launcher/recipes/dsv41-flash-4x-aiden11/recipe.toml

# Aiden's published production-1.1 DCP2 stack (prebuilt image, no build).
#
# Image: aidendle94/sparkrun-vllm-dsv41-gb10:production-1.1 (Docker Hub, 23.4 GB,
# published 2026-09-11T16:54Z). Internals (verified):
#   vllm 0.28.1rc1.dev388+g8a728663c (base vllm/vllm-openai nightly 8a728663,
#   NOT our e47aa780 pin) + b12x 1.3.0 + FlashInfer 0.7.0rc1 + torch 2.13.0+cu130.
#   The 24-file DCP2 patch set is baked into the tree (manifest at
#   /opt/dsv41-dcp2-mounts.txt) + FI source with trtllm_dcp_alltoall.cu /
#   cake_dcp.py. DCP knobs are env-gated, OFF by default in-image:
#   DSV41_DCP_STAGE=3 (default), DSV41_DCP_COMBINE=ag_rs, DSV41_DCP_PREFILL_GATHER=0,
#   DSV41_ENGRAM_MMAP=0, DSV41_ENGRAM_PREFETCH=0.
#
# This recipe = their DCP2 core (TP4 + --decode-context-parallel-size 2 +
# --dcp-comm-backend a2a + 8 GiB pinned KV, their measured 300K-ctx row) on our
# serving config (DSpark k=5 exact capture, max-num-seqs 8, thinking+max
# default, deepseek-v4-flash alias) so the A/B isolates the DCP2 delta.
#
# A/B baseline: dsv41-flash-4x-b12x (167/176 @ 512K ctx, DSV41_ENGRAM_LOCAL on
# our clone-at-pin stack). Context differs on purpose: their top-k/indexer
# validation row is 300K.
#
# Engram: our per-node local copies (47.2 GiB each, /var/tmp/engram-local) —
# the engine reads its rank's row ranges from there, no NFS at runtime.
#
# Prebuilt: no docker_image_dir — the image is pulled manually to the nodes
# (sm121 does not pull external images).

engine = "vllm"
name = "DeepSeek-V4.1-Flash MXFP4 (Aiden production-1.1 img, 4x, DCP2)"
docker_image = "aidendle94/sparkrun-vllm-dsv41-gb10:production-1.1"
model = "deepseek-ai/DeepSeek-V4.1-Flash"
port = 30000
required_nodes = 4
tensor_parallel_size = 4
pipeline_parallel_size = 1
distributed_backend = "mp"
shm_size = "32gb"

args = [
    "--served-model-name", "deepseek-v4-flash",
    "--trust-remote-code",
    "--distributed-executor-backend", "mp",
    "--gpu-memory-utilization", "0.85",
    # Aiden's validated row: 300K ctx (their top-k fix is validated <=300K)
    "--max-model-len", "300000",
    "--max-num-seqs", "8",
    "--max-num-batched-tokens", "8192",
    "--block-size", "128",
    # Our exact capture sizes for DSpark k=5, SEQS=8 (no padding batches)
    "--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",
    "--linear-backend", "b12x",
    "--kv-cache-dtype", "fp8",
    # Their DCP2 core: 2-way decode context parallelism over the a2a path
    "--decode-context-parallel-size", "2",
    "--dcp-comm-backend", "a2a",
    # Their pinned KV (8 GiB = 1.26M tokens at 300K, ~1.89M with DCP2)
    "--kv-cache-memory-bytes", "8589934592",
    "--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 on node-local sparse copies (disk mode, their launch defaults)
DSV41_ENGRAM_DISK = "1"
DSV41_ENGRAM_DISK_THREADS = "32"
DSV41_ENGRAM_DISK_CHUNK = "16"
DSV41_ENGRAM_LOCAL = "1"
DSV41_ENGRAM_DIR = "/engram-local"
# DCP2 activation (image defaults are off; Aiden's launcher enables these)
DSV41_DCP_STAGE = "3"
DSV41_DCP_COMBINE = "a2a"
DSV41_DCP_PREFILL_GATHER = "1"
DSV41_ENGRAM_MMAP = "1"
DSV41_ENGRAM_PREFETCH = "1"
# Runtime JIT wedge countermeasure + arch pins (same stack family)
MAX_JOBS = "2"
OMP_NUM_THREADS = "2"
PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True"
TORCH_CUDA_ARCH_LIST = "12.1a"
FLASHINFER_CUDA_ARCH_LIST = "12.1a"
VLLM_ENGINE_READY_TIMEOUT_S = "3600"

# 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 (fresh tree => first boot re-JITs, then cached)
[[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)
[[extra_mounts]]
host = "~/sm121-cache/jit-121f/b12x"
container = "/root/.cache/b12x"