# DeepSeek-V4.1-Flash on 4x DGX Spark with Aiden's SGLang production-1.6 # (aidenle.com/recipes/deepseek-v4-1-flash-sglang-4x-dgx-spark). # # Image: aidendle94/sparkrun-sglang-dsv41-gb10:production-1.6 (Docker Hub, # arm64, 33.3 GB) = lmsysorg/sglang:dev-dsv41 (SGLang team's own DSV4.1 # image, unmodified) + Aiden's MIT overlay (import hooks: Engram row store # from node-local/shard files inside the CUDA graph, b12x SM12x FP8 dense # kernels, SM120 prefill paging, allocator flush, one-shot RDMA collectives, # request guard). Lineage: replaced the vLLM stack as Aiden's production # 2026-09-17 (prefill +30-45%, KV pool +23%); 1.2: 16 concurrent; 1.5: rank # crash fix + 2 GB10 speedups; 1.6 current. Measured by Aiden on 4x: decode # 107/87/36 tok/s (count/code/prose C1), prefill 3.3K tok/s @32K, KV pool # ~4.2M, 524K context, ~10 min to healthy. # # Model: deepseek-ai/DeepSeek-V4.1-Flash (510 GB, fp8 native) on the NAS — # serve with --hf-cache /mnt/nas/ai/huggingface. The checkpoint snapshot # dba1be0a... is also the Engram row source (SPARK_ENGRAM_DIR): all four # ranks read their rows directly from the NAS shards (Aiden's fallback path; # his measured config uses 47.2 GiB node-local sparse copies, which needs the # private engram_local.py — the NFS read costs ~3-5 ms/step extra, acceptable # for a validation run). # # Per-rank Engram row ranges (the table has 384,006,168 rows per layer, # layers 1 and 14; four contiguous ranges partition each table): # rank 0: 0..96001542 rank 1: 96001542..192003084 # rank 2: 192003084..288004626 rank 3: 288004626..384006168 # Computed per node by /home/kilork/aiden4x-launch.sh from $NODE_RANK # (the launcher exports NODE_RANK per rank; a full range on every rank would # double-count row contributions in the reduce, so the range is mandatory — # the overlay refuses a full snapshot without it). # # B12X_ROCE_GID_INDEX reuses the launcher's per-node NCCL GID auto-detect # (the wrapper exports it), so no pinned GID anywhere. engine = "sglang" name = "DeepSeek-V4.1-Flash (SGLang 4x, Aiden production-1.6)" docker_image = "aidendle94/sparkrun-sglang-dsv41-gb10:production-1.6" model = "deepseek-ai/DeepSeek-V4.1-Flash" parent_model = "deepseek-ai/DeepSeek-V4.1-Flash" port = 30000 required_nodes = 4 tensor_parallel_size = 4 distributed_backend = "mp" shm_size = "32gb" # The wrapper (mounted read-only from the host via extra_mounts below) sets # SPARK_ENGRAM_ROWS (+ B12X_ROCE_GID_INDEX) from NODE_RANK and execs the # engine with the launcher's full flag list in "$@". entrypoint = ["sh", "/aiden4x-launch.sh"] args = [ "--trust-remote-code", "--served-model-name", "deepseek-v4.1-flash", "--load-format", "safetensors", # Aiden's measured shape (production-1.6); --tp 4 / --nnodes 4 / # --node-rank / --dist-init-addr are appended by the launcher "--ep-size", "4", "--attention-backend", "dsv4", "--moe-runner-backend", "flashinfer_mxfp4", # required for the MXFP8 route (selects the block-scaled path the hook wraps; # K=576 projections fall back to CUTLASS per shape) "--fp8-gemm-backend", "flashinfer_cutlass", "--mem-fraction-static", "0.80", "--chunked-prefill-size", "2048", "--context-length", "524288", "--max-running-requests", "16", "--cuda-graph-max-bs-decode", "16", "--random-seed", "0", "--enable-decoder-swa-bounded-replay", "--speculative-algorithm", "DSPARK", "--speculative-dspark-block-size", "5", "--tool-call-parser", "deepseekv41", "--reasoning-parser", "deepseek-v41", "--watchdog-timeout", "1800", "--min-free-slots-delay", "1", ] # The launcher also appends: --model-path , --nnodes 4, # --node-rank N, --dist-init-addr :25000, --host 0.0.0.0, --port 30000. # NOTE: no --default-chat-template-kwargs '{"thinking":true}' — the bench # harness controls thinking; Aiden's default is for his interactive fleet. # The per-rank Engram range wrapper, mounted read-only (bare keys above; the # table header must come after them or `args`/`[env]` would parse into it). [[extra_mounts]] host = "/home/kilork/aiden4x-launch.sh" container = "/aiden4x-launch.sh" read_only = true # Engram row source: the full checkpoint snapshot on the NAS (container-side # path — the launcher mounts --hf-cache /mnt/nas/ai/huggingface at # /root/.cache/huggingface). No engram-local.json manifest in a full snapshot, # so per-rank ranges come from SPARK_ENGRAM_ROWS (wrapper). [env] SPARK_ENGRAM_DIR = "/root/.cache/huggingface/hub/models--deepseek-ai--DeepSeek-V4.1-Flash/snapshots/dba1be0a40aa45a94ad051997016db3960a90277" SPARK_ENGRAM_MODE = "hostnode" SPARK_ENGRAM_EARLY = "1" SPARK_ENGRAM_PREFETCH = "1" SPARK_PAGE_CACHE_RELEASE = "1" SPARK_LATE_TAIL_SKIP = "1" SPARK_LATE_TAIL_SKIP_ALL = "1" SPARK_SM120_REAL_HEADS = "1" SPARK_ENGRAM_THREADS = "64" SPARK_ENGRAM_MAX_IDS = "65536" SPARK_MXFP8_BACKEND = "b12x" SPARK_ROCE_AR = "1" SPARK_ROCE_AR_MAX = "1MB" SPARK_ROCE_AG_MAX = "16MB" # b12x RoCE runtime (GID index set by the wrapper from the detected value) B12X_ROCE_HCA = "rocep1s0f0,roceP2p1s0f0" B12X_ROCE_SPIN_LIMIT = "300000000" # image defaults Aiden overrides SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE = "0" SGLANG_FLASHINFER_MOE_FUSED_FINALIZE = "0" SGLANG_DSV41_REASONING_EFFORT = "100" PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:False" CUDA_DEVICE_ORDER = "PCI_BUS_ID" # Aiden's NCCL pins (4x 200G RoCE fabric, 192.168.200.0/24) — [env] wins over # the config's [nccl."4nodes-switch"] where they differ NCCL_NET = "IB" NCCL_IB_DISABLE = "0" NCCL_IB_HCA = "rocep1s0f0,roceP2p1s0f0" NCCL_SOCKET_IFNAME = "enp1s0f0np0" GLOO_SOCKET_IFNAME = "enp1s0f0np0" NCCL_IB_ROCE_VERSION_NUM = "2" NCCL_IB_ADDR_FAMILY = "AF_INET" NCCL_IB_ADDR_RANGE = "192.168.200.0/24" NCCL_IB_SUBNET_PREFIX_LEN = "24" NCCL_CROSS_NIC = "1" NCCL_IB_MERGE_NICS = "0" NCCL_IB_SUBNET_AWARE_ROUTING = "1" NCCL_CUMEM_ENABLE = "0" NCCL_P2P_DISABLE = "1" NCCL_SHM_DISABLE = "1" NCCL_BUFFSIZE = "1048576" NCCL_LL128_BUFFSIZE = "262144" NCCL_PROTO = "^LL128" NCCL_MAX_NCHANNELS = "8"