TensorRT-LLM Learning Path — 12-Week Battle Plan

Total Duration: 12 weeks | ~15 hours weekly (1.5h × 5 weekdays + 4h × 2 weekend days)
GPU Cost: Weeks 1–4, 10–12 use Google Colab Free ($0); Weeks 5–9 use RunPod RTX 3090 (~$15)

Phase One: PyTorch + HuggingFace (Week 1–4)¶

Week 1: Tensor, Autograd, Complete Training Loop¶

Goal: Write trainable neural networks from blank notebooks independently

• Day 1 (Mon): Environment setup + tensor basics (1.5hr)
• Day 2 (Tue): Autograd and gradient calculation (1.5hr)
• Day 3 (Wed): Dataset and DataLoader (1.5hr)
• Day 4 (Thu): Build first neural network (1.5hr)
• Day 5 (Fri): Training loop—Loss + Optimizer (1.5hr)
• Day 6 (Sat): Complete MNIST classifier project (4hr)
• Day 7 (Sun): Review + paper intro (4hr)

Week 1 Output: GitHub repo, MNIST MLP, test acc > 97%

Week 2: HuggingFace Transformers¶

Goal: Run pretrained models for inference; feel the gap between large and small models

(Details on Tokenizer deep-dive, Model + forward pass, AutoModelForSequenceClassification, inference optimization, classification pipeline project)

Week 2 Output: Text classification inference pipeline, CPU vs GPU benchmarks

Week 3: Fine-tuning Intro¶

Goal: Full fine-tuning and LoRA on pretrained models; understand parameter update differences

(Details on fine-tuning concepts, Trainer API, manual training loops, LoRA theory and implementation, benchmark LoRA vs full fine-tuning)

Week 3 Output: LoRA fine-tuning pipeline with comparison metrics

Week 4: Integration + GitHub Portfolio¶

Goal: Integrate three weeks into complete GitHub repo; prep for Week 5 inference frameworks

(Repo structure, inference benchmarking, README, technical blog post, vLLM environment familiarization)

Week 4 Output: Complete GitHub repo + first technical blog

Phase Two: LLM Inference Frameworks (Week 5–9)¶

Week 5: Inference Principles¶

• KV cache mechanism and implementation
• Batching strategies: static vs continuous
• Quantization methods (GPTQ, AWQ, INT8 differences)
• PagedAttention conceptual understanding
• Related research papers

Week 6–7: vLLM Hands-On Deployment¶

• Install vLLM on RunPod RTX 3090
• Deploy Llama-3.1-8B, first inference request
• Systematic benchmark: tokens/sec, latency, GPU usage
• Test quantization impact on speed and quality
• Understand OpenAI-compatible API server

Week 8: llama.cpp Hands-On¶

• Install llama.cpp, download GGUF models
• vLLM vs llama.cpp: applicable scenarios, latency, CPU availability
• GGUF format and quantization levels (Q4_K_M, Q8_0, etc)

Week 9: Integration Project + Technical Blog¶

• Connect traffic violation system with LLM for violation classification
• Complete benchmark: vLLM vs llama.cpp across batch sizes
• Second technical blog post

Phase Two Output: LLM inference service project + benchmarks + second blog

Phase Three: CUDA Basics (Week 10–12)¶

Week 10–11: CUDA Core Concepts¶

• GPU architecture: SM, warp, thread/block/grid relationships
• Memory hierarchy: global/shared/register/local memory
• First vector addition kernel with numba.cuda
• Coalesced memory access and bank conflicts
• GEMM (matrix multiplication) tiling optimization

Week 12: Build a Showcase Kernel¶

• Implement tiled matrix multiplication in CUDA C or numba.cuda
• Compare naive vs tiled versions
• Explain kernel parallelization logic to interviewers
• Push to GitHub—proof of "CUDA fundamentals" on resume

Phase Three Output: CUDA kernel implementation + performance comparisons

Three Months Later: Resume Additions¶

New Skills:
PyTorch, HuggingFace Transformers, LoRA fine-tuning, vLLM, llama.cpp, CUDA basics, LLM inference optimization

New Projects:
* MNIST MLP Classifier (97%+ test accuracy)
* LLM Inference Benchmark (vLLM vs llama.cpp with throughput/latency data)
* Traffic Violation AI (integrated LLM violation classification)
* LoRA Fine-tuning Pipeline (full vs LoRA comparison)

Blog Articles:
* PyTorch Basics to LoRA Fine-tuning: 4-Week Learning Log
* Deploying LLM Inference with vLLM: Benchmarks & Insights