#!/usr/bin/env bash
set -euo pipefail

echo "==> Show Python / Torch / CUDA"
python - <<'PY'
import sys, torch
print("Python:", sys.version.split()[0])
print("Torch:", torch.__version__)
print("Torch CUDA tag:", torch.version.cuda)
print("CUDA available:", torch.cuda.is_available())
PY

# 选择 CUDA 工具链：优先 /usr/local/cuda-11.8，其次现有的 /usr/local/cuda-12.8
if [ -d /usr/local/cuda-11.8 ]; then
  export CUDA_HOME=/usr/local/cuda-11.8
  echo "==> Using CUDA_HOME=${CUDA_HOME} (preferred for cu118)"
elif [ -d /usr/local/cuda-12.8 ]; then
  export CUDA_HOME=/usr/local/cuda-12.8
  echo "==> Using CUDA_HOME=${CUDA_HOME} (nvcc 12.8)"
else
  echo "!! 未找到 /usr/local/cuda-11.8 或 /usr/local/cuda-12.8，请安装 CUDA toolkit (dev)。"
  exit 1
fi
export PATH="${CUDA_HOME}/bin:${PATH}"

echo "==> nvcc version"
nvcc --version || { echo "nvcc not found via CUDA_HOME=${CUDA_HOME}"; exit 1; }

# 3090 = sm_86
export TORCH_CUDA_ARCH_LIST="8.6"

echo "==> Install build deps via mamba"
mamba install -y -c conda-forge cmake ninja pybind11 libaio git

echo "==> Upgrade Python build tools"
pip install -U pip setuptools wheel

# 固定到较稳的 DeepSpeed tag；需要最新版可改为 --branch master
DS_TAG="v0.14.3"
if [ ! -d DeepSpeed ]; then
  git clone --branch ${DS_TAG} https://github.com/microsoft/DeepSpeed.git
fi
cd DeepSpeed

echo "==> Build & Install DeepSpeed (training kernels only)"
export DS_BUILD_OPS=1
export DS_BUILD_AIO=1
export DS_BUILD_FUSED_ADAM=1
export DS_BUILD_CPU_ADAM=1
# 不启用推理/transformer内核，降低不必要的编译/兼容风险
# export DS_BUILD_TRANSFORMER=1

# 某些环境需要强制找 CUDA：export DS_FORCE_CUDA=1
# export DS_FORCE_CUDA=1

pip install .

echo "==> Verify DeepSpeed env"
python -m deepspeed.env_report

echo "==> Smoke test: import FusedAdam"
python - <<'PY'
import deepspeed, torch
from deepspeed.ops.adam import FusedAdam
print("DeepSpeed:", deepspeed.__version__)
print("Torch:", torch.__version__, "CUDA tag:", torch.version.cuda)
print("GPU:", torch.cuda.get_device_name(0) if torch.cuda.is_available() else None)
print("FusedAdam OK")
PY

echo "==> Create minimal HF Trainer single-GPU test (tiny model)"
cd ..
cat > ds_config_stage2_single.json <<'JSON'
{
  "train_batch_size": 8,
  "train_micro_batch_size_per_gpu": 1,
  "gradient_accumulation_steps": 8,
  "zero_optimization": { "stage": 2, "overlap_comm": true, "contiguous_gradients": true },
  "fp16": { "enabled": true },
  "aio": {
    "block_size": 1048576,
    "queue_depth": 16,
    "thread_count": 1,
    "single_submit": false,
    "overlap_events": true,
    "verbose": false
  },
  "gradient_clipping": 1.0,
  "steps_per_print": 1000,
  "wall_clock_breakdown": false
}
JSON

cat > train_single_gpu_min.py <<'PY'
from datasets import Dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments, DataCollatorForLanguageModeling
import torch

model_name = "sshleifer/tiny-gpt2"   # 极小模型，快速验证
tok = AutoTokenizer.from_pretrained(model_name)
if tok.pad_token is None:
    tok.pad_token = tok.eos_token

data = ["hello world", "deepspeed single gpu", "trainer test", "fast check"] * 200
ds = Dataset.from_dict({"text": data})
def enc(e): return tok(e["text"], truncation=True, max_length=128)
ds = ds.map(enc)
collator = DataCollatorForLanguageModeling(tok, mlm=False)

model = AutoModelForCausalLM.from_pretrained(model_name)
args = TrainingArguments(
    output_dir="out-ds-single",
    per_device_train_batch_size=1,
    gradient_accumulation_steps=8,
    learning_rate=5e-4,
    num_train_epochs=1,
    logging_steps=10,
    save_steps=0,
    fp16=True,
    deepspeed="ds_config_stage2_single.json"
)
trainer = Trainer(model=model, args=args, tokenizer=tok,
                  train_dataset=ds, data_collator=collator)
trainer.train()
print("OK: single-GPU training finished.")
PY

echo "==> Run minimal single-GPU training with DeepSpeed Stage-2"
CUDA_VISIBLE_DEVICES=0 python train_single_gpu_min.py

echo "==> ALL DONE (single-GPU)."