Puneet Mangla

Table of Contents Toward DeepSeek-R1: Building Thinking Models Using GRPO How Group Relative Policy Optimization (GRPO) Works Core GRPO Objective Group Relative Advantage Estimation Clipped Surrogate Objective KL Divergence Regularization Why GRPO Excels in Mathematical Reasoning Intrinsic Support for Verifiable…

Table of Contents DeepSeek-V3 from Scratch: Implementing the DeepSeek Trainer Training Infrastructure with Hugging Face Trainer Implementation: Custom Trainer and Data Handling Training Configuration Data Preparation: TinyStories Dataset Training DeepSeek-V3 from Scratch and Saving Checkpoints Testing Generation: Bringing the Model…

Table of Contents DeepSeek-V3 from Scratch: Building the Architecture in PyTorch The Transformer Block: Combining MLA and MoE The Complete Model: From Tokens to Logits Weight Tying: Sharing Token Embeddings and Output Weights Multi-Token Prediction Integration Implementing the DeepSeek-V3 Architecture…

Table of Contents Building a Multimodal Chatbot with Qwen3-VL Instruct and Thinking Models Qwen3-VL Vision-Language Model: Architecture, Training, and Capabilities Qwen3-VL Architecture Overview: SigLIP2 Vision Encoder and Multimodal Transformer Design Training Pipeline: Multimodal Pretraining with Image-Text and Video-Text Data Performance…

Table of Contents Building an Intelligent Chatbot with Qwen3 Instruct and Thinking Models Qwen3 Model Details Qwen3 Model Architecture: Dense Models, MoE Variants, and Dual-Mode Reasoning Qwen3 Training Data and Post-Training Pipeline for Reasoning and Instruction Following Qwen3 Benchmark Performance:…

Table of Contents Building and Training a Kimi-K2 Model Using DeepSeek-V3 Components Kimi-K2 vs DeepSeek-V3: Key Architecture Differences in LLM Design Mixture of Experts Scaling in Kimi-K2: Model Size, Sparsity, and Efficiency Attention Head Optimization in Kimi-K2 for Efficient Long-Context…

Table of Contents Autoregressive Model Limits and Multi-Token Prediction in DeepSeek-V3 Why Next-Token Prediction Limits DeepSeek-V3 Multi-Token Prediction in DeepSeek-V3: Predicting Multiple Tokens Ahead DeepSeek-V3 Architecture: Multi-Token Prediction Heads Explained Gradient Insights for Multi-Token Prediction in DeepSeek-V3 DeepSeek-V3 Training vs.…

Table of Contents DeepSeek-V3 from Scratch: Mixture of Experts (MoE) The Scaling Challenge in Neural Networks Mixture of Experts (MoE): Mathematical Foundation and Routing Mechanism SwiGLU Activation in DeepSeek-V3: Improving MoE Non-Linearity Shared Expert in DeepSeek-V3: Universal Processing in MoE…

Table of Contents Build DeepSeek-V3: Multi-Head Latent Attention (MLA) Architecture The KV Cache Memory Problem in DeepSeek-V3 Multi-Head Latent Attention (MLA): KV Cache Compression with Low-Rank Projections Query Compression and Rotary Positional Embeddings (RoPE) Integration Attention Computation with Multi-Head Latent…