jdopensource/JoyAI-LLM-Flash

1# JoyAI-LLM Flash Deployment Guide2 3> [!Note]4> This guide offers a selection of deployment command examples for JoyAI-LLM Flash, which may not be the optimal configuration. Given the rapid evolution of inference engines, we recommend referring to their official documentation for the latest updates to ensure peak performance.5 6> Support for JoyAI-LLM Flash’s dense MTP architecture is currently being integrated into vLLM and SGLang. Until these PRs are merged into a stable release, please use the nightly Docker image for access to these features.7 8## vLLM Deployment9 10Here is the example to serve this model on a H200 single node via vLLM:11 121. pull the Docker image.13```bash14docker pull jdopensource/joyai-llm-vllm:v0.15.1-joyai_llm_flash15```162. launch JoyAI-LLM Flash model with dense MTP.17```bash18# TP1 for memory efficiency19vllm serve ${MODEL_PATH} -tp 1 --trust-remote-code \20   --tool-call-parser qwen3_coder --enable-auto-tool-choice \21   --speculative-config $'{"method": "mtp", "num_speculative_tokens": 3}'22 23# TP8 for extreme speed and long context24vllm serve ${MODEL_PATH} -tp 8 --trust-remote-code \25  --tool-call-parser qwen3_coder --enable-auto-tool-choice \26  --speculative-config $'{"method": "mtp", "num_speculative_tokens": 3}'27```28**Key notes**29- `--tool-call-parser qwen3_coder`: Required for enabling tool calling30 31## SGLang Deployment32 33Similarly, here is the example to run on a H200 single node via SGLang:34 351. pull the Docker image.36```bash37docker pull jdopensource/joyai-llm-sglang:v0.5.8-joyai_llm_flash38```392. launch JoyAI-LLM Flash model with dense MTP.40 41```bash42# TP1 for memory efficiency43python3 -m sglang.launch_server --model-path ${MODEL_PATH} --tp-size 1 --trust-remote-code \44  --tool-call-parser qwen3_coder \45  --speculative-algorithm EAGLE --speculative-draft-model-path ${MTP_MODEL_PATH} \46  --speculative-num-steps 2 --speculative-eagle-topk 2 --speculative-num-draft-tokens 347 48# TP8 for extreme speed and long context49python3 -m sglang.launch_server --model-path ${MODEL_PATH} --tp-size 8 --trust-remote-code \50  --tool-call-parser qwen3_coder \51  --speculative-algorithm EAGLE \52  --speculative-num-steps 2 --speculative-eagle-topk 2 --speculative-num-draft-tokens 353```54**Key notes:**55- `--tool-call-parser qwen3_coder`: Required when enabling tool usage.56