How to accelerate my corpus embedding to the chromadb

I have the corpus.jsonl which has 6.5gb storage.And i use the one h100 gpu to embedding the corpus to the chromadb,but it seems very slowly.I want to find how can i accelerate the progress(gpu,cpu,io).Here is my raw code:

import os
import torch
import chromadb
from tqdm import tqdm
import json
from sentence_transformers import SentenceTransformer

# ===================== 配置 =====================
# 输入：直接使用第一阶段生成的预处理文件
PREPROCESSED_FILE = "./preprocessed_chunks.jsonl"
EMBED_MODEL_PATH = "./minicpm"  # 你的 MiniCPM 模型路径
CHROMA_DIR = "./chroma"  # 持久化路径
COLLECTION_NAME = "my_rag"
BATCH_SIZE = 2048
NORMALIZE = True


# ===================== 主程序：建立索引 =====================
def build_chroma_index():
    embedder = SentenceTransformer(EMBED_MODEL_PATH, trust_remote_code=True)
    client = chromadb.PersistentClient(path=CHROMA_DIR)
    collection = client.get_or_create_collection(COLLECTION_NAME)

    # --- 单GPU版本提示 ---
    if torch.cuda.is_available():
        print("检测到可用的 GPU。将使用默认的单个 GPU (cuda:0) 进行嵌入。")
    else:
        print("警告：未检测到可用的 GPU。将使用 CPU 进行嵌入，速度会较慢。")
    # ---

    all_chunks = []
    all_ids = []
    total = 0

    print(f"📖 从预处理文件 {PREPROCESSED_FILE} 读取文本块...")

    # 主进程的工作就是逐行读取预处理好的数据
    with open(PREPROCESSED_FILE, "r", encoding="utf-8") as f:
        for line in tqdm(f, desc="构建向量索引"):
            chunk_text = json.loads(line)["text"]

            all_chunks.append(chunk_text)
            all_ids.append(f"chunk-{total}")
            total += 1

            # 当累积的文本块数量达到 BATCH_SIZE 时，进行一次批处理
            if len(all_chunks) >= BATCH_SIZE:
                # 直接调用 encode 函数，无需 pool
                # show_progress_bar=True 可以在处理大批次时显示一个内部进度条
                vecs = embedder.encode(
                    all_chunks,
                    normalize_embeddings=NORMALIZE,
                    batch_size=512,  # 这是模型内部计算时使用的批大小，对于GPU很重要
                    show_progress_bar=True
                ).tolist()

                collection.add(documents=all_chunks, embeddings=vecs, ids=all_ids)
                all_chunks.clear()
                all_ids.clear()

    # 处理最后一批不足 BATCH_SIZE 的剩余文本块
    if all_chunks:
        vecs = embedder.encode(
            all_chunks,
            normalize_embeddings=NORMALIZE,
            batch_size=512,
            show_progress_bar=True
        ).tolist()
        collection.add(documents=all_chunks, embeddings=vecs, ids=all_ids)

    print(f"✅ 索引建立完成。总共处理了 {total} 个文本块。")


# (查询和命令行接口部分与之前完全相同，无需改动)
# ===================== 查询 =====================
def search(query: str, k: int = 5):
    embedder = SentenceTransformer(EMBED_MODEL_PATH, trust_remote_code=True)
    client = chromadb.PersistentClient(path=CHROMA_DIR)
    collection = client.get_or_create_collection(COLLECTION_NAME)

    q_vec = embedder.encode([query], normalize_embeddings=NORMALIZE).tolist()
    results = collection.query(query_embeddings=q_vec, n_results=k)

    print("🔍 Top-{} results for: {}".format(k, query))
    for i, doc in enumerate(results['documents'][0]):
        clean_doc = doc[:100].replace('\n', ' ').replace('\r', '')
        print("[{}] {}...".format(i + 1, clean_doc))


# ===================== CLI =====================
if __name__ == "__main__":
    import argparse

    parser = argparse.ArgumentParser("MiniCPM + ChromaDB RAG Pipeline (Python 3.10 Compatible)")
    sub = parser.add_subparsers(dest="cmd", required=True)

    b = sub.add_parser("build", help="构建 Chroma 向量库")
    s = sub.add_parser("search", help="搜索")
    s.add_argument("--query", required=True)
    s.add_argument("--topk", type=int, default=5)

    args = parser.parse_args()
    if args.cmd == "build":
        build_chroma_index()
    elif args.cmd == "search":
        search(args.query, args.topk)

And below is my gpu、cpu、memory usage:

this is my raw corpus analysis: What is the limitation of my whole progress?Should i use the more gpus or Does the chromadb write too slow or Does the progress reads the data too slow and so on.