# DeepSeek-V4.1-Flash (552B MoE, MXFP4 experts, Engram-on-disk, DCP2) — 4x
# DGX Spark.
#
# DCP2 variant of dsv41-flash-4x-b12x (the 167/176 lane): same config plus
# decode context parallelism (4 TP ranks = 2 DCP groups of 2) with Aiden's
# a2a all-to-all combine — the production-1.1 overlay ported onto our tree
# (docker-images/dsv41-flash-dcp2/, 23-file superset; provenance + triage in
# top-level dcp2/README.md).
#
# A/B target: DSV4.1 quality/speed at DCP2's KV capacity. On Aiden's own
# stack DCP2 gave 2.5M-token KV @300K but 158/176 (−9) and slower single-
# stream; this variant measures the same trade on our stack.
#
# Image: dsv41-flash-dcp2:local (docker-images/dsv41-flash-dcp2/ — thin
# FROM over dsv41-flash-b12x:local, 23-file DCP2 overlay COPY, verified
# symbol-complete at build time).
#
# Model: deepseek-ai/DeepSeek-V4.1-Flash (476 GB on the NFS HF cache,
# refs/main). The two 203 GB FP8 Engram n-gram tables are NEVER loaded into
# HBM — the baked engram.py patch keeps them on disk (DSV41_ENGRAM_DISK=1)
# and reads rows on demand. Node-local sparse copies of each rank's rows
# (48 GB, engram-local.json) accelerate decode once created (DSV41_ENGRAM_DIR
# below, added after the copies exist — first boot reads straight from NFS).
#
# Config mirrors their boot-10 serving setup (launch/dsv41-tp4.sh) with
# local deviations:
#   CUDA graphs FULL_AND_PIECEWISE + exact DSpark capture sizes (no padded
#   rows — padded speculative batches hang SM120 sparse MLA, FlashInfer
#   #5015), DSpark k=5 with the checkpoint's own draft layers, tools +
#   vision on.
#   * GMU 0.85 (their 0.80): the indexer prefill buffer grows with max
#     context (~5.2 GiB at 1M, ~2.6 GiB at 512K); the extra budget keeps
#     the KV pool at ~11 GiB (≈1.6M tokens) with host headroom.
#   * 512K ctx (2026-09-11, operator): matches the GLM prod convention
#     (524288) and stays inside the row width their top-k fix was
#     validated for (300K measured, 1M-wide never re-run even by them).
#     The 1.6M-token pool gives 3× headroom at 512K.
#   * thinking ON + reasoning_effort=max by default (request can override
#     via chat_template_kwargs), served as "deepseek-v4-flash" (gateway
#     alias, no config churn).
#
# Reference: Aiden's write-up of the same stack + their deltas (B12X MoE
# backend, explicit KV budget, DCP2 port) —
# https://aidenle.com/recipes/deepseek-v4-1-flash-4x-dgx-spark/
# (page says production-1.1 in places; the DCP2-baked image is
# production-1.2 — the page was not updated to the end).

engine = "vllm"
name = "DeepSeek-V4.1-Flash MXFP4 (dsv41 img, 4x, DSpark k=5, Engram-on-disk, DCP2 a2a)"
docker_image = "dsv41-flash-dcp2:local"
docker_image_dir = "dsv41-flash-dcp2"
model = "deepseek-ai/DeepSeek-V4.1-Flash"
required_nodes = 4
tensor_parallel_size = 4
pipeline_parallel_size = 1
distributed_backend = "mp"
shm_size = "32gb"
# NO recipe entrypoint: the base image bakes ENTRYPOINT ["vllm","serve"]
# (vllm/vllm-openai), so the launcher's CMD (model + flags) completes it.

args = [
    # Stable gateway alias — clients/gateways keep the DeepSeek-V4-Flash
    # model name; no per-deployment config churn.
    "--served-model-name", "deepseek-v4-flash",
    "--trust-remote-code",
    "--distributed-executor-backend", "mp",
    "--gpu-memory-utilization", "0.85",
    "--max-model-len", "524288",
    "--max-num-seqs", "8",
    "--max-num-batched-tokens", "8192",
    # boot-4 fix: the V4 indexer backend refuses the smallest-listed block
    "--block-size", "128",
    # boot-10: exact capture sizes for DSpark k=5, SEQS=8 (multiples of 5
    # and 6 up to 48 — every decode batch gets an exact FULL graph, no padding)
    "--compilation-config", """
'{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[5,6,10,12,15,18,20,24,25,30,35,36,40,42,48]}'""",
    "--speculative-config", """
'{"method":"dspark","num_speculative_tokens":5,"draft_sample_method":"probabilistic","rejection_sample_method":"block","enable_adaptive_verification":false}'""",
    "--moe-backend", "b12x",
    # DCP2: sharded KV across DCP groups (2 ranks each), the new
    # all-to-all + LSE combine path (Aiden production-1.1's runtime config).
    "--decode-context-parallel-size", "2",
    "--dcp-comm-backend", "a2a",
    "--tool-call-parser", "deepseek_v41",
    "--enable-auto-tool-choice",
    "--reasoning-parser", "deepseek_v41",
    "--default-chat-template-kwargs", """
'{"thinking": true, "reasoning_effort": "max"}'""",
]

[env]
HF_HUB_OFFLINE = "1"
# Engram tables stay on disk (the baked engram.py patch); 32 parallel row
# reads, 16 MB chunks (their launch defaults)
DSV41_ENGRAM_DISK = "1"
DSV41_ENGRAM_DISK_THREADS = "32"
DSV41_ENGRAM_DISK_CHUNK = "16"
# Node-local sparse Engram copy (per-rank row ranges from the first boot's
# "Engram DISK mode" log lines; built with tools/engram_local.py, verified
# byte-for-byte). NVMe instead of NFS for the 23.6 GiB/rank tables.
DSV41_ENGRAM_LOCAL = "1"
DSV41_ENGRAM_DIR = "/engram-local"
# CUDA graphs: the Engram rows are staged in prepare_inputs (baked
# model_state.py prestage patch) so the whole decode step is capturable
VLLM_USE_BREAKABLE_CUDAGRAPH = "1"
VLLM_HAS_FLASHINFER_CUBIN = "1"
VLLM_USE_RUST_FRONTEND = "0"
# runtime JIT wedge countermeasure (their boot-3 lesson): kernels are
# prebuilt in the image, but any residual JIT must not OOM the box
MAX_JOBS = "2"
VLLM_ENGINE_READY_TIMEOUT_S = "3600"
PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True"
TORCH_CUDA_ARCH_LIST = "12.1a"
FLASHINFER_CUDA_ARCH_LIST = "12.1a"
OMP_NUM_THREADS = "2"

# Per-node sparse Engram copy (model-00047/00048 shards + engram-local.json)
[[extra_mounts]]
host = "/var/tmp/engram-local/DeepSeek-V4.1-Flash"
container = "/engram-local"
read_only = true

# JIT caches persist per-node (self-namespace by engine version key; the
# image prebakes the flashinfer 0.7.0rc1 SM120 kernels, first boots re-JIT
# the rest)
[[extra_mounts]]
host = "~/sm121-cache/jit-121f/flashinfer"
container = "/root/.cache/flashinfer"

[[extra_mounts]]
host = "~/sm121-cache/jit-121f/torchinductor"
container = "/root/.cache/torchinductor"

[[extra_mounts]]
host = "~/sm121-cache/jit-121f/vllm"
container = "/root/.cache/vllm"

# B12X kernel JIT cache (B12X_CACHE_DIR; prebuilt into the image, persists
# per node like the other JIT caches)
[[extra_mounts]]
host = "~/sm121-cache/jit-121f/b12x"
container = "/root/.cache/b12x"