launcher/recipes/dsv4-vision-aiden-373/recipe.toml

# DeepSeek-V4-Flash-Vision (aiden production-3.73-vision image, 2x Spark).
#
# Image: aidendle94/sparkrun-vllm-ds4-gb10:production-3.73-vision (Docker Hub,
# arm64, pushed 2026-09-06). The "vision" aiden line: their pinned
# serve-ds4-flash-spark.sh launcher (env-driven; ds4-vision 0.20 branch)
# wraps a compiled vLLM wheel (B12X MLA sparse attention, b12x MoE/linear,
# instanttensor load, FULL_AND_PIECEWISE cudagraphs, async scheduling,
# deepseek_v4 tokenizer/tool/reasoning parsers).
#
# Model: deepseek-ai/DeepSeek-V4-Flash-Vision-Exp (157 GB, 78.5 GiB/node at
# 2x; cached in the NAS hub at snapshot 6821d6ad). Their launcher resolves
# the model from the MODEL_PATH / MODEL_REVISION_OVERRIDE env vars, so this
# recipe leaves `model` EMPTY — the launcher skips the positional model for
# entrypoint-driven images (added capability, see AGENTS.md).
#
# Launcher env contract (read from the image's /opt/vllm-src/serve-ds4-
# flash-spark.sh): requires NODE_RANK, MASTER_ADDR, VLLM_HOST_IP,
# NCCL_IB_HCA, NCCL_IB_GID_INDEX, NCCL_SOCKET_IFNAME, GLOO_SOCKET_IFNAME,
# UCX_NET_DEVICES, TP_SOCKET_IFNAME (all injected by the sm121 docker path;
# NODE_RANK/MASTER_ADDR as env since their launcher doesn't parse them as
# CLI flags). Defaults used here: TP=2, no spec (MTP/DSpark both off),
# GMU 0.80, KV cap 24.4 GiB (SPARK_KV_CACHE_MEMORY_BYTES), max-model-len
# 500K, max-num-seqs 4, max-num-batched-tokens 8192, B12X_MLA_SPARSE
# attention, prefix caching on, flashinfer autotune on.
#
# SPARK_HF_HOME points their HF_HOME (which the launcher overrides to
# ~/.cache/vllm-huggingface) at the NAS mount our docker path binds at
# /root/.cache/huggingface.
#
# Boot markers: "Loading safetensors" (instanttensor, ~20 min from NAS) →
# "GPU KV cache size" → "Application startup complete".
# Bench model name: DeepSeek-V4-Flash (their SERVED_MODEL_NAME default).

engine = "vllm"
name = "DeepSeek-V4-Flash-Vision (aiden prod-3.73-vision img, 2x Spark)"
docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-3.73-vision"
model = ""
port = 30000
required_nodes = 2
tensor_parallel_size = 2
pipeline_parallel_size = 1  # explicit: launcher default would be PP=nnodes=2
 # (this image's vllm fork asserts the TP×PP world size against the device count)
distributed_backend = "mp"
shm_size = "32gb"
entrypoint = []

args = []

[env]
MODEL_PATH = "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp"
MODEL_REVISION_OVERRIDE = "6821d6ad3681a4b137b066b76094fa82ebd0a380"
SPARK_HF_HOME = "/root/.cache/huggingface"
HF_HUB_OFFLINE = "1"
VLLM_ENGINE_READY_TIMEOUT_S = "3600"
VLLM_NO_USAGE_STATS = "1"