launcher/recipes/dsv41-flash-jovian-r37-4x/recipe.toml

# DeepSeek-V4.1-Flash — local-inference-lab r37 (jovian-judgement, 2026-09-13), 4x.
#
# Clone of dsv41-flash-jovian-r36-4x retargeted at the r37 source lock — the
# lab's newest published DS4.1 qualification. r37's headline is a
# kernel-dispatch overhead removal: no more repeated CPU argument-list
# expansion, mutable-schema inspection, scalar dtype allocations or
# tensor-equality calls per dispatch (no quantization or sampling change);
# plus measured B12X projection tiles, four FP32 projection partials, visible
# index-score tiles, a compact mHC prefill config, removed attention-output
# copies and bounded short-index scratch. Their RTX-quartet qualification
# (public-source control -> r37): uncached 32K prefill 14,896 -> 21,399 tok/s
# (+43.65%), C1 output 238.80 -> 281.08 tok/s (+17.70%, budget-50 three-cell
# median), C1 verifier +11.92%, Sieve +8.17%. r37 also corrects the named
# `high` reasoning budget 50 -> 75 (behaviour, not speed — our benches pass
# reasoning_effort explicitly). Carries the B12X small-tile barrier fix (#363)
# and TP3 prefill-head support (TP3 is not serving-qualified here).
#
# Image: our own build for GB10 sm_121a from docker-images/glm53-jovian-r27
# with profile jovian-r37-ds41 (vLLM c687594b8a8 + b12x f1c4e9dd5b1; LMCache
# 29bc5a2e / FlashKDA 3b225bf2 / DeepGEMM 8b1392b9 unchanged; torch 2.13.0 +
# flashinfer 0.6.17 = the same GB10 pins as r35/r36, dep gates re-checked
# against the r37 tree; liburing 2.5 + dev headers already in the apt lock).
# The published lab image is CUDA 13.3 x86 / sm_120 — ours is the documented
# CUDA 13.2 aarch64 / sm_121a deviation.
#
# Serving profile, GB10 deviations and the Engram-on-node-NVMe mechanism are
# unchanged from r36 (see that recipe's header + history/night-shifts/
# 2026-09-12.md "Day follow-up 2/3"): TP4/DCP1, native DSpark K7 (greedy
# draft, standard rejection, adaptive verification), 512K (our floor — their
# 131K qualified profile is not usable for us), GMU 0.88, B12X attn/MoE/linear,
# block 256, instanttensor, Engram table on disk via
# [composed_model].local_files hardlinked from the eugr lane's per-rank sparse
# copies in /var/tmp/engram-local (48 GB real per node, zero extra disk).
#
# Measured reference points (dev65/seed 42, d0 c1):
#   r36 + local Engram @512K   154/176 (88), 3,774 pp / 45.7 tg /  822 ms TTFT
#   r36 + local Engram @131K   149/176 (85), 3,654 pp / 43.9 tg /  826 ms
#   r36 + Engram over NFS      158/176 (90),   401 pp / 33.7 tg / 5,061 ms
#   eugr dsv41 prod (K5)       167/176 (95), 1,642 pp / 47.0 tg / 1,738 ms
# What r37 targets: the same r36 profile with their +43% prefill fix. The A/B
# vs the eugr prod decides (quality gap vs speed lead) — plan:
# history/night-shifts/2026-09-13.md.
#
# PREREQUISITE: weights in the shared NAS HF cache; per-node local Engram
# sparse copies under /var/tmp/engram-local/DeepSeek-V4.1-Flash.
engine = "vllm"
name = "DeepSeek-V4.1-Flash (jovian r37 img, 4x DCP1, DSpark K7 greedy + Engram disk)"
docker_image = "jovian-r37-ds41:local"
docker_image_dir = "glm53-jovian-r27"
build_profile = "jovian-r37-ds41"
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).
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"