Komputasi GPU →

Halaman lesson per-babak + tracking progress.