sglang-0.5.14-minimax-h3-0801 推荐部署参数是？

请问sglang-0.5.14-minimax-h3-0801镜像8卡的推荐docker部署参数是？

尊敬的开发者您好，参考

sglang serve \
  --model-path xx \
  --num-gpus 8 \
  --tp-size 4 \
  --ulysses-degree 2 \
  --performance-mode speed \
  --host 0.0.0.0 \
  --port 30010 \
  --model-variant fl2va

[08-04 09:29:12] Initializing distributed environment with world_size=8, device=cuda:0, timeout=3600
[08-04 09:29:22] Found nccl from library libmccl.so ← Metax 的 MCCL(NCCL 兼容实现)
...
[08-04 09:32:31] Rank 0 scheduler is dead. Please check if there are relevant logs.
[08-04 09:32:31] Exit code: -9 ← worker 被 SIGKILL
node002:148:1288 [...] MCCL WARN socketProgressOpt:
Call to recv from 10.1.32.2<32919> failed : Connection reset by peer ← MCCL bootstrap 连 10.1.32.2 被对端 reset
node002:150:1287 [...] MCCL WARN socketProgressOpt:
Call to recv from 10.1.32.2<32919> failed : Connection reset by peer
...
[主进程] data = reader.recv() → raise EOFError ← 主进程与 worker 的管道断开,整体退出

3) 错误逐行解读

• Initializing distributed environment with world_size=8:8 个 GPU worker 进程开始建组。
• Found nccl from library libmccl.so:走的是 Metax 的 MCCL 集合通信库。
• Call to recv from 10.1.32.2<32919> failed: Connection reset by peer:MCCL 做 bootstrap 时,某个 worker 试图连 10.1.32.2:32919(= bond0,这台机的数据网 IP)被对端 reset。说明 8 个 worker 之间没法建立 TCP 通信通道。
• Rank 0 scheduler is dead. Exit code: -9:worker 0 收到 SIGKILL(-9),崩溃。
• EOFError:主进程等着从 worker 收消息,管道断了,整个服务退出。
  启动参数是 docker run -d \
  --name minimax-h3-sglang \
  --privileged \
  --network=host \
  --ipc=host --shm-size=32g \
  -v /data4/model/MiniMax-H3:/model \
  -e TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9;9.0" \
  -e PYTORCH_ALLOC_CONF=expandable_segments:True \
  -e SGLANG_NSA_FLASHMLA_BACKEND_DECODE_COMPUTE_FP8=0 \
  -e SGLANG_SAIL_DSV4_USE_FLASH_MLA_SPARSE_FWD=0 \
  -e SGLANG_NSA_DUAL_STREAM=0 \
  -e ENABLE_ACEXT_W4A8_EP=1 \
  -e SGLANG_ENABLE_JIT_DEEPGEMM=1 \
  -e PYTHONUNBUFFERED=1 \
  -e MCCL_IB_DISABLE=1 \
  cr.metax-tech.com/public-init/minimax-h3/sglang-0.5.14-minimax-h3-0801:latest \
  python -m sglang.multimodal_gen.runtime.entrypoints.cli.main serve \
  --model-path /model/FL2VA \
  --num-gpus 8 \
  --ulysses-degree 8 \
  --port 30010 \
  --host 0.0.0.0 \
  --trust-remote-code

尊敬的开发者您好，请使用以下命令

export MODEL_PATH="$MINIMAX_H3_FL2VA_MODEL_PATH" 
export PORT=30010 
export SERVER_OUTPUT_DIR="$MINIMAX_H3_OUTPUT_ROOT/server-$PORT" 
mkdir -p "$SERVER_OUTPUT_DIR" 

docker run -it  --device=/dev/dri --device=/dev/mxcd \
--device=/dev/infiniband --privileged=true --group-add video \
--device=/dev/mem \
--network=host --security-opt seccomp=unconfined \
--security-opt apparmor=unconfined --shm-size '100gb' \
--ulimit memlock=-1 -v /usr/local/:/usr/local/ \
$image_id /bin/bash

python3 -m sglang.multimodal_gen.runtime.launch_server \
  --model-path "$MODEL_PATH" \
  --num-gpus 8 \
  --sp-degree 8 \
  --ulysses-degree 8 \
  --use-fsdp-inference true \
  --hsdp-replicate-dim 1 \
  --hsdp-shard-dim 8 \
  --dit-cpu-offload true \
  --dit-layerwise-offload false \
  --vae-cpu-offload true \
  --trust-remote-code \
  --warmup-mode off \
  --host 0.0.0.0 \
  --port "$PORT" \
  --strict-ports \
  --output-path "$SERVER_OUTPUT_DIR"

请问在8卡C500环境下，有运行更快的参数吗？

另一个问题是，每次运行一次生成视频后，第二次发送请求sglang进程就会crash，然后就要重启sglang，请问是我的命令有问题吗？
video_id=$(
curl -sS -X POST http://127.0.0.1:30000/v1/videos \
-H "Content-Type: application/json" \
-d '{
"model": "MiniMaxAI/MiniMax-H3",
"prompt": "A young man walks on the street.",
"seconds": 5,
"task": "t2va",
"conditions": [],
"target": {
"short_edge": 768, "aspect_ratio": "9:16", "duration_seconds": 5.0
},
"num_outputs_per_prompt": 1, "num_inference_steps": 50,
"flow_shift": 12.0, "audio_flow_shift": 3.0, "seed": 1101
}' |
jq -r '.id'
)
while true; do
status=$(curl -sS "http://127.0.0.1:30000/v1/videos/${video_id}" | jq -r '.status')
[ "$status" = "completed" ] && break
[ "$status" = "failed" ] && exit 1
sleep 10
done
curl -sS -L "http://127.0.0.1:30000/v1/videos/${video_id}/content" -o minimax-h3-t2va.mp4

尊敬的开发者您好，请参考developer.metax-tech.com/forum/t/fa-tie-qian-bi-kan-jing-xiang-shi-yong-wen-ti-ti-wen-mo-ban/267/，详细描述您的环境信息和启动日志

启动方式就是参考docker run -it --device=/dev/dri --device=/dev/mxcd \
--device=/dev/infiniband --privileged=true --group-add video \
--device=/dev/mem \
--network=host --security-opt seccomp=unconfined \
--security-opt apparmor=unconfined --shm-size '100gb' \
--ulimit memlock=-1 -v /usr/local/:/usr/local/ \
$image_id /bin/bash

python3 -m sglang.multimodal_gen.runtime.launch_server \
--model-path "$MODEL_PATH" \
--num-gpus 8 \
--sp-degree 8 \
--ulysses-degree 8 \
--use-fsdp-inference true \
--hsdp-replicate-dim 1 \
--hsdp-shard-dim 8 \
--dit-cpu-offload true \
--dit-layerwise-offload false \
--vae-cpu-offload true \
--trust-remote-code \
--warmup-mode off \
--host 0.0.0.0 \
--port "$PORT" \
--strict-ports \
--output-path "$SERVER_OUTPUT_DIR"

环境是8卡C500, 驱动sdk都是最新的3.8.1
第一次请求正常，但是第二次请求就会出现

尊敬的开发者您好，等待后续镜像修复此问题。若需加速修复，请通过商务渠道申请。

我也是报这个错，环境： 128核 768G内存 8*C500