From 111f8b1d04fe0efb8193211514b39eeb64903673 Mon Sep 17 00:00:00 2001
From: odashi <yusuke.oda@predicate.jp>
Date: Sat, 26 Oct 2024 21:00:54 +0900
Subject: [PATCH 1/2] add converter script

---
 .../v3-13b-exp4-cpt-lr-sakura/run_convert.sh  | 36 +++++++++++++++++++
 1 file changed, 36 insertions(+)
 create mode 100644 pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/run_convert.sh

diff --git a/pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/run_convert.sh b/pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/run_convert.sh
new file mode 100644
index 0000000..e571792
--- /dev/null
+++ b/pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/run_convert.sh
@@ -0,0 +1,36 @@
+#!/bin/bash
+
+src_root=/home/shared/experiments/0066_v3-13b-cpt-lr/checkpoints
+dest_root=/home/shared/experiments/0066_v3-13b-cpt-lr/checkpoints_hf
+
+for src_exp_dir in ${src_root}/exp*; do
+    exp_rel=$(basename ${src_exp_dir})
+
+    for src_ckpt_dir in ${src_exp_dir}/iter_???????; do
+        ckpt_rel=${exp_rel}/$(basename ${src_ckpt_dir})
+        dest_ckpt_dir=${dest_root}/${ckpt_rel}
+
+        if [ -e ${dest_ckpt_dir} ]; then
+            echo "Exists: ${ckpt_rel}"
+            continue
+        fi
+
+        mkdir -p ${dest_ckpt_dir}
+
+        sbatch \
+            --job-name=0066_convert \
+            --partition=gpu-small \
+            --priority=1 \
+            scripts/pretrain/scripts/v3-converter/convert.sh \
+            ${src_ckpt_dir} \
+            ${dest_ckpt_dir}
+        sbatch_result=$?
+
+        if [ ${sbatch_result} -eq 0 ]; then
+            echo "Queued: ${ckpt_rel}"
+        else
+            echo "Error: ${ckpt_rel}"
+            rmdir ${dest_ckpt_dir}
+        fi
+    done
+done

From 4dc08919631b8bd32328b3305fb7ca3905b33676 Mon Sep 17 00:00:00 2001
From: odashi <yusuke.oda@predicate.jp>
Date: Sat, 26 Oct 2024 21:17:58 +0900
Subject: [PATCH 2/2] fix script

---
 .../v3-13b-exp4-cpt-lr-sakura/convert.sh      | 72 +++++++++++++++++++
 .../v3-13b-exp4-cpt-lr-sakura/run_convert.sh  |  3 +-
 2 files changed, 74 insertions(+), 1 deletion(-)
 create mode 100644 pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/convert.sh

diff --git a/pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/convert.sh b/pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/convert.sh
new file mode 100644
index 0000000..b683a65
--- /dev/null
+++ b/pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/convert.sh
@@ -0,0 +1,72 @@
+#!/bin/bash
+# Model conversion script for converting Megatron format checkpoints into Hugging Face format
+#
+# This script needs one node on the `gpu` partition of the cluster.
+# However, a GPU is necessary to verify CUDA functionality, even though no VRAM will be used.
+#
+# Usage:
+# On a cluster with SLURM:
+#   Run `sbatch --partition {partition} convert.sh SOURCE_DIR TARGET_DIR`
+# On a cluster without SLURM:
+#   Run `bash convert.sh SOURCE_DIR TARGET_DIR TEMPORAL_DIR > outpus/convert.out 2> outputs/convert.err`
+# - SOURCE_DIR: Megatron checkpoint directory including `iter_NNNNNNN`
+# - TARGET_DIR: Output directory for the Hugging Face format
+#
+# Example:
+# sbatch convert.sh /data/experiments/{exp-id}/checkpoints/iter_0001000 /data/experiments/{exp-id}/hf_checkpoints/iter_0001000
+#
+#SBATCH --job-name=ckpt-convert
+#SBATCH --partition=<FIX_ME>
+#SBATCH --nodes=1
+#SBATCH --gpus=1
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=8
+#SBATCH --mem=400G
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.err
+
+set -e
+
+MEGATRON_CHECKPOINT_DIR=${1%/}
+HF_CHECKPOINT_DIR=$2
+
+ENV_DIR=environment_train
+
+source ${ENV_DIR}/scripts/environment.sh
+source ${ENV_DIR}/venv/bin/activate
+
+TOKENIZER_MODEL_DIR=${ENV_DIR}/src/llm-jp-tokenizer/hf/ver3.0/llm-jp-tokenizer-100k.ver3.0b2
+
+TARGET_ITER_DIR=$(basename $MEGATRON_CHECKPOINT_DIR) # iter_NNNNNNN
+ITER=$(( 10#$(echo $TARGET_ITER_DIR | sed 's/^iter_//') )) # NNNNNNN (no 0 padding)
+echo ITER=$ITER
+
+if [[ -z "$ITER" || ! "$ITER" =~ ^[0-9]+$ ]]; then # check if directory is valid
+  >&2 echo "Error: ITER=$ITER is not a valid number. Exiting."
+  exit 1
+fi
+
+# Create a unique temporal working directory to avoid affecting the original directory and
+# to allow multiple runs to execute simultaneously.
+TMP_DIR=$(mktemp -d "${HOME}/ckpt_convert.XXXXXXXX")
+>&2 echo TMP_DIR=$TMP_DIR 
+ln -s $(readlink -f $MEGATRON_CHECKPOINT_DIR) ${TMP_DIR}/${TARGET_ITER_DIR}
+echo $ITER > "${TMP_DIR}/latest_checkpointed_iteration.txt"
+
+echo "Converting $MEGATRON_CHECKPOINT_DIR"
+
+python ${ENV_DIR}/src/Megatron-LM/tools/checkpoint/convert.py \
+  --model-type GPT \
+  --loader mcore \
+  --saver llama2_hf \
+  --load-dir $TMP_DIR \
+  --save-dir $HF_CHECKPOINT_DIR \
+  --hf-tokenizer-path $TOKENIZER_MODEL_DIR \
+  --save-dtype bfloat16 \
+  --loader-transformer-impl "transformer_engine" \
+  --megatron-path ${ENV_DIR}/src/Megatron-LM
+
+cp ${TOKENIZER_MODEL_DIR}/* $HF_CHECKPOINT_DIR
+
+rm -r $TMP_DIR
+echo "Done"
diff --git a/pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/run_convert.sh b/pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/run_convert.sh
index e571792..bd51ff5 100644
--- a/pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/run_convert.sh
+++ b/pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/run_convert.sh
@@ -21,7 +21,8 @@ for src_exp_dir in ${src_root}/exp*; do
             --job-name=0066_convert \
             --partition=gpu-small \
             --priority=1 \
-            scripts/pretrain/scripts/v3-converter/convert.sh \
+            --mem=200G \
+            scripts/pretrain/scripts/v3-13b-exp4-cpt-lr-sakura/convert.sh \
             ${src_ckpt_dir} \
             ${dest_ckpt_dir}
         sbatch_result=$?