MCP Technical Overview
- Concept, Evolution, and Significance of MCP
- MCP Architecture, Core Components, and Function Types
- MCP Client and MCP Server
- How Different Roles Use MCP
About 8 min
EasyR1 + Verl + Ray + QwenVL + GRPO
- Background Introduction
- GRPO Four Main Steps
- Implementation of GRPO Training Code Using EasyR1
- Practical Record of GRPO Training Details
About 5 min
SFTTrainer Source Code Exploration: Prepare Train
- Prepare Train Overall Logic
- Prepare Train Code Details
- _inner_training_loop
- training_step
- compute_loss
- PeftModelForCausalLM.forward
- Linear4bit.forward
About 5 min
SFTTrainer Source Code Exploration: Prepare Dataset
- Prepare Dataset Overall Logic
- Prepare Dataset Code Details
- SFTTrainer.init
- DataCollatorForLanguageModeling
- _prepare_dataset
About 3 min
SFTTrainer Source Code Exploration: Prepare Model
- Prepare Model Overall Logic
- Prepare Model Code Details
- _prepare_peft_model
- PeftModelForCausalLM.init
- PeftModel.init
- LoraModel.init
- Linear4bit.init
- LoraLayer.init(self, base_layer)
About 5 min
QLoRA Code Implementation and Process Analysis
- Background Introduction: QLoRA / Base Model / Dataset
- QLoRA Code Implementation
- QLoRA Process Analysis
- QLoRA Application Value
- QLoRA Questions and Thoughts
- QLoRA Details Supplement
About 11 min
Lightweight Visualization Tool for Deep Learning: wandb
- What is wandb
- Common Functions
- How to Use
About 2 min
GRPO + Unsloth + vLLM
- How GRPO Works
- GRPO vs PPO
- Three Revolutionary Designs of GRPO
- GRPO Code Implementation
About 10 min
Distributed Training Part 5: Introduction to GPU
- GPU Architecture
- How to Improve Performance with Kernels
- Fused Kernels
- Flash Attention
About 6 min
Distributed Training Part 4: Parallel Strategies
- Five Dimensions of Parallelization Strategies
- batch dimension
- hidden_state dimension
- sequence dimension
- model_layer dimension
- model_expert dimension
- Optimal Training Configuration
- Tensor Parallelism（TP）
- Sequence Parallelism (SP)
- Context Parallelism (CP)
- Pipeline parallelism (PP)
- Expert Parallelism (PP)
About 9 min