マルチ LLM 対応の NVIDIA NIM による合成データ SFT (Seed あり / なし) の効果分析

uv run python examples/converters/convert_dcp_to_hf.py \
  --config="<path/to/checkpoints/config.yaml>" \
  --dcp-ckpt-path="<path/to/checkpoints/step*/policy/weights>" \
  --hf-ckpt-path="$HF_CKPT_PATH"

import argparse
import shutil
import os
from transformers import AutoModelForCausalLM
from huggingface_hub import hf_hub_download


def copy_files_from_hub(original_model_name: str, hf_ckpt_path: str):
    """必要なファイルを HF Hub からコピー"""
    files_to_copy = [
        "modeling_nemotron_h.py",
        "configuration_nemotron_h.py",
        "special_tokens_map.json",
        "tokenizer_config.json",
        "tokenizer.json",
    ]

    os.makedirs(hf_ckpt_path, exist_ok=True)

    for filename in files_to_copy:
        try:
            src_file = hf_hub_download(
                repo_id=original_model_name,
                filename=filename,
            )
            dst_file = os.path.join(hf_ckpt_path, filename)
            shutil.copy(src_file, dst_file)
            print(f"[OK] Copied {filename}")
        except Exception as e:
            print(f"[WARN] Failed to copy {filename}: {e}")


def convert_to_safetensors(hf_ckpt_path: str):
    """HF checkpoint → safetensors に変換"""
    print(f"\n=== Loading model from: {hf_ckpt_path} ===")
    model = AutoModelForCausalLM.from_pretrained(hf_ckpt_path, trust_remote_code=True)

    output_dir = hf_ckpt_path + "_safetensors"
    print(f"=== Saving safetensors to: {output_dir} ===")
    model.save_pretrained(output_dir, safe_serialization=True)

    # tokenizer / tokens を safetensors 側にもコピー
    files_to_copy = [
        "special_tokens_map.json",
        "tokenizer_config.json",
        "tokenizer.json",
    ]
    for filename in files_to_copy:
        src_file = os.path.join(hf_ckpt_path, filename)
        dst_file = os.path.join(output_dir, filename)
        shutil.copy(src_file, dst_file)
        print(f"[OK] Copied {filename} to safetensors directory")

    print("\n=== Conversion completed ===")


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--hf-ckpt-path",
        type=str,
        required=True,
        help="Path to HF checkpoint directory"
    )
    parser.add_argument(
        "--original-model-name",
        type=str,
        default="nvidia/NVIDIA-Nemotron-Nano-9B-v2",
        help="Original model name used for training"
    )

    args = parser.parse_args()

    hf_ckpt_path = args.hf_ckpt_path
    original_model_name = args.original_model_name

    print(f"Using HF checkpoint path: {hf_ckpt_path}")
    print(f"Original model name: {original_model_name}")

    copy_files_from_hub(original_model_name, hf_ckpt_path)
    convert_to_safetensors(hf_ckpt_path)


if __name__ == "__main__":
    main()

uv run python examples/convert_hf_safetensors.py --hf-ckpt-path="$HF_CKPT_PATH"

export CONTAINER_NAME=LLM-NIM
export Repository=nim/nvidia/llm-nim
export TAG=1.15.0
export IMG_NAME="nvcr.io/$Repository:$TAG"
export LOCAL_MODEL_DIR={HuggingFace Safetensors path}
export NIM_MODEL_NAME=$LOCAL_MODEL_DIR
export NIM_SERVED_MODEL_NAME=custom_nemotron_nano_v2_9b

# Choose a path on your system to cache the downloaded models
export LOCAL_NIM_CACHE=~/.cache/nim
mkdir -p "$LOCAL_NIM_CACHE"

# Add write permissions to the NIM cache for downloading model assets
chmod -R a+w "$LOCAL_NIM_CACHE"

# Start the LLM NIM
docker run -it --rm --name=$CONTAINER_NAME \
   --gpus '"device=0"' \
   --shm-size=124GB \
   -e NIM_MODEL_NAME="/opt/models/local_model" \
   -e NIM_SERVED_MODEL_NAME=$NIM_SERVED_MODEL_NAME \
   -v "$LOCAL_MODEL_DIR:/opt/models/local_model" \
   -v "$LOCAL_NIM_CACHE:/opt/nim/.cache" \
   -e NIM_FORCE_TRUST_REMOTE_CODE=1 \
   -e NIM_MANIFEST_ALLOW_UNSAFE=1 \
   -u $(id -u) \
   -e NIM_MAX_NUM_SEQS=256 \
   -e NIM_MAX_BATCH_SIZE=16 \
   -e NIM_MAX_MODEL_LEN=4096 \
   -e NIM_KVCACHE_PERCENT=0.5 \
   -p 8000:8000 \
   $IMG_NAME

datasets:
  - jcommonsenseqa
#  - hle
categories:
  CR:
    description: "nemotron"
    default_metric: exact_match
    metrics: {}
    datasets:
      - jcommonsenseqa

dataset_info_overrides: {}

run_name: null # 最優先で適用されるrun_name。指定しない場合は`wandb.run_name`またはモデル名に実行日時を結合したデフォルト値が用いられる。
output_dir: 'local_files'
eval_dataset_config_path: './eval_configs/datasets_jcommonsense.yaml' # 評価データセットの設定ファイルのパス
include_non_commercial: false # true の場合、商用利用不可のデータセットも含めて評価します。

max_num_samples: -1 # 評価データセットのサンプル数の上限。デフォルトは10件。-1の場合は全件評価します。

exporters:
  local:
    export_output_table: true # 出力結果をテーブル形式で保存するかどうか
    output_top_n: null           # 出力結果の上位何件を保存するか

# HTTP Requestなどによる同期的推論を行う場合の設定
online_inference_config:
  provider: vllm-openai
  max_concurrent: 200 
  hostname: localhost:8000
  model_name: "custom_nemotron_nano_v2_9b"
  generation_config:
    temperature: 0.0

custom_prompt_template: |
  /no_think
  {%- if dataset_instruction -%}
  ### 指示
  {{ dataset_instruction.strip() }}
  {% endif -%}
  {%- if answer_pattern -%}
  ### 回答形式
  {{ answer_pattern.strip() }}
  {% endif -%}
  ### 入力:
  {{ input.strip() }}


default_answer_extract_pattern: "(?s)^(.*?)(?=\\n\\n|\\Z)"

output_length_delta: 0

uv run scripts/evaluate_llm.py eval --config configs/config_sft_jcommonsense.yaml