# DeepSeek-V4.1-Flash — eugr b12x nightly-20260913 (lab `dev/infernal-invocation`), 4x. # # P1.2 of the 2026-09-18 night: the r37 typo A/B. r37 (09-13, vllmc687594b8a8) # garbled DS4.1 text in the morning bisect (A/B/C: flags excluded; D infeasible # — engram-off 4x OOMs, the table is part of the weights and only fits in DRAM # at 8x). This runs the lab's NEWER published b12x build (09-13 DockerHub, cut # from dev/infernal-invocation — newer than r37's 3fb88fdf tree) with the SAME # measured r37 contract (GMU 0.88, 512K, DCP1, block 256, B12X attn/linear/moe, # K7 greedy/adaptive, engram disk, instanttensor/lazy). Garbled -> the r37 # garbling is a lab-tree trait (not our build); clean -> r37-specific # regression (file the exact bisect with the lab). # # Image already pulled on dgx1-4 (24.5 GB). Same model/weights/engram staging # as the r37 recipe (NAS snapshot + per-node /var/tmp/engram-local hardlinks). engine = "vllm" name = "DeepSeek-V4.1-Flash (eugr b12x 0913 = lab dev/infernal-invocation, 4x DCP1, DSpark K7)" docker_image = "eugr/spark-vllm-b12x:nightly-20260913" # prebuilt (no docker_image_dir) model = "deepseek-ai/DeepSeek-V4.1-Flash" parent_model = "deepseek-ai/DeepSeek-V4.1-Flash" required_nodes = 4 tensor_parallel_size = 4 pipeline_parallel_size = 1 distributed_backend = "mp" shm_size = "32gb" # Our r27 build dir images ship with ENTRYPOINT [] — set it explicitly. entrypoint = ["vllm", "serve"] args = [ "--served-model-name", "dsv41", "--trust-remote-code", "--distributed-executor-backend", "mp", # GB10 startup-gate value (see header); KV pool is engine-computed from # the remainder (no pin — their profile pins nothing either). "--gpu-memory-utilization", "0.88", # 512K context (our floor — their qualified 131K is not usable for us; # the KV pool is engine-computed from the memory budget, unchanged): "--max-model-len", "524288", "--decode-context-parallel-size", "1", "--max-num-seqs", "4", "--max-num-batched-tokens", "4096", "--prefill-compute-share", "0.4", "--block-size", "256", "--enable-chunked-prefill", "--enable-prefix-caching", "--kv-cache-dtype", "fp8", "--attention-backend", "B12X", "--linear-backend", "b12x", "--moe-backend", "b12x", "--compilation-config", """ '{"cudagraph_mode":"FULL_AND_PIECEWISE"}'""", "--max_cudagraph_capture_size", "128", # Native DSpark K7 — the draft weights live inside the target checkpoint # (no separate draft repo). Greedy draft, standard rejection, adaptive # verification (the lab's r36 qualified trio vs our prod's K5 # probabilistic/block). "--speculative-config", """ '{"method":"dspark","num_speculative_tokens":7,"draft_tensor_parallel_size":4,"attention_backend":"B12X","draft_sample_method":"greedy","rejection_sample_method":"standard","enable_adaptive_verification":true}'""", # Engram: the model's 196B learned ngram lookup storage. disk mode = # io_uring reads of the table rows from the checkpoint dir (bounded # staging; NOT pinned RAM). cpu_offload stays false (no generic CPU # model offload). "--engram-config", """ '{"cpu_offload":false,"table_memory":"disk"}'""", "--load-format", "instanttensor", "--safetensors-load-strategy", "lazy", "--generation-config", "vllm", "--override-generation-config", """ '{"temperature":1.0,"top_p":0.95}'""", "--limit-mm-per-prompt", """ '{"image":2}'""", "--reasoning-parser", "deepseek_v41", "--tool-call-parser", "deepseek_v41", "--enable-auto-tool-choice", "--jit-monitor-mode", "error", ] [env] # Their serve-ds41-jovian.sh runtime env (multi-node-adapted; the launcher # injects NCCL_SOCKET_IFNAME/IB_HCA/GID + VLLM_HOST_IP per node). # 2026-09-14 (night shift, operator's typo report): the lab tree's runtime # LM-head quantization (vocab_parallel_embedding.py) quantizes an unquantized # LM head to MXFP8 BY DEFAULT (envs.py: VLLM_MXFP8_LM_HEAD, default "1"). # The eugr 2026-09-10 nightly (our prod image) has NO such feature and keeps # the BF16 head - the coarsened 8-bit vocab-logit layer is the suspected cause # of the typo-rate regression (close-token logits distorted). Opting out. VLLM_MXFP8_LM_HEAD = "0" NCCL_NCHANNELS = "16" NCCL_MIN_NCHANNELS = "16" NCCL_MAX_NCHANNELS = "16" NCCL_BUFFSIZE = "2097152" OMP_NUM_THREADS = "8" CUDA_DEVICE_MAX_CONNECTIONS = "32" # GB10 arch (theirs sm_120a on RTX PRO 6000): CUTE_DSL_ARCH = "sm_121a" TORCH_CUDA_ARCH_LIST = "12.1a" # Their V2 model-runner + AOT compile pipeline (baked in their image env): VLLM_USE_V2_MODEL_RUNNER = "1" VLLM_USE_AOT_COMPILE = "1" VLLM_USE_STANDALONE_COMPILE = "1" VLLM_USE_MEGA_AOT_ARTIFACT = "1" VLLM_USE_BREAKABLE_CUDAGRAPH = "0" VLLM_USE_FLASHINFER_SAMPLER = "1" VLLM_WORKER_MULTIPROC_METHOD = "spawn" VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS = "900" HF_HUB_OFFLINE = "1" # Deliberately NOT set: VLLM_ENABLE_PCIE_ALLREDUCE (single-host PCIe in # their profile; cross-node on the switch fabric uses the launcher's NCCL), # VLLM_HOST_IP (launcher per-node), NCCL_SOCKET_IFNAME (launcher per cluster). # --- Engram from node-local NVMe, the rest from the shared NAS --- # [composed_model] (standard launcher feature) builds a per-node /tmp dir: # the whole DSV4.1 snapshot symlinked into the HF cache + the two 101.5 GB # Engram table shards (model-00047/00048) HARDLINKED from the pre-existing # per-rank sparse copies in /var/tmp/engram-local (48 GB real per node, built # for the eugr/aiden lanes — offset-preserving full-size files, only this # rank's rows materialized; hardlink = zero extra disk, same fs as /tmp). # The r36 io_uring PLE reader then reads this rank's table rows from local # NVMe (no NFS); weights stream from the NAS at load time (one-shot, same as # prod). Local files run after the repo symlinks and win at the same name. [composed_model] mount = "/model" [[composed_model.sources]] repo = "deepseek-ai/DeepSeek-V4.1-Flash" # whole snapshot, symlinked into the HF cache [[composed_model.local_files]] # hardlinks win over the symlinks above host = "/var/tmp/engram-local/DeepSeek-V4.1-Flash/model-00047-of-00048.safetensors" [[composed_model.local_files]] host = "/var/tmp/engram-local/DeepSeek-V4.1-Flash/model-00048-of-00048.safetensors"