launcher/recipes/deepseek-v4-flash-dspark38/recipe.toml

# DeepSeek-V4-Flash-0731 DSpark — 2× GB10 / DGX Spark, TP=2 over RoCE.
#
# Variant targeting aidendle94's production-3.8 image ("dsv4/apostolic" b12x
# harness). This is the BUG-2026-08-03T204037 bring-up recipe — the two images
# are *different b12x forks*, not a version bump from production-3.75, so the
# launch contract differs:
#
#   - production-3.8 is a clean-overlay base (null ENTRYPOINT, CMD sleep
#     infinity) whose intended launcher is /usr/local/bin/dsv4-vllm-entrypoint.
#     We set that as the recipe entrypoint so the JIT toolchain env
#     (DG_JIT_NVCC_COMPILER, NVCC_PREPEND_FLAGS + CCCL include path, CUDAHOSTCXX)
#     and the /cache JIT dir layout are active. The 3.75 recipe ran vLLM directly
#     because that image was self-contained with everything baked in.
#   - VLLM_USE_V2_MODEL_RUNNER is dropped (experimental in this fork; the
#     working glm52 recipe does not set it).
#   - tiny-decode (VLLM_B12X_TINY_DECODE) is enabled at GMU 0.78 + CHUNK=8
#     (the load-time repack needs headroom; it OOM'd at GMU 0.9).
#   - /cache is mounted so the wrapper's JIT caches persist across restarts.

engine = "vllm"
name = "DeepSeek V4 Flash DSpark (production-3.8)"
# Local overlay of the production-3.8 base built from
# docker-images/deepseek-v4-flash-dspark38 (root-cause patch for the b12x
# cutlass from_dlpack grad bug that blocked boot; see the Dockerfile).
docker_image = "sparkrun-vllm-ds4-gb10:dspark38"
model = "deepseek-ai/DeepSeek-V4-Flash-0731"
required_nodes = 2
tensor_parallel_size = 2
pipeline_parallel_size = 1
distributed_backend = "mp"
shm_size = "32gb"

# production-3.8 launch contract: the clean-base image must be started via the
# dsv4-vllm-entrypoint wrapper, which sets the JIT toolchain + /cache dirs and
# `exec vllm "$@"`. `serve <recipe-args/auto flags>` land as "$@".
entrypoint = ["/usr/local/bin/dsv4-vllm-entrypoint", "serve"]

args = [
    "--kv-cache-dtype", "fp8",
    "--block-size", "256",
    "--load-format", "auto",
    "--max-model-len", "524288",
    "--max-num-seqs", "8",
    "--max-num-batched-tokens", "8192",
    "--max-cudagraph-capture-size", "64",
    # 0.78 is the safe GMU ceiling on GB10 (weights ~77.7 GiB of ~119.6 GiB);
    # higher boots but dies under traffic. Also leaves headroom for the tiny
    # decode kernel's GPU load-time repack.
    "--gpu-memory-utilization", "0.78",
    "--served-model-name", "deepseek-v4-flash",
    # Engine backends — b12x across attention / MoE / linear (this fork's path)
    "--attention-backend", "B12X_MLA_SPARSE",
    "--moe-backend", "b12x",
    "--linear-backend", "b12x",
    # Prefix caching
    "--enable-prefix-caching",
    # Compilation
    "--compilation-config", """
'{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}'""",
    # DSpark speculative decoding — k=5 (DeepSeek-V4-Flash native
    # dspark_block_size). draft_sample_method greedy.
    "--speculative-config", """
'{"method":"dspark","num_speculative_tokens":5,"draft_sample_method":"greedy"}'""",
    # Tokenizer / reasoning
    "--tokenizer-mode", "deepseek_v4",
    "--tool-call-parser", "deepseek_v4",
    "--reasoning-parser", "deepseek_v4",
    "--enable-auto-tool-choice",
    "--generation-config", "vllm",
    "--default-chat-template-kwargs.thinking=true",
    "--default-chat-template-kwargs.reasoning_effort=high",
    "--disable-custom-all-reduce",
    "--trust-remote-code",
]

# Persistent JIT/compile caches. The wrapper `mkdir -p`s under /cache
# (DG_JIT_CACHE_DIR=/cache/deep_gemm_jit, VLLM_CACHE_ROOT=/cache/vllm,
# TRITON_CACHE_DIR=/cache/triton, etc.), so /cache must be a mounted volume to
# survive restarts. The HF cache is already mounted at /root/.cache/huggingface
# by the launcher.
[[extra_mounts]]
host = "~/.cache/ds4-38-vllm"
container = "/cache"

[env]
# --- b12x fork env family (shared with the proven eugr b12x recipe) ---
VLLM_ATTENTION_BACKEND = "B12X_MLA_SPARSE"
VLLM_USE_B12X_MOE = "1"
VLLM_USE_B12X_MHC = "1"
VLLM_USE_B12X_WO_PROJECTION = "1"
VLLM_USE_B12X_FP8_GEMM = "1"
VLLM_USE_B12X_SPARSE_INDEXER = "1"
VLLM_USE_FLASHINFER_SAMPLER = "1"
VLLM_USE_AOT_COMPILE = "1"
VLLM_USE_MEGA_AOT_ARTIFACT = "0"
VLLM_USE_BREAKABLE_CUDAGRAPH = "0"
VLLM_MEMORY_PROFILE_INCLUDE_ATTN = "1"
VLLM_PREFIX_CACHE_RETENTION_INTERVAL = "4096"
B12X_MLA_SM120_UNIFIED = "1"
B12X_DENSE_SPLITK_TURBO = "1"
B12X_W4A16_TC_DECODE = "1"
# Tiny A8 decode kernel: DISABLED — its GPU load-time MoE-weight repack, on
# top of the tight host RAM (~32 GiB free) + dual-node 155 GiB load, OOM'd both
# hosts (had to power-cycle via HA). Revisit separately if host RAM headroom is
# available. Keep the CHUNK pin for reference.
# VLLM_B12X_TINY_DECODE = "1"
# VLLM_B12X_TINY_CHUNK = "8"

# --- arch / memory / GPU isolation (operational on the sm121 cluster) ---
CUTE_DSL_ARCH = "sm_121a"
TORCH_CUDA_ARCH_LIST = "12.1a"
PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True"
OMP_NUM_THREADS = "4"
CUDA_VISIBLE_DEVICES = "0"
CUDA_DEVICE_ORDER = "PCI_BUS_ID"
CUDA_CACHE_MAXSIZE = "4294967296"

# --- offline ---
HF_HUB_OFFLINE = "1"

# --- NCCL overrides (image-agnostic; launcher injects IB/HCA/socket from
#     cluster config, these are recipe-level extras) ---
NCCL_P2P_LEVEL = "SYS"
NCCL_CUMEM_ENABLE = "0"
NCCL_IB_TIMEOUT = "24"
NCCL_IB_QPS_PER_CONNECTION = "8"
NCCL_BUFFERSIZE = "16777216"
NCCL_SOCKET_FAMILY = "inet"

VLLM_ENABLE_PCIE_ALLREDUCE = "0"
VLLM_PCIE_ALLREDUCE_BACKEND = "cpp"