MiMo-V2.5-DFlash-MXFP4A16

Benchmarks

TerminalBench 2.1

• 64%

GSM8K

Docker Image

cstechdev/vllm:eldritch-enlightenment-mimo25-dflash-mxfp4-v2-cu132-20260705

Tested on 4x RTX 6000 (Blackwell SM120)

docker run --rm \
  --gpus all \
  --ipc=host \
  --shm-size=32g \
  --ulimit memlock=-1 \
  --ulimit stack=67108864 \
  --ulimit nofile=1048576:1048576 \
  --network host \
  --init \
  -e CUDA_VISIBLE_DEVICES=0,1,2,3 \
  -e CUDA_DEVICE_ORDER=PCI_BUS_ID \
  -e CUTE_DSL_ARCH=sm_120a \
  -e NCCL_IB_DISABLE=1 \
  -e NCCL_P2P_LEVEL=SYS \
  -e NCCL_PROTO=LL,LL128,Simple \
  -e PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
  -e VLLM_ENABLE_PCIE_ALLREDUCE=1 \
  -e VLLM_PCIE_ALLREDUCE_BACKEND=b12x \
  -e VLLM_PCIE_ONESHOT_ALLREDUCE_MAX_SIZE=64KB \
  -e SAFETENSORS_FAST_GPU=1 \
  -e VLLM_USE_V2_MODEL_RUNNER=1 \
  -e VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1 \
  -e VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS=1 \
  -v /path/to/MiMo-V2.5-DFlash-MXFP4A16:/model:ro \
  cstechdev/vllm:eldritch-enlightenment-mimo25-dflash-mxfp4-v1-cu132-20260705 \
  /bin/sh -lc 'exec vllm serve /model \
    --served-model-name mimo-v2.5 \
    --trust-remote-code \
    --host 0.0.0.0 \
    --port 8000 \
    --kv-cache-dtype auto \
    --block-size 64 \
    --tensor-parallel-size 4 \
    --gpu-memory-utilization 0.90 \
    --max-num-seqs 64 \
    --max-num-batched-tokens 16384 \
    --max-model-len 500000 \
    --max-cudagraph-capture-size 128 \
    --attention-backend TRITON_ATTN \
    --kernel-config.moe_backend flashinfer_cutlass \
    --kernel-config.linear_backend b12x \
    --reasoning-parser mimo \
    --tool-call-parser mimo \
    --enable-auto-tool-choice \
    --compilation-config "{\"cudagraph_mode\":\"FULL_AND_PIECEWISE\",\"custom_ops\":[\"all\"]}" \
    --async-scheduling \
    --no-scheduler-reserve-full-isl \
    --enable-chunked-prefill \
    --enable-prefix-caching \
    --speculative-config "{\"model\":\"/model/dflash\",\"method\":\"dflash\",\"num_speculative_tokens\":7}"'