Skip to content

vLLM Minimax 2.7 startart error #9

@shivajid

Description

@shivajid

While trying to bring up Minimax 2.7 on single node 8 GPU instance of B40, the process is hanging. This minimax 2.5 earlier worked in the same env,

(APIServer pid=1) DEBUG 04-12 21:38:28 [v1/engine/utils.py:1143] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=1) DEBUG 04-12 21:38:38 [v1/engine/utils.py:1143] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=1) DEBUG 04-12 21:38:48 [v1/engine/utils.py:1143] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=1) DEBUG 04-12 21:38:58 [v1/engine/utils.py:1143] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=1) DEBUG 04-12 21:39:08 [v1/engine/utils.py:1143] Waiting for 1 local, 0 remote core engine proc(s) to start.
(APIServer pid=1) DEBUG 04-12 21:39:18 [v1/engine/utils.py:1143] Waiting for 1 local, 0 remote core engine proc(s) to start.

This is from kubernetes yaml.

 containers:
      - name: vllm-container
        image: vllm/vllm-openai:nightly
        env:
        - name: VLLM_LOGGING_LEVEL
          value: "DEBUG" # Shows detailed model loading progress
        - name: NCCL_DEBUG
          value: "INFO"  # "TRACE" is very noisy, "INFO" is usually enough
        - name: VLLM_WORKER_MULTIPROC_METHOD
          value: "fork"
        args: [
          "MiniMaxAI/MiniMax-M2.7",
          "--tensor-parallel-size", "8",
          "--enable-expert-parallel",
          "--tool-call-parser", "minimax_m2",
          "--reasoning-parser", "minimax_m2_append_think",
          "--enable-auto-tool-choice",
          "--compilation-config", '{"mode":1,"pass_config":{"fuse_minimax_qk_norm":true}}' ,
          "--trust-remote-code",
          "--compilation-config", '{"cudagraph_mode": "PIECEWISE"}',
          "--disable-custom-all-reduce",
          "--enforce-eager"
        ]

I have tried the instructions here - https://docs.vllm.ai/projects/recipes/en/latest/MiniMax/MiniMax-M2.html#using-docker
All of them seems to hang in the above steps. Any hints.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions