Update README.md to include SGLang cookbook

README.md

@@ -81,6 +81,7 @@ Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated sys
 **Supported Runtime Engine(s):**   
 
 - vLLM
+- SGLang
 
 **Supported Hardware Microarchitecture Compatibility:**   
 
@@ -136,7 +137,7 @@ No calibration dataset was used. The MXFP4-to-NVFP4 expert conversion used `inpu
 
 ## Inference:
 
-**Acceleration Engine:** vLLM   
+**Acceleration Engine:** vLLM, SGLang   
 
 **Test Hardware:** 8 NVIDIA Blackwell B300 GPUs
 
@@ -146,10 +147,12 @@ No calibration dataset was used. The MXFP4-to-NVFP4 expert conversion used `inpu
 
 ## Post Training Quantization
 
-This model was obtained from Kimi-K3 without calibration data. The source MXFP4 routed-expert weights were converted to NVFP4 using `input_scale=1.0`, while the supported attention projection weights in KDA and MLA were quantized to 128×128 per-block FP8. Other checkpoint tensors retain their original precision. The checkpoint is ready for inference with vLLM.
+This model was obtained from Kimi-K3 without calibration data. The source MXFP4 routed-expert weights were converted to NVFP4 using `input_scale=1.0`, while the supported attention projection weights in KDA and MLA were quantized to 128×128 per-block FP8. Other checkpoint tensors retain their original precision. The checkpoint is ready for inference with vLLM and SGLang.
 
 ## Usage
 
+### vLLM
+
 This checkpoint was validated on 8 NVIDIA B300 GPUs with the public [`vllm/vllm-openai:kimi-k3`](https://hub.docker.com/r/vllm/vllm-openai/tags?name=kimi-k3) image. At validation time, the image included vLLM `0.1.dev19262+gb6bbf29dd.d20260727`. The compatibility patch included in this repository is required for this serving configuration.
 
 Native vLLM support is being upstreamed to replace the compatibility patch. Generic mixed `FP8_PB_WO` dispatch and DeepGEMM preparation are tracked in [vLLM PR #50617](https://github.com/vllm-project/vllm/pull/50617), Kimi-K3 fused-projection support in [vLLM PR #52406](https://github.com/vllm-project/vllm/pull/52406), and the TRTLLM NVFP4 SiTU scale correction in [vLLM PR #52405](https://github.com/vllm-project/vllm/pull/52405). Continue using the bundled patch until those changes are included in a released vLLM build.
@@ -200,6 +203,42 @@ The first launch downloads approximately 1.6 TB into the mounted Hugging Face ca
 
 The 196,608-token limit is the validated TP8 setting rather than the model's architectural maximum. Adjust context length and concurrency together based on available memory.
 
+### SGLang
+
+This checkpoint serves on 8 NVIDIA B300 GPUs with [`lmsysorg/sglang:dev-dev-kimi-k3-nvfp4`](https://hub.docker.com/r/lmsysorg/sglang/tags?name=dev-dev-kimi-k3-nvfp4), a CUDA 13 build cut from the head of the SGLang support PR. No runtime patch and no registry authentication are required.
+
+Native SGLang support for the mixed NVFP4 / `FP8_PB_WO` checkpoint is tracked in [SGLang PR #35077](https://github.com/sgl-project/sglang/pull/35077). Until that lands in a released build, use the image above — a pip-installed SGLang cannot load this checkpoint.
+
+`--moe-runner-backend flashinfer_trtllm` is required rather than optional: the automatic resolution never engages the TRT-LLM deferred-finalize path, and `flashinfer_cutlass` has no SiTU kernel for the routed experts. The quantization scheme is read from the checkpoint's own config, so no `--quantization` flag is needed.
+
+```sh
+docker run --rm --gpus all --ipc=host --network=host \
+  --shm-size=64g \
+  -e HF_TOKEN \
+  -v "$HOME/.cache/huggingface:/root/.cache/huggingface" \
+  lmsysorg/sglang:dev-dev-kimi-k3-nvfp4 \
+  sglang serve \
+    --trust-remote-code \
+    --model-path nvidia/Kimi-K3-NVFP4 \
+    --served-model-name kimi-k3-nvfp4 \
+    --tp-size 8 \
+    --dcp-size 8 \
+    --mem-fraction-static 0.85 \
+    --reasoning-parser kimi_k3 \
+    --tool-call-parser kimi_k3 \
+    --host 0.0.0.0 \
+    --port 30000 \
+    --moe-runner-backend flashinfer_trtllm \
+    --speculative-algorithm DSPARK \
+    --speculative-draft-model-path RadixArk/Kimi-K3-DSpark \
+    --speculative-dspark-block-size 7 \
+    --enable-linear-replayssm-spec
+```
+
+The three `--speculative-*` flags turn on DSPARK speculative decoding, the cookbook's default operating point for this model; drop them to serve without speculation. Context length is left at the checkpoint's own value — narrow it with `--context-length`, and tune it together with concurrency against available memory.
+
+Further operating points — the low-latency and high-throughput tiers, GB200/GB300 multi-node recipes, and prefill/decode disaggregation — are generated by the [SGLang Kimi-K3 cookbook](https://docs.sglang.io/cookbook/autoregressive/Moonshotai/Kimi-K3), where NVFP4 is the **Quantization** row of the deployment panel.
+
 
 
 ## Evaluation