From 8b5b9a0e90c0fb86a1a7b2da2f7862769a2fecd7 Mon Sep 17 00:00:00 2001
From: wozeparrot <wozeparrot@gmail.com>
Date: Wed, 1 Apr 2026 06:46:16 +0800
Subject: [PATCH] llama: run_and_time (#15533)

---
 .../tinybox_8xMI350X/run_and_time.sh          | 48 +++++++++++++++++++
 1 file changed, 48 insertions(+)
 create mode 100755 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/run_and_time.sh

diff --git a/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/run_and_time.sh b/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/run_and_time.sh
new file mode 100755
index 0000000000..8b05f248f9
--- /dev/null
+++ b/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/run_and_time.sh
@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="."
+export DEV=AMD
+export EMULATE="AMD_CDNA4"
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export HK_FLASH_ATTENTION=1
+export ALL2ALL=1
+export USE_ATOMICS=1
+export ASM_GEMM=1
+export WQKV=0
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=8 MP=1 BS=8 EVAL_BS=8 GRADIENT_ACC_STEPS=4
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=8B
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=8192
+export MASTER_WEIGHTS=1
+
+export SEED=$RANDOM
+export DATA_SEED=$SEED
+
+export JITBEAM=3
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+export LOGMLPERF=1
+
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="llama31_8b_8xMI350x_${DATETIME}_${SEED}.log"
+
+# beam
+FAKEDATA=1 BENCHMARK=10 INITMLPERF=1 LLAMA_LAYERS=2 python3 examples/mlperf/model_train.py | tee "$LOGFILE"
+
+# run
+RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a "$LOGFILE"