Add SGLang serving instructions

Adds SGLang as a supported runtime and documents the high-throughput aggregated TP4 serving command from the GLM-5.3-Flash cookbook.

Validation:
- SGLang 0.0.0.dev1+g008403017 on 4x NVIDIA GB300 GPUs
- High-throughput: full GSM8K test split, 1319/1319 examples, 97.19% accuracy, 0 request errors
- Low-latency adaptive EAGLE: full GSM8K test split, 1319/1319 examples, 96.66% accuracy, 0 request errors

This PR intentionally documents the high-throughput command. Addresses discussion #2.

README.md

@@ -61,6 +61,7 @@ Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated sys
 ## Software Integration:
 **Supported Runtime Engine(s):** <br>
 * vLLM <br>
+* SGLang <br>
 
 **Supported Hardware Microarchitecture Compatibility:** <br>
 * NVIDIA Blackwell <br>
@@ -95,7 +96,7 @@ The model version is NVFP4 1.0 version and is quantized with nvidia-modelopt **v
 
 
 ## Inference:
-**Acceleration Engine:** vLLM <br>
+**Acceleration Engine:** vLLM, SGLang <br>
 **Test Hardware:** NVIDIA Blackwell GB200 <br>
 
 ## Post Training Quantization
@@ -127,6 +128,26 @@ vllm serve /checkpoint \
   --gpu-memory-utilization 0.90
 ```
 
+### SGLang
+
+To serve this checkpoint with [SGLang](https://github.com/sgl-project/sglang) on 4 NVIDIA Blackwell GPUs, run:
+
+```sh
+sglang serve \
+  --model-path nvidia/GLM-5.3-Flash-NVFP4 \
+  --quantization modelopt_fp4 \
+  --tp-size 4 \
+  --dsa-prefill-backend trtllm \
+  --dsa-decode-backend trtllm \
+  --kv-cache-dtype fp8_e4m3 \
+  --moe-runner-backend flashinfer_cutlass \
+  --reasoning-parser glm45 \
+  --tool-call-parser glm47 \
+  --mem-fraction-static 0.85 \
+  --host 0.0.0.0 \
+  --port 30000
+```
+
 ## Evaluation
 The accuracy benchmark results are presented in the table below:
 <table>