# DeepSeek-V4-Flash-Vision-Exp on 2× GB10 — AIDEN'S IMAGE (production-3.72-vision).
#
# aiden's vision guide translated to our 2-node ds4 launcher, RUN THROUGH HIS
# OWN env-script entrypoint (/usr/local/bin/deepseek-v4-flash-2x-spark-
# entrypoint → /tmp copy of /opt/vllm-src/serve-ds4-flash-spark.sh). The script
# generates the full vllm serve command from the env below; do NOT add args —
# the mod drops our launcher tail ($@) so his env stays the single source.
#
# Image: aidendle94/sparkrun-vllm-ds4-gb10:production-3.72-vision. Their TWO
# vllm wheels: dist-packages (plain, no vision) and /opt/vllm-src/.venv
# (multimodal-wrapped DeepseekV4 + DSpark/DCP fixes) — only his script picks
# the right one. Production profile at TP4/DCP2: 66-71 tok/s @ acceptance
# 3.5-3.8; this 2x run is his "DSpark, DCP1" profile (73.8-75.2 @ TP4).
#
# Required envs his script demands: NODE_RANK (mod derives from VLLM_HOST_IP
# vs rank-0's IB IP), MASTER_ADDR (mod; ds4 = 192.168.200.17), VLLM_HOST_IP +
# NCCL_IB_HCA/GID/SOCKET/GLOO/UCX/TP (launcher [nccl.ds4] + auto-GID).
#
# Boot markers: "DSpark draft model loaded: 97 params" → "Model loading took
# ~82 GiB" → "GPU KV cache size" ≈ 600K (12 GiB pin at this GMU) → "Application
# startup complete". Smoke: 47*83 → 3901 (thinking off) + image_url color ask.
#
# Known quirk: first image request ~6 s cold, ~1-2 s warm (partition kernel).

engine = "vllm"
name = "DeepSeek V4 Flash Vision-Exp (aiden image + env-script, 2x Spark, DSpark-6)"
docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-3.72-vision"
model = "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp"
required_nodes = 2
tensor_parallel_size = 2
pipeline_parallel_size = 1
distributed_backend = "mp"
shm_size = "32gb"
# his script copies /opt/vllm-src/serve-ds4-flash-spark.sh to /tmp and execs
# it; the mod block below seds the "$@" tail off the SOURCE before that copy
# (our launcher's appended model/args would otherwise duplicate his generated
# ones), and derives NODE_RANK/MASTER_ADDR.
entrypoint = ["/usr/local/bin/deepseek-v4-flash-2x-spark-entrypoint"]

# (no args: the env-script builds everything; the launcher tail is dropped by
# the mod below — if the sed ever fails, vllm errors on duplicate positional
# and the boot aborts loudly instead of misconfiguring.)

[mods]
commands = [
    # rank-0 IB ip on ds4 (config.toml [cluster.ds4] rank 0 = dgx3 192.168.200.17)
    "export MASTER_ADDR=\"192.168.200.17\"",
    "if [ \"${VLLM_HOST_IP:-}\" = \"192.168.200.17\" ]; then export NODE_RANK=0; else export NODE_RANK=1; fi",
    "export MASTER_PORT=\"29511\"",
    "export PORT=\"30000\"",
    "export NODE_RANK=\"${NODE_RANK:-0}\"",
    # drop the forwarded "$@" his launcher appends to its generated vllm call
    # (our tail would duplicate his model positional); fail loudly if absent
    "sed -i '$s/ \\\"\\$@\\\"$//' /opt/vllm-src/serve-ds4-flash-spark.sh || exit 1",
    "grep -q 'vllm.entrypoints.cli.main serve' /opt/vllm-src/serve-ds4-flash-spark.sh",
]

[env]
# his script's inputs (same names verbatim; all other B12X_*/AOT/JIT knobs
# the script exports itself)
SPARK_HF_HOME = "/opt/local-hf"
HF_HUB_OFFLINE = "1"
TRANSFORMERS_OFFLINE = "1"
NNODES = "2"
TP_SIZE = "2"
DCP_SIZE = "1"
MODEL_ID = "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp"
DSPARK_MODEL_ID = "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp"
DSPARK_MODEL_REVISION = "e46e16bf6035c6f317eb2ac7458eb0362926d402"
SERVED_MODEL_NAME = "deepseek-v4-vision"
GPU_MEMORY_UTILIZATION = "0.80"
KV_CACHE_MEMORY_BYTES = "12884901888"
MAX_MODEL_LEN = "393216"
MAX_NUM_SEQS = "8"
MAX_NUM_BATCHED_TOKENS = "8192"
MAX_CUDAGRAPH_CAPTURE_SIZE = "56"
ENABLE_PREFIX_CACHING = "1"
VLLM_ENABLE_DSPARK = "1"
VLLM_ENABLE_MTP = "0"
NUM_SPECULATIVE_TOKENS = "6"
DSPARK_DRAFT_SAMPLE_METHOD = "greedy"
DSPARK_SPS_CURVE = "auto"
VLLM_CUDAGRAPH_MODE = "FULL_AND_PIECEWISE"
VISION_MHC = "1"
# KDA/CKV gather is a DCP-side lane; DCP1 needs none of the VLLM_DSV4_CKV_*

# Weights from per-node NVMe copies (fast path). extra_mounts stay AFTER args.
[[extra_mounts]]
host = "~/dsv4vision-hf"
container = "/opt/local-hf"
env = "HF_HOME"