run_sd3_rectified_sampling_old.sh

#!/bin/bash
set -euo pipefail

# 分布式采样：指定 LoRA 与 Rectified(SIT) 权重

export CUDA_VISIBLE_DEVICES="0,1,2,3"
export NCCL_DEBUG=INFO
export NCCL_DEBUG_SUBSYS=ALL
export NCCL_IB_DISABLE=1
export NCCL_P2P_LEVEL=SYS

PRETRAINED_MODEL="/gemini/space/hsd/project/pretrained_model/huggingface/hub/models--stabilityai--stable-diffusion-3-medium-diffusers/snapshots/ea42f8cef0f178587cf766dc8129abd379c90671"
#"/gemini/space/zhaozy/zhy/hsd/project/pretrained_model/models--stabilityai--stable-diffusion-3-medium-diffusers"
LORA_PATH="/gemini/space/gzy_new/models/Sida/sd3-lora-finetuned-batch-4/checkpoint-500000"                # 可为空
RECTIFIED_WEIGHTS="/gemini/space/gzy_new/models/Sida/rectified-noise-batch-2/checkpoint-120000"  # 可为空（若不用 Rectified）

CAPTIONS_JSONL="/gemini/space/hsd/project/dataset/cc3m-wds/validation/metadata.jsonl"
SAMPLE_DIR="./sd3_rectified_samples_new_batch_2"

NUM_INFERENCE_STEPS=40
GUIDANCE_SCALE=7.0
HEIGHT=512
WIDTH=512
PER_PROC_BATCH_SIZE=32
IMAGES_PER_CAPTION=3
MAX_SAMPLES=30000
GLOBAL_SEED=42
MIXED_PRECISION="fp16"  # no / fp16 / bf16
NUM_SIT_LAYERS=1      # 需与训练一致

ARGS=(
  --pretrained_model_name_or_path "$PRETRAINED_MODEL"
  --captions_jsonl "$CAPTIONS_JSONL"
  --sample_dir "$SAMPLE_DIR"
  --num_inference_steps $NUM_INFERENCE_STEPS
  --guidance_scale $GUIDANCE_SCALE
  --height $HEIGHT
  --width $WIDTH
  --per_proc_batch_size $PER_PROC_BATCH_SIZE
  --images_per_caption $IMAGES_PER_CAPTION
  --max_samples $MAX_SAMPLES
  --global_seed $GLOBAL_SEED
  --num_sit_layers $NUM_SIT_LAYERS
  --mixed_precision $MIXED_PRECISION
)

if [ -n "$LORA_PATH" ]; then
  ARGS+=(--lora_path "$LORA_PATH")
fi

if [ -n "$RECTIFIED_WEIGHTS" ]; then
  ARGS+=(--rectified_weights "$RECTIFIED_WEIGHTS")
fi

echo "[run_sd3_rectified_sampling.sh] start torchrun: $(date)"

# 先尝试 4 卡模式，如果失败则退到单卡模式
if ! torchrun --nproc_per_node=4 --master_port=25913 sample_sd3_rectified_ddp.py "${ARGS[@]}"; then
  ret=$?
  echo "[run_sd3_rectified_sampling.sh] 4卡运行失败(退出码 ${ret})，尝试单卡模式"
  if ! torchrun --nproc_per_node=1 --master_port=25913 sample_sd3_rectified_ddp.py "${ARGS[@]}"; then
    ret2=$?
    echo "[run_sd3_rectified_sampling.sh] 单卡运行也失败(退出码 ${ret2})，请查看具体错误信息。"
    exit $ret2
  fi
  echo "[run_sd3_rectified_sampling.sh] 单卡运行成功，建议降低 per_proc_batch_size 或使用单卡配置继续。"
fi

wait

echo "Sampling done. Output at: $SAMPLE_DIR"
# nohup bash run_sd3_rectified_sampling.sh > run_sd3_rectified_sampling.log 2>&1 &