Recommend public vLLM image for Kimi-K3

Signed-off-by: Zhiyu Cheng <zhiyuc@nvidia.com>

README.md

@@ -150,13 +150,11 @@ This model was obtained from Kimi-K3 without calibration data. The source MXFP4
 
 ## Usage
 
-This checkpoint was validated on 8 NVIDIA B300 GPUs with [`nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1`](https://catalog.ngc.nvidia.com/orgs/nvidia/ai-dynamo/containers/vllm-runtime/1.4.0-kimi-k3-dev.1), which includes vLLM `0.1.dev19251+g13c59a3da.d20260726`. The image requires NGC authentication. The compatibility patch included in this repository is required for this serving configuration.
+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.
 
 ```sh
-docker login nvcr.io
-
 hf download nvidia/Kimi-K3-NVFP4 \
   runtime_patches/sitecustomize.py \
   --local-dir kimi-k3-runtime
@@ -173,7 +171,7 @@ docker run --rm --gpus all --ipc=host --network=host \
   -e VLLM_USE_RUST_FRONTEND=0 \
   -v "$HOME/.cache/huggingface:/root/.cache/huggingface" \
   -v "$PWD/kimi-k3-runtime/runtime_patches:/k3-runtime-patches:ro" \
-  nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1 \
+  vllm/vllm-openai:kimi-k3 \
   vllm serve nvidia/Kimi-K3-NVFP4 \
     --served-model-name kimi-k3-nvfp4 \
     --host 0.0.0.0 \
@@ -182,10 +180,16 @@ docker run --rm --gpus all --ipc=host --network=host \
     --quantization modelopt_mixed \
     --tensor-parallel-size 8 \
     --moe-backend flashinfer_trtllm \
+    --gpu-memory-utilization 0.90 \
     --max-model-len 196608 \
+    --max-num-batched-tokens 8192 \
+    --max-num-seqs 32 \
     --kv-cache-dtype fp8 \
     --attention-config '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}' \
     --attention-backend FLASHINFER_MLA \
+    --enable-prefix-caching \
+    --enable-chunked-prefill \
+    --max-cudagraph-capture-size 32 \
     --enable-auto-tool-choice \
     --tool-call-parser kimi_k3 \
     --reasoning-parser kimi_k3