Document Sync by Tina

Author: Chivier

docs/stable/intro.md

@@ -9,6 +9,10 @@ sidebar_position: 1
 
 ServerlessLLM is a **fast** and **easy-to-use** serving system designed for **affordable** multi-LLM serving, also known as LLM-as-a-Service. ServerlessLLM is ideal for environments with multiple LLMs that need to be served on limited GPU resources, as it enables efficient dynamic loading of LLMs onto GPUs. By elastically scaling model instances and multiplexing GPUs, ServerlessLLM can significantly reduce costs compared to traditional GPU-dedicated serving systems while still providing low-latency (Time-to-First-Token, TTFT) LLM completions.
 
+ServerlessLLM now supports NVIDIA and AMD GPUs, including following hardware:
+* NVIDIA GPUs: Compute Capability 7.0+ (e.g, V100, A100, RTX A6000, GeForce RTX 3060)
+* AMD GPUs: ROCm 6.2.0+ (tested on MI100s and MI200s)
+
 ## Documentation
 
 ### Getting Started
@@ -25,6 +29,7 @@ ServerlessLLM is a **fast** and **easy-to-use** serving system designed for **af
 ### ServerlessLLM Store
 
 - [Quickstart](./store/quickstart.md)
+- [ROCm Installation(Experimental)](./store/installation_with_rocm.md)
 
 ### ServerlessLLM CLI
 

docs/stable/store/installation_with_rocm.md

@@ -4,6 +4,15 @@ sidebar_position: 1
 
 # Installation with ROCm (Experimental)
 
+## Latest Tested Version
++ v0.5.1
+
+## Tested Hardware
++ OS: Ubuntu 22.04
++ ROCm: 6.2
++ PyTorch: 2.3.0
++ GPU: MI100s (gfx908), MI200s (gfx90a)
+
 ## Build the wheel from source and install
 ServerlessLLM Store (`sllm-store`) currently provides experimental support for ROCm platform. Due to an internal bug in ROCm, serverless-llm-store may face a GPU memory leak in ROCm before version 6.2.0, as noted in [issue](https://github.com/ROCm/HIP/issues/3580).
 
@@ -162,13 +171,6 @@ cd ServerlessLLM/sllm_store/build
 ctest --output-on-failure
 ```
 
-## Tested Hardware
-+ OS: Ubuntu 22.04
-+ ROCm: 6.2
-+ PyTorch: 2.3.0
-+ GPU: MI100s (gfx908), MI200s (gfx90a)
-
-
 ## Known issues
 
 1. GPU memory leak in ROCm before version 6.2.0.