# DeepSeek-V4.1-Flash (552B MoE, MXFP4 experts, Engram-on-disk) — 4x DGX Spark.
#
# Image: dsv41-flash:local (docker-images/dsv41-flash/ — tonyd2wild overlay
# chain as clone-at-pin build: vllm dsv41-feat@e47aa780 + sm121
# _C_stable_libtorch + FlashInfer 0.7.0rc1 + 7 boot-10 patches baked in,
# MIT provenance).
#
# 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):
#   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, GMU 0.80,
#   300K ctx (KV pool ~1.07M tokens = 3.57x headroom), tools + vision on,
#   thinking off by default.

engine = "vllm"
name = "DeepSeek-V4.1-Flash MXFP4 (dsv41 img, 4x, DSpark k=5, Engram-on-disk)"
docker_image = "dsv41-flash:local"
docker_image_dir = "dsv41-flash"
model = "deepseek-ai/DeepSeek-V4.1-Flash"
required_nodes = 4
tensor_parallel_size = 4
pipeline_parallel_size = 1
distributed_backend = "mp"
shm_size = "32gb"
entrypoint = ["vllm", "serve"]

args = [
    "--served-model-name", "dsv41",
    "--trust-remote-code",
    "--distributed-executor-backend", "mp",
    "--gpu-memory-utilization", "0.80",
    "--max-model-len", "300000",
    "--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}'""",
    "--tool-call-parser", "deepseek_v41",
    "--enable-auto-tool-choice",
    "--reasoning-parser", "deepseek_v41",
    "--default-chat-template-kwargs", """
'{"thinking": false}'""",
]

[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"
# 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"

# 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"