From aec1ae0de1aaa290fe4a9db284e80877d389e58d Mon Sep 17 00:00:00 2001 From: nimlgen <138685161+nimlgen@users.noreply.github.com> Date: Thu, 29 Jan 2026 01:40:00 +0300 Subject: [PATCH] llama: set manual_seed (#14409) --- examples/mlperf/model_train.py | 2 ++ .../llama8b/implementations/tinybox_8xMI350X/dev_run.sh | 2 +- 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/examples/mlperf/model_train.py b/examples/mlperf/model_train.py index 98debf7000..5b33e5e880 100644 --- a/examples/mlperf/model_train.py +++ b/examples/mlperf/model_train.py @@ -1321,6 +1321,8 @@ def train_llama3(): opt_base_learning_rate = LR opt_end_learning_rate = END_LR + Tensor.manual_seed(SEED) # seed for weight initialization + # ** init wandb ** WANDB = getenv("WANDB") if WANDB: 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 e2dfbc3cba..5d3492fb58 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 @@ -22,7 +22,7 @@ export LR="2.5e-4" END_LR="2.5e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS)) export SAMPLES=$((MAX_STEPS * GBS)) -export SEED=5760 +export SEED=${SEED:-5760} export JITBEAM=${JITBEAM:-3} export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5