[Bug] FlashInfer sparse MLA attention fails with "Unsupported sparse-MLA prefill configuration" during DSpark CUDA Graph capture

[BrightMoon-000]

using the alternative dspark-mtp0 mode (by setting MODE=dspark-mtp0) allows the service to start successfully. The failure only reproduces with the default MODE=dspark setting.

---

Environment

• Hardware: 8 × NVIDIA RTX 5090 (32 GB VRAM each)
• Software:
  • Docker image: voipmonitor/vllm:infernal-invocation-vllmd6cf36a-b12xf6dc512-fi1ac6942-cu133-torch213-20260827-r21

---

Reproduction (Docker run command)

docker run -d \
  --name deepseek-v4-flash \
  --gpus all \
  --runtime nvidia \
  --network host \
  --ipc host \
  --shm-size 48g \
  --init \
  --restart unless-stopped \
  --ulimit memlock=-1 \
  --ulimit stack=67108864 \
  --ulimit nofile=1048576:1048576 \
  -v /fast/models/DeepSeek-V4-Flash-0731:/app/model:ro \
  -v /fast/cache/vllm:/cache:rw \
  -v /fast/cache/tmp:/container-tmp:rw \
  -e CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
  -e OMP_NUM_THREADS=2 \
  -e PORT=8000 \
  -e SERVED_MODEL_NAME=DeepSeek-V4-Flash \
  -e MODEL_PATH=/app/model \
  -e MODE=dspark \
  -e BACKEND=b12x-a8-dglin \
  -e TP_SIZE=8 \
  -e DCP_SIZE=1 \
  -e ALLREDUCE_MODE=auto \
  -e DSPARK_DEPTH_MODE=fixed \
  -e DSPARK_TOKENS=5 \
  -e DRAFT_SAMPLE_METHOD=probabilistic \
  -e MAX_NUM_SEQS=16 \
  -e GRAPH=auto \
  -e MAX_MODEL_LEN=262144 \
  -e MAX_NUM_BATCHED_TOKENS=4096 \
  -e GPU_MEMORY_UTILIZATION=0.95 \
  -e LMCACHE_MODE=ram \
  -e LMCACHE_L1_GB=32 \
  -e LMCACHE_HTTP_PORT=8099 \
  -e LOAD_FORMAT=instanttensor \
  -e INSTANTTENSOR_BACKEND=BUFFERED \
  -e PYTHONHASHSEED=0 \
  --entrypoint /usr/local/bin/lmcache-mp-wrapper.sh \
  voipmonitor/vllm:infernal-invocation-vllmd6cf36a-b12xf6dc512-fi1ac6942-cu133-torch213-20260827-r21 \
  /usr/local/bin/serve-ds4-flash.sh

The container script internally invokes vllm serve ... --attention-backend B12X_MLA_SPARSE ... --speculative-config ... --method dspark.

---

Error log (core excerpt)

(Worker_TP6 pid=1996) ERROR 08-28 16:17:44 [multiproc_executor.py:1048] WorkerProc hit an exception.
...
tvm.error.InternalError: Check failed: (ok) is false: Unsupported sparse-MLA prefill configuration: model=DSV4 num_heads=8 topk=512 page_block_size=64 topk_extra=0 extra_page_block_size=0
...
RuntimeError: Worker failed with error 'Check failed: (ok) is false: Unsupported sparse-MLA prefill configuration: model=DSV4 num_heads=8 topk=512 page_block_size=64 topk_extra=0 extra_page_block_size=0'

Full stack trace points to the CUDA graph capture phase inside self.speculator.capture() calling flashinfer.sparse_mla_sm120_paged_attention.

[BrightMoon-000]

This issue can be resolved by using the latest version of flashinfer v0.6.18rc10.