playground/run_h200.sh

#!/bin/bash
set -ex

# curl -fsSL https://claude.ai/install.sh | bash

# Activate conda env with torch 2.6.0 + flash-attn 2.7.4 (compatible)
source /mlx/users/jiashuo.fan/miniconda3/etc/profile.d/conda.sh
conda activate abbie

export DISABLE_ZB=1
export USE_DETERMINISTIC_BACKWARD=0
export TORCH_NCCL_ENABLE_MONITORING=0
export WANDB_MODE=disabled
export WANDB_DISABLED=true
export UCX_ERROR_SIGNALS=""
export NCCL_IB_DISABLE=0
export UCX_TLS=rc,sm,self

# pip3 install byted-thoth-arpeggio==0.2.0c6

# git clone -b feat/qwen3-vl-lbh --single-branch git@code.byted.org:Thoth/Abbie.git
# unzip /mlx/users/jiashuo.fan/playground/Abbie-feat_qwen3-vl-lbh-sft.zip -d /opt/tiger/Abbie
cd /opt/tiger/Abbie
# conda activate
DATA_PATHS='[
  ["/mnt/hdfs/byte_tt_data_cu_vagcp/haogeng.liu/new_policy7w_v2_reformat","100%"]
]'
DATA_PATHS=$(echo "$DATA_PATHS" | tr -d ' \n\t')

# hdfs dfs -get hdfs://harunava/home/byte_tt_data_cu_vagcp/libohan.1024/models/TiViLa-Qwen3-VL-8B-Instruct-Compliance /home/tiger/.cache/
MODEL_PATH=/home/tiger/.cache/TiViLa-Qwen3-VL-8B-Instruct-Compliance
PROJNAME=qwen3_vl_cpt
EXPNAME=new_policy7w_v2_reformat
OUTDIR=/mnt/bn/bohanzhainas1/jiashuo/exp/new_policy7w_v2_reformat_h200

# ls /mnt/bn/bohanzhainas1/jiashuo/viz

MASTER_PORT=$(python - <<'EOF'
import socket
s = socket.socket()
s.bind(('', 0))
print(s.getsockname()[1])
s.close()
EOF
)

echo "Using port: $MASTER_PORT"

LOG_DIR=/mlx/users/jiashuo.fan/playground/.claude/run_$(date +%Y%m%d_%H%M%S)
mkdir -p "$LOG_DIR"
echo "Logs: $LOG_DIR"

torchrun \
    --nproc_per_node=8 \
    --master_port=$MASTER_PORT \
    --log-dir="$LOG_DIR" \
tivila_trainer.py \
    model.pretrained_path=${MODEL_PATH} \
    optim.lr="1e-5" \
    optim.visual_lr="1e-5" \
    optim.lr_warmup_steps_ratio=0.03 \
    trainer.output_path=${OUTDIR} \
    trainer.project_name=${PROJNAME} \
    trainer.experiment_name=${EXPNAME} \
    trainer.log_interval=10 \
    trainer.checkpoint_interval=850 \
    trainer.checkpoint_hf_model=true \
    data.patterns=${DATA_PATHS} \
    data.transform_extra_kwargs.image_max_pixels=16777216 \
    data.transform_extra_kwargs.video_max_pixels=307200 \
    data.transform_extra_kwargs.video_max_frames=32 \
    data.max_seq_len=8192 \
    data.is_continuous_batch=true \
    data.num_training_steps=850 \
    data.chunks_per_step=4 \
    model.recompute_attn=true \
    model.activation_offloading=false \
    model.visual_activation_offloading=true

# torchrun \
#     --nproc_per_node=${ARNOLD_WORKER_GPU} \
#     --nnodes=${ARNOLD_WORKER_NUM} \
#     --node_rank=${ARNOLD_ID} \
#     --master_addr=${METIS_WORKER_0_HOST} \
#     --master_port=${METIS_WORKER_0_PORT} \
# tivila_trainer.py \
#     model.pretrained_path=${MODEL_PATH} \
#     optim.lr="1e-5" \
#     optim.visual_lr="1e-5" \
#     optim.lr_warmup_steps_ratio=0.03 \
#     trainer.output_path=${OUTDIR} \
#     trainer.project_name=${PROJNAME} \
#     trainer.experiment_name=${EXPNAME} \
#     trainer.log_interval=10 \
#     trainer.checkpoint_interval=850 \
#     trainer.checkpoint_hf_model=true \
#     data.patterns=${DATA_PATHS} \
#     data.transform_extra_kwargs.image_max_pixels=16777216 \
#     data.transform_extra_kwargs.video_max_pixels=307200 \
#     data.transform_extra_kwargs.video_max_frames=32 \
#     data.max_seq_len=8192 \
#     data.is_continuous_batch=true \
#     data.num_training_steps=850 \
#     data.chunks_per_step=4 \
#     model.recompute_attn=true \
#     model.activation_offloading=false \
#     model.visual_activation_offloading=true