zen5-engine/lab.yaml

# zen5 home lab — 4 heterogeneous nodes, 448 GB unified memory total.
#
# Hostnames are mDNS names on the local network. Confirm reachable with:
#   bin/lab self-test
#
#   spark.local — DGX Spark GB10 Blackwell, 128 GB, 200 GbE ConnectX-7  → coordinator + worker
#   dbc.local   — MacBook Pro M4 Max, 128 GB                            → worker (MLX)
#   evo.local   — GMKtec EVO-X2 (Ryzen AI Max+ 395, Strix Halo 8060S)   → worker (ROCm)
#   m1max.local — MacBook Pro M1 Max, 64 GB                             → worker (MLX)
#
# Spark wins coordinator because (a) ConnectX-7 200 GbE — 20× any other NIC,
# (b) Blackwell tensor cores are 5–10× the others for trim-mean reductions,
# (c) it's the only box that can serve aggregates + train concurrently.
# It's also a worker — too much memory + Blackwell to leave idle.
#
# Sync at LoRA-delta granularity over HTTP — no cross-vendor NCCL/Gloo.
# Aggregation: byzantine_robust (trimmed mean, tolerates 1 bad worker).
# Auth: ZEN_LAB_SECRET_<NAME> env vars per node; null for local-only dev.

job_dir: /var/lib/zen-fed
sync_interval_steps: 8
aggregation: byzantine_robust

nodes:
  - name: spark
    host: spark.local
    role: hybrid                         # coordinator + worker
    backend: cuda                        # GB10 Blackwell, CUDA via PyTorch
    memory_gb: 128
    nic_gbps: 200                        # ConnectX-7
    tflops: 250                          # BF16 dense; FP4 ≈ 1 PFLOP (unused for router)
    pin_experts: []                      # auto-pinned by scheduler at run time
    pin_modalities: [text]
    auth_token: ${ZEN_LAB_SECRET_SPARK}

  - name: dbc
    host: dbc.local
    role: worker
    backend: mlx                         # MLX is ~2× MPS — required at 128 GB
    memory_gb: 128
    nic_gbps: 10                         # 10 GbE; bump to 80 if TB5 networking
    tflops: 34                           # M4 Max GPU
    pin_experts: []
    pin_modalities: [vision, video, 3d]
    auth_token: ${ZEN_LAB_SECRET_DBC}

  - name: evo
    host: evo.local
    role: worker
    backend: rocm                        # EVO-X2 = Ryzen AI Max+ 395 / Radeon 8060S
    memory_gb: 128
    nic_gbps: 10
    tflops: 60
    pin_experts: []
    pin_modalities: [text, audio]
    auth_token: ${ZEN_LAB_SECRET_EVO}

  - name: m1max
    host: m1max.local
    role: worker
    backend: mlx
    memory_gb: 64
    nic_gbps: 10
    tflops: 10                           # M1 Max GPU (32-core variant)
    pin_experts: [qwen35_0_8b, qwen35_9b]   # T0+T1 small dense — fit cleanly
    pin_modalities: [text]
    auth_token: ${ZEN_LAB_SECRET_M1MAX}