docs: update README file

README.md

@@ -1,23 +1,196 @@
 # InfiniTrain
 
-## 🚀 Getting Started
+A from-scratch C++ training framework for large-scale models with multi-dimensional distributed parallelism.
 
-### 🛠️ Build Instructions
+## 🚀 Quick Start
+
+### System Requirements
+
+#### Hardware Requirements
+
+- **Recommended**: NVIDIA Ampere-class GPUs (A100/A800) or newer
+
+#### Software Requirements
+
+- **CUDA / NCCL**: Latest stable versions
+- **gcc / g++**: Version **13+**
+- **CMake**: Version **3.13+**
+
+### Installation
 
 ```bash
 mkdir build
 cd build
-cmake ..           # Use -DUSE_CUDA=ON to enable CUDA support
-make
+cmake .. -DUSE_CUDA=ON -DUSE_NCCL=ON
+make -j
 ```
 
-## 🧪 Running Examples
+Build Options:
+
+- `USE_CUDA=ON`
+
+  Enable CUDA backend support.
+
+- `USE_NCCL=ON`
+
+  Enable NCCL-based distributed communication.
+
+> Both options are optional and can be disabled for CPU-only builds.
+
+## ✨ InfiniTrain Overview
+
+### ✔ Model Support
+
+- GPT-2
+- LLaMA 3
+
+### ✔ Precision Support
+
+- Full-precision training
+  - FP32
+  - BF16
+- Mixed-precision training
+  - BF16 computation with FP32 accumulation
+  - Autocast-based precision control
+
+### ✔ Distributed Training
+
+InfiniTrain provides a flexible distributed execution model with multiple parallelism strategies.
+
+- Data Parallelism
+  - Parameter-server-style Data Parallel (DP)
+  - Collective-based Data Parallel (DDP)
+- Model Parallelism
+  - Tensor Parallelism (TP)
+  - Sequence Parallelism (SP)
+  - Pipeline Parallelism (PP)
+    - GPipe scheduling
+    - 1F1B scheduling
+    - Virtual Pipeline Parallelism (vPP)
+- Arbitrary combination of **DDP + TP + SP + PP**
+
+### ✔ Core Components
+
+- Multi-backend support
+  - CPU
+  - CUDA
+- Multi-node distributed training
+- Kernel registration and dispatch mechanism
+- Automatic differentiation (Autograd)
+- Automatic mixed precision (Autocast)
+
+### ✔ Performance Optimizations
+
+- Computation-communication overlap
+  - Explicit scheduling to hide communication latency
+- DDP gradient bucketing
+  - Gradient synchronization is deferred and bucketed
+  - Reducer-controlled bucket scheduling for overlap optimization
+
+### ✔ Training & Execution Modes
+
+- Standard training mode
+- `no_grad` inference mode
+  - Forward-only execution without gradient tracking
+
+### ✔ Debugging & Tooling
+
+- Built-in profiler
+  - Native framework profiler for kernel-level time analysis
+- Automated performance testing
+  - One-click execution of all benchmark cases
+  - Automatic log analysis
+  - Performance metrics aggregated and reported to Feishu spreadsheets
+
+## 🏋️ Training
 
 Each model in the `example/` directory is compiled into an independent executable.  
-For instance, the `mnist` example will produce a binary named `mnist`.
+For example, the `llama3` example produces a binary named `llama3`.
+
+To view available runtime options:
+
+```bash
+./llama3 --help
+```
 
-You can view the available runtime options by executing:
+### Getting Started
+
+The following examples demonstrate **LLaMA 3 supervised fine-tuning (SFT)** using InfiniTrain.
+
+#### Single-node Training Example
+
+```bash
+./llama3 \
+  --device cuda \
+  --input_bin [training_data_path] \
+  --llmc_filepath [model_path] \
+  --num_iteration 10
+
+```
+
+#### Multi-nodes Training Example (3D parallel)
+
+```bash
+./infini_run \
+  --nnodes=2 \
+  --nproc_per_node=1 \
+  --node_rank=[rank_id] \
+  -- ./llama3 \
+     --device cuda \
+     --input_bin [training_data_path] \
+     --llmc_filepath [model_path] \
+     --num_iteration 10 \
+     --nthread_per_process 8 \
+     --batch_size 40 \
+     --total_batch_size 10240 \
+     --tensor_parallel 2 \
+     --pipeline_parallel 2 \
+     --sequence_parallel
+```
+
+### Parallelism Strategies
+
+#### Distributed Data Parallelism (DDP)
+
+```bash
+--nthread_per_process 8 	# ddp_size = nthread_per_process / (tensor_parallel × pipeline_parallel)
+```
+
+#### Tensor Parallelism (TP)
 
 ```bash
-./mnist --help
-```
+--tensor_parallel 4        # 4-way tensor parallelism
+--sequence_parallel        # Enable sequence parallelism (recommended with TP)
+```
+
+#### Pipeline Parallelism (PP)
+
+```bash
+--pipeline_parallel 8     		# 8 pipeline stages
+--virtual_pipeline_parallel 4  	# Virtual pipeline for better load balancing
+```
+
+#### Combining Parallelism Strategies
+
+Multiple parallelism strategies (DDP, TP, SP, PP) can be freely combined to scale training across devices and nodes.
+
+## 🗺 Roadmap
+
+- **2025/03/10** — InfiniTrain **v0.1.0**
+
+  Initial framework prototype with MNIST CPU training.
+
+- **2025/04/30** — InfiniTrain **v0.3.0**
+
+  Added Autograd support and GPT-2 training on CPU/CUDA.
+
+- **2025/07/09** — InfiniTrain **v0.4.0**
+
+  Introduced kernel registration, LLaMA training on CPU/CUDA, BF16 precision, and Data Parallelism.
+
+- **2025/12/xx** — InfiniTrain **v0.5.0**
+
+  Added Autocast, multi-dimensional distributed parallelism
+   (DDP, TP, SP, PP with GPipe / 1F1B / vPP),
+   multi-node training, `no_grad` mode,
+   and communication–computation overlap with bucketed gradient synchronization.