diff --git a/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh b/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh index dab0f584a9..daa1448754 100755 --- a/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh +++ b/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh @@ -7,6 +7,7 @@ export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000 export DEBUG=${DEBUG:-0} export FLASH_ATTENTION=${FLASH_ATTENTION:-1} +export ALL2ALL=${ALL2ALL:-1} export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16" export DP=8 BS=8 EVAL_BS=8 GRADIENT_ACC_STEPS=1