add attention backend supporting matrix in the doc

docs/backend/attention_backend.md

@@ -0,0 +1,42 @@
+# Attention Backend
+
+## Supporting matrix for different attention backend
+
+| **Backend**              | **Page Size > 1** | **Spec Decoding** | **MLA** | **Sliding Window** |
+|--------------------------|-------------------|-------------------|--------|--------------------|
+| **FlashInfer (Default)** | ✅                | ✅                | ✅     | ✅                 |
+| **FA3**                  | ✅                | ✅                | ✅     | ✅                 |
+| **FlashMLA**             | ✅                | ❌                | ✅     | ❌                 |
+| **Triton**               | ❌                | ✅                | ✅     | ❌                 |
+| **Torch Native**         | ❌                | ❌                | ❌     | ❌                 |
+
+*Note: FlashMLA only supports page size = 64 case.*
+
+## User guide
+
+#### Launch command for different attention backend.
+
+- FlashInfer (Default)
+```bash
+python3 -m sglang.launch_server --model meta-llama/Meta-Llama-3.1-8B-Instruct
+```
+
+- FlashAttention 3
+```bash
+python3 -m sglang.launch_server --model meta-llama/Meta-Llama-3.1-8B-Instruct --attention-backend fa3
+```
+
+- FlashMLA
+```bash
+python -m sglang.launch_server --model-path deepseek-ai/DeepSeek-R1-Distill-Qwen-7B --enable-flashmla
+```
+
+- Triton
+```bash
+python3 -m sglang.launch_server --model meta-llama/Meta-Llama-3.1-8B-Instruct --attention-backend triton
+```
+
+- Torch Native
+```bash
+python3 -m sglang.launch_server --model meta-llama/Meta-Llama-3.1-8B-Instruct --attention-backend torch_native
+```