Jack Taylor
ebcac26530
* Add gpt-fast pytorch all reduce benchmark script
* Update readme instructions
* Minor changes
[ROCm/rccl commit: 5f2b88bc28]
14 строки
546 B
Bash
14 строки
546 B
Bash
#!/bin/bash
|
|
|
|
SEQUENCE_LENGTHS=(50 128 256 512 1024 2048 4096)
|
|
ALL_REDUCE_ALGOS=(1 2 3 4)
|
|
|
|
HIP_DEV_FORCE_KERNARG=1
|
|
|
|
for SEQ_LEN in "${SEQUENCE_LENGTHS[@]}"; do
|
|
for ALGO in "${ALL_REDUCE_ALGOS[@]}"; do
|
|
echo "Running sequence length $SEQ_LEN with intra-node all_reduce $ALGO"
|
|
ENABLE_INTRA_NODE_COMM=1 rocprofv3 --hip-trace --kernel-trace --stats --output-format PFTRACE -d rocprof_trace/intranode_input"$SEQ_LEN"_allreduce"$ALGO"/ -- torchrun --nproc_per_node=8 all_reduce.py --sequence_lengths $SEQ_LEN --all_reduce $ALGO --tracing
|
|
done
|
|
done
|