# EXPERIMENTAL — DeepSeek-V4-Flash-0731 DSpark, built FROM SOURCE from the
# newest local-inference-lab b12x fork ("eldritch"), instead of the prebuilt
# aidendle94 image.
#
#   fork branch: feat/gg-ds4-0731-launcher (e63a190ee) — torch 2.12.0,
#                flashinfer 0.6.15, cutlass-dsl 4.6.0, sm_121a
#   image:       vllm-ds4-dspark-src:probe (built on demand via docker_image_dir)
#
# Serving flags/env mirror the fork's authoritative serve-ds4-flash-spark.sh at
# that commit (DSpark k=5, draft attention backend B12X_MLA_SPARSE, b12x MoE/MLA,
# cudagraph capture size = max_num_seqs * (k+1)). Adapt cluster-side: TP=2,
# 1 node per rank, no DCP.
#
# ⚠️ NOT production. The image is a multi-hour from-source build. Boot may need
# tuning; verify acceptance/throughput before any real traffic.

engine = "vllm"
name = "DeepSeek V4 Flash DSpark (EXPERIMENTAL, from source)"
docker_image = "vllm-ds4-dspark-src:probe"
model = "deepseek-ai/DeepSeek-V4-Flash-0731"
required_nodes = 2
tensor_parallel_size = 2
pipeline_parallel_size = 1
distributed_backend = "mp"
shm_size = "32gb"
docker_image_dir = "deepseek-v4-flash-dspark-src"

entrypoint = ["vllm", "serve"]

args = [
    "--kv-cache-dtype", "fp8",
    "--block-size", "256",
    "--load-format", "auto",
    "--max-model-len", "524288",
    "--max-num-seqs", "4",
    "--max-num-batched-tokens", "8192",
    # fork rule: max_num_seqs * (k+1) = 4 * 6
    "--max-cudagraph-capture-size", "24",
    # 0.78, not 0.80: the SvangenStudios agent-serving benchmark (v0.1.1) on
    # this exact stack found "0.80 boots but dies under traffic". GMU is a KV
    # lever, not a safety lever, when weights dominate memory. See
    # specs/tech-architecture/dgx-spark-agent-serving-findings.md.
    "--gpu-memory-utilization", "0.78",
    "--served-model-name", "deepseek-v4-flash", "default",
    "--attention-backend", "B12X_MLA_SPARSE",
    "--moe-backend", "b12x",
    "--enable-chunked-prefill",
    "--enable-prefix-caching",
    "--enable-flashinfer-autotune",
    "--compilation-config", """
'{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}'""",
    "--async-scheduling",
    "--no-scheduler-reserve-full-isl",
    # DSpark k=5 (dspark_block_size == 5); draft attention backend per fork
    "--speculative-config", """
'{"method":"dspark","num_speculative_tokens":5,"attention_backend":"B12X_MLA_SPARSE"}'""",
    "--generation-config", "vllm",
    "--default-chat-template-kwargs.thinking=true",
    "--default-chat-template-kwargs.reasoning_effort=high",
    "--default-chat-template-kwargs.temperature=1",
    "--default-chat-template-kwargs.top_p=1",
    "--disable-custom-all-reduce",
    "--trust-remote-code",
]

[env]
# --- from the fork's serve-ds4-flash-spark.sh (e63a190ee) ---
VLLM_ALLOW_LONG_MAX_MODEL_LEN = "1"
VLLM_USE_AOT_COMPILE = "1"
VLLM_USE_BREAKABLE_CUDAGRAPH = "0"
VLLM_USE_MEGA_AOT_ARTIFACT = "1"
VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS = "1"
VLLM_USE_FLASHINFER_SAMPLER = "1"
VLLM_USE_B12X_WO_PROJECTION = "1"
VLLM_USE_B12X_MHC = "1"
VLLM_USE_B12X_FP8_GEMM = "1"
VLLM_USE_B12X_MOE = "1"
VLLM_USE_B12X_SPARSE_INDEXER = "1"
VLLM_USE_V2_MODEL_RUNNER = "1"
VLLM_PCIE_ALLREDUCE_BACKEND = "b12x"
VLLM_ENABLE_PCIE_ALLREDUCE = "1"
B12X_MLA_SM120_UNIFIED = "1"
B12X_DENSE_SPLITK_TURBO = "1"
B12X_W4A16_TC_DECODE = "1"
B12X_MOE_FORCE_A8 = "1"

# --- build/arch + memory ---
CUTE_DSL_ARCH = "sm_121a"
PYTORCH_CUDA_ALLOC_CONF = "expandable_segments:True"
DG_JIT_USE_NVRTC = "0"
USE_CUDNN = "1"
CUDA_VISIBLE_DEVICES = "0"
CUDA_DEVICE_ORDER = "PCI_BUS_ID"
OMP_NUM_THREADS = "4"

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