From 5cc31e23e603268d6e388e1e90677fa92379e088 Mon Sep 17 00:00:00 2001
From: wozeparrot <wozeparrot@gmail.com>
Date: Thu, 9 Jul 2026 23:50:20 -0400
Subject: [PATCH] gptoss: scripts (#16962)

---
 .../tinybox_8xMI350X/dev_beam.sh              | 44 +++++++++++++++++++
 .../tinybox_8xMI350X/dev_run.sh               | 39 ++++++++++++++++
 2 files changed, 83 insertions(+)
 create mode 100755 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_beam.sh
 create mode 100755 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_run.sh

diff --git a/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_beam.sh b/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_beam.sh
new file mode 100755
index 0000000000..aa3c67acfc
--- /dev/null
+++ b/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_beam.sh
@@ -0,0 +1,44 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
+export DEV=${DEV:-AMD}
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-2}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export MXFP8=${MXFP8:-1}
+export ZERO_OPTIM=${ZERO_OPTIM:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="gptoss"
+export BASEDIR="/raid/datasets/c4-8b/"
+export EVAL_TARGET=3.34 EVAL_FREQ=12288
+export END_LR="4e-5" WARMUP_STEPS=128 MAX_STEPS=1200000
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-5760}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
+if [ -z "$FULL_LAYERS" ]; then
+  export LAYERS=${LAYERS:-2}
+fi
+
+python3 examples/mlperf/model_train.py
diff --git a/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_run.sh b/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_run.sh
new file mode 100755
index 0000000000..62daa7ec1e
--- /dev/null
+++ b/examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_run.sh
@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
+export DEV=${DEV:-AMD}
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-0}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export MXFP8=${MXFP8:-1}
+export ZERO_OPTIM=${ZERO_OPTIM:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="gptoss"
+export BASEDIR="/raid/datasets/c4-8b/"
+export EVAL_TARGET=3.34 EVAL_FREQ=12288
+export END_LR="4e-5" WARMUP_STEPS=128 MAX_STEPS=1200000
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-$RANDOM}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+python3 examples/mlperf/model_train.py