cluster.env.example

# Copy to cluster.env on every node. Same file everywhere: rank is derived from NODES.

# Rank order. Rank 0 owns the API port and (if you share the checkpoint) the writable copy.
# These addresses only carry TCPStore/Gloo rendezvous - use the L2 all three nodes share.
NODES="192.168.2.36 192.168.2.37 192.168.2.38"
IFACE=enP7s7

# Tensor traffic. TRANSPORT=ib needs the RDMA devices from `ls /sys/class/infiniband`.
# Use mesh (upstream nccl-mesh plugin) if subnet-aware routing cannot handle your wiring,
# socket for debugging only.
TRANSPORT=ib
IB_HCAS=rocep1s0f0,rocep1s0f1,roceP2p1s0f0,roceP2p1s0f1

# Same path on every node. One shared read-only copy over NFS is fine.
MODEL_DIR=/home/user/.cache/model-shelf/models/safetensors/LibertAIDAI/GLM-5.3-Flash-NVFP4
UPSTREAM_DIR=/home/user/src/GLM-5.3-Flash-3x-DGX-Sparks
OVERLAY_DIR=/home/user/src/glm53-overlay

IMAGE=glm53:sm121-v8
PORT=8045

# 524288 x 2.5 concurrency x 8793 B/token/rank (MTP-4). Unset gives ~18x and wastes ~27 GiB.
MAX_MODEL_LEN=524288
KV_CACHE_MEMORY=11523000000
MTP_TOKENS=4

# Speculation: mtp (native head, default) | dflash (DFlash2 drafter) | none
# For the DFlash2 lane, build docker/Dockerfile.glm53-sm121-v12-dflash2, extract
# nvrtc.h with scripts/extract-nvrtc-header.sh, and patch the TP=3 GLM model overlay
# with docker/dflash2-overlay/patch_glm_aux_capture.py. The drafter must be head-padded
# first -- 32 q / 8 kv heads divide neither by 3, and draft_tensor_parallel_size is
# ignored on this path. See scripts/pad-dflash2-drafter.py and results/dflash2-tp3-*.md.
#   scripts/pad-dflash2-drafter.py ~/glm53-dflash2-draft ~/glm53-dflash2-draft-tp3
# IMAGE=glm53:sm121-v12-dflash2
# SPEC_METHOD=dflash
# DFLASH_TOKENS=7                              # block_size 8 minus the target's own token
# DFLASH2_DIR=$HOME/glm53-dflash2-draft-tp3    # the 48 q / 12 kv padded drafter

# Optional: default reasoning verbosity (low|high; unset means the model default, max).
# REASONING_EFFORT=low