From ba90e1b52ebf783dfcf37129b288359ebdd3b3cd Mon Sep 17 00:00:00 2001 From: wozeparrot Date: Tue, 20 Jan 2026 12:44:06 -0800 Subject: [PATCH] feat: script to run llama8b training (#14239) --- .../tinybox_8xMI350X/dev_run.sh | 23 +++++++++++++++++++ 1 file changed, 23 insertions(+) create mode 100755 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh 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 new file mode 100755 index 0000000000..09e6bbfacf --- /dev/null +++ b/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh @@ -0,0 +1,23 @@ +#!/usr/bin/env bash + +export PYTHONPATH="." AMD=1 +export IGNORE_OOB=1 +export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000 + +export DEBUG=${DEBUG:-0} +export FLASH_ATTENTION=1 + +export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16" +export DP=8 BS=8 EVAL_BS=8 + +export MODEL="llama3" +export BASEDIR="/raid/datasets/c4-8b/" +export SMALL=1 LLAMA3_SIZE="8B" +export EVAL_TARGET=3.3 EVAL_FREQ=12288 +export LR="1e-3" END_LR="1e-4" WARMUP_STEPS=1024 MAX_STEPS=1200000 + +export SEED=5760 + +export JITBEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 + +python3 examples/mlperf/model_train.py