af94058203
* [LLM] Support CPU Deepspeed distributed inference * Update run_deepspeed.py * Rename * fix style * add new codes * refine * remove annotated codes * refine * Update README.md * refine doc and example code
18 lines
615 B
Bash
18 lines
615 B
Bash
#/bin/bash
|
|
source bigdl-nano-init
|
|
unset OMP_NUM_THREADS # deepspeed will set it for each instance automatically
|
|
source /opt/intel/oneccl/env/setvars.sh
|
|
export WORLD_SIZE=2 # run 1 instance per SPR socket, thus 2 instances on 2 sockets, 96 cores
|
|
export MASTER_ADDR=127.0.0.1
|
|
export CCL_ZE_IPC_EXCHANGE=sockets
|
|
export DS_ACCELERATOR="cpu"
|
|
export CCL_WORKER_AFFINITY=auto
|
|
unset KMP_AFFINITY # deepspeed will set it for each instance automatically
|
|
export FI_PROVIDER=tcp
|
|
export CCL_ATL_TRANSPORT=ofi
|
|
export CCL_PROCESS_LAUNCHER=none
|
|
|
|
deepspeed \
|
|
--bind_cores_to_rank \
|
|
--bind_core_list 0-95 \
|
|
deepspeed_autotp.py
|