File size: 929 Bytes
0693474 | 1 2 3 4 5 6 7 8 9 10 11 12 13 | 15c15
< export FAB_IF=${FAB_IF:-enp1s0f0np0} FAB_HCA=${FAB_HCA:-rocep1s0f0} # fabric netdev and its RoCE device (DGX Spark names)
---
> export FAB_IF=${FAB_IF:-enp1s0f1np1} FAB_HCA=${FAB_HCA:-rocep1s0f1} # sparkhead/sparkworker: port f1 is the cabled one (f0 DOWN) # fabric netdev and its RoCE device (DGX Spark names)
17c17
< export MODEL_DIR=$HOME/models/DeepSeek-V4.1-Flash-1017-EXL3-2.75bpw-DGX-Spark # incl. the two 1017 Engram shards (scripts/download_model.sh)
---
> export MODEL_DIR=${MODEL_DIR:-/srv/dsv41/model} # node-local NVMe on both (head: symlink to ~/models/...; worker: real dir; ~/models on sparkworker is NFS) # incl. the two 1017 Engram shards (scripts/download_model.sh)
39c39
< NCCL_IB_HCA=rocep1s0f0,roceP2p1s0f0 EXL3_RDMA_DEVICES=rocep1s0f0 NCCL_PROTO=^LL128 NCCL_CUMEM_ENABLE=0 \
---
> NCCL_IB_HCA=rocep1s0f1,roceP2p1s0f1 EXL3_RDMA_DEVICES=rocep1s0f1 NCCL_PROTO=^LL128 NCCL_CUMEM_ENABLE=0 \
|