19 lines
670 B
Bash
19 lines
670 B
Bash
source /opt/intel/oneapi/setvars.sh
|
|
export no_proxy=localhost
|
|
export FI_PROVIDER=tcp
|
|
export OMP_NUM_THREADS=32
|
|
|
|
export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so
|
|
basekit_root=/opt/intel/oneapi
|
|
source $basekit_root/setvars.sh --force
|
|
source $basekit_root/ccl/latest/env/vars.sh --force
|
|
|
|
export USE_XETLA=OFF
|
|
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=2
|
|
export TORCH_LLM_ALLREDUCE=0
|
|
|
|
export MODEL_PATH=YOUR_MODEL_PATH
|
|
export NUM_GPUS=2
|
|
export BIGDL_QUANTIZE_KV_CACHE=1
|
|
|
|
CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS pipeline_serving.py --repo-id-or-model-path $MODEL_PATH --low-bit fp8 --max-num-seqs 4
|