ai-learning-repo

This repo contains code related to my learnings in GenAI. I've following these tasks and I will be updating each task in a python file

Tasks

Code. Plot. Break. Repeat.

---

Tokenization & Embeddings

• Build a byte-pair encoder (BPE) and train your own subword vocab
• Write a token visualizer to map words/chunks → token IDs
• Compare one-hot vs learned embeddings, and
  • Plot cosine distances between token vectors

---

Positional Embeddings

• Implement & compare:
  • Sinusoidal
  • Learned
  • RoPE
  • ALiBi
• Animate a toy sequence being position-encoded in 3D
• Ablate positions — watch attention collapse

---

Self-Attention & Multihead Attention

• Hand-wire dot-product attention for one token
• Scale to multi-head attention, plot per-head weight heatmaps
• Mask out future tokens, verify causal property

---

Transformers: Q, K, V & Stacking

• Stack your Attention + LayerNorm + Residual → single-block transformer
• Generalize to n-block “mini-former” on toy data
• Dissect Q, K, V — swap them, break them, see what explodes

---

Sampling Parameters: Temperature / Top-k / Top-p

• Build a sampler dashboard — interactively tune temp/k/p
• Plot entropy vs output diversity
• Set temp = 0 (argmax) — watch repetition set in

---

KV Cache (Fast Inference)

• Record & reuse KV states; measure speedup vs no-cache
• Build a cache hit/miss visualizer for token streams
• Profile cache memory cost for long vs short sequences

---

Long-Context Tricks: Infini-Attention / Sliding Window

• Implement sliding-window attention; measure loss on long docs
• Benchmark memory-efficient attention (recompute, flash)
• Plot perplexity vs context length, find the context collapse point

---

Mixture of Experts (MoE)

• Code a 2-expert router layer; route tokens dynamically
• Plot expert utilization histograms
• Simulate sparse vs dense routing, measure FLOP savings

---

Grouped Query Attention

• Convert your mini-former to grouped query layout
• Measure speed vs vanilla multi-head on large batch
• Ablate number of groups, plot latency

---

Normalization & Activations

• Hand-implement:
  • LayerNorm
  • RMSNorm
  • SwiGLU
  • GELU
• Ablate each — observe train/test loss impact
• Plot activation distributions layer-wise

---

Pretraining Objectives

• Train:
  • Masked LM
  • Causal LM
  • Prefix LM
• Plot loss curves — compare who learns “English” faster
• Generate samples from each — note quirks

---

Finetuning vs Instruction Tuning vs RLHF

• Fine-tune on a small custom dataset
• Instruction-tune by prepending tasks (e.g., “Summarize: …”)
• RLHF:
  • Hack a reward model
  • Run PPO for 10 steps
  • Plot reward curve

---

Scaling Laws & Model Capacity

• Train tiny / small / medium models
• Plot loss vs size
• Benchmark time, VRAM, throughput
• Extrapolate scaling curve — how “dumb” can you go?

---

Quantization

• Implement PTQ (Post-Training Quantization)
• Implement QAT (Quantization-Aware Training)
• Export to GGUF / AWQ
• Plot accuracy drop vs compression ratio

---

Inference / Training Stacks

• Port a model from HuggingFace → Deepspeed → vLLM → ExLlama
• Profile throughput, VRAM, latency across all three

---

Synthetic Data

• Generate toy data, add noise, dedupe, create eval splits
• Visualize model learning curves on real vs synthetic data

---