dma-cli 0.1.0__py3-none-any.whl

dma_cli/submit.py

@@ -0,0 +1,673 @@
+"""submit: 7 步链路 + 单个/批量并发 + 统计 + check_machine_type 预检。
+
+7 步链路（移植自 dma_job_submit_10000.jmx 的权威请求体，HAR 同构）：
+  1. POST /irpapi/szc/v1/job/                          body {"client":3,"createSource":2}      -> data.jobId
+  2. GET  /irpapi/szc/v1/job/{jobId}/input/token                                               -> data.token
+  3. POST /dmaapi/szc/v1/file/upload/multipart/init     X-Storage-Token; 网页 attachment 体   -> data.initialKey
+  4. POST /dmaapi/szc/v1/file/upload/multipart/upload   X-Storage-Token+X-Storage-Param; 原文 -> data.partString
+  5. POST /dmaapi/szc/v1/file/upload/multipart/complete X-Storage-Token; {initialKey,partString[]}
+  6. POST /irpapi/szc/v1/job/{jobId}/snapshot           HAR 同构体（当前选择 + 覆盖）
+  7. POST /irpapi/szc/v1/job/commit/{jobId}             Content-Type: x-www-form-urlencoded; 空 body
+每步断言 code==0，失败记录「哪步 + code/msg」。
+"""
+import base64
+import json
+import os
+import sys
+import time
+from collections import Counter
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+from typing import Optional
+
+import requests
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+
+from .client import DmaClient, STATUS_MAP
+
+# 内置默认任务脚本（无 --input_directory 时上传它）
+DEFAULT_INPUT = Path(__file__).parent / "calc.py"
+
+
+def find_sku_by_machine_type(client: DmaClient, machine_type: str) -> dict:
+    """根据机型名（skuName）查找 sku 信息。返回 {skuId, skuName, cpu, memory, gpuModel, gpuNum}。"""
+    for kind in ["cpu", "gpu"]:
+        resp = client.get("/dmaapi/szc/v1/softwares/skus",
+                         params={"kind": kind, "create_source": 2})
+        if resp.get("code") != 0:
+            continue
+        for item in resp.get("data", {}).get("items", []):
+            if item.get("skuName") == machine_type:
+                return {
+                    "skuId": item["skuId"],
+                    "skuName": item["skuName"],
+                    "cpu": item["cpu"],
+                    "memory": item["memory"],
+                    "gpuModel": item["gpuModel"],
+                    "gpuNum": item["gpuNum"],
+                }
+    raise RuntimeError(f"未找到机型 {machine_type}，可用 'dma sku list --kind cpu|gpu' 查询")
+
+
+def find_software_by_image_address(client: DmaClient, image_address: str) -> dict:
+    """根据镜像地址查找软件信息。返回 {softwareId, softwareVersionId, default_cmd}。
+    匹配顺序：
+    1. 版本的 url 字段精确匹配
+    2. softwareVersion 字段精确匹配
+    3. softwareName 模糊匹配
+    """
+    resp = client.get("/dmaapi/szc/v1/softwares/deep-mining", params={"bizPlatform": 2})
+    if resp.get("code") != 0:
+        raise RuntimeError(f"查询软件目录失败: code={resp.get('code')}")
+
+    items = resp.get("data", {}).get("items", [])
+
+    # 1. 先尝试按 url 精确匹配
+    for item in items:
+        for version in item.get("versions", []):
+            if version.get("url") == image_address:
+                cmds = version.get("cmds", [])
+                default_cmd = cmds[0]["cmd"] if cmds else None
+                return {
+                    "softwareId": item["softwareId"],
+                    "softwareVersionId": version["softwareVersionId"],
+                    "softwareName": item["softwareName"],
+                    "softwareVersion": version["softwareVersion"],
+                    "default_cmd": default_cmd,
+                }
+
+    # 2. 尝试按 softwareVersion 精确匹配
+    for item in items:
+        for version in item.get("versions", []):
+            if version.get("softwareVersion") == image_address:
+                cmds = version.get("cmds", [])
+                default_cmd = cmds[0]["cmd"] if cmds else None
+                return {
+                    "softwareId": item["softwareId"],
+                    "softwareVersionId": version["softwareVersionId"],
+                    "softwareName": item["softwareName"],
+                    "softwareVersion": version["softwareVersion"],
+                    "default_cmd": default_cmd,
+                }
+
+    # 3. 再尝试按 softwareName 模糊匹配
+    for item in items:
+        if image_address in item.get("softwareName", "") or item.get("softwareName") in image_address:
+            version = item.get("versions", [{}])[0]
+            cmds = version.get("cmds", [])
+            default_cmd = cmds[0]["cmd"] if cmds else None
+            return {
+                "softwareId": item["softwareId"],
+                "softwareVersionId": version.get("softwareVersionId"),
+                "softwareName": item["softwareName"],
+                "softwareVersion": version.get("softwareVersion"),
+                "default_cmd": default_cmd,
+            }
+
+    raise RuntimeError(f"未找到镜像 {image_address}，可用 'dma image list' 查询")
+
+
+def load_json_config(config_file: str) -> dict:
+    """加载 JSON 配置文件。"""
+    if not os.path.exists(config_file):
+        raise RuntimeError(f"配置文件不存在: {config_file}")
+    with open(config_file, "r", encoding="utf-8") as f:
+        data = json.load(f)
+        # 如果是数组形式（批量任务），外部做处理，此处包装成 dict 格式避免 attribute error
+        if isinstance(data, list):
+            return {"_is_batch_list": True, "_batch_data": data}
+        return data
+
+
+def check_machine_type(client, sku_name, n_node=1) -> dict:
+    """预检机型合法性，返回完整响应（data.type 如 QUICK）。"""
+    return client.get("/irpapi/szc/v1/job/check_machine_type",
+                      params={"machineType": sku_name, "from": "web", "nNode": n_node})
+
+
+def build_snapshot_body(client, job_id, user_id, *, name, cmd, overrides: dict) -> dict:
+    """用当前项目 + 覆盖参数拼 HAR 同构 snapshot 体。"""
+    project = client.current_project()
+    defaults = client.defaults
+
+    def ov(key, current):
+        v = overrides.get(key)
+        return v if v is not None else current
+
+    project_id = ov("project_id", project and project["projectId"])
+    if not project_id:
+        raise RuntimeError("未设置项目，请先 'dma project use <projectId>' 或指定 --project_id")
+
+    # 从 overrides 中获取 sku 和 software 信息（已通过自动查找填充）
+    sku_id = ov("sku_id", None)
+    cpu = ov("cpu", None)
+    memory = ov("memory", None)
+    gpu_model = ov("gpu_model", None)
+    gpu_num = ov("gpu_num", None)
+
+    if not all([sku_id, cpu is not None, memory is not None]):
+        raise RuntimeError("未设置机型，请指定 --machine_type 或 --sku-id")
+
+    software_id = ov("software_id", None)
+    software_version_id = ov("software_version_id", None)
+
+    if not all([software_id, software_version_id]):
+        raise RuntimeError("未设置镜像，请指定 --image_address 或 --software-id")
+
+    # 处理 max_run_time（分钟 → maxRunTime + maxRunTimeUnit）
+    max_run_time_min = ov("max_run_time", 0)
+    if max_run_time_min and max_run_time_min > 0:
+        max_run_time = max_run_time_min
+        max_run_time_unit = "m"
+    else:
+        max_run_time = defaults.get("max_run_time", 15)
+        max_run_time_unit = defaults.get("max_run_time_unit", "d")
+
+    # 处理 log_files 和 out_files
+    log_files = []
+    if overrides.get("log_file"):
+        log_files = [overrides["log_file"]]
+
+    out_files = []
+    if overrides.get("out_files"):
+        out_files.extend(overrides["out_files"])
+    if overrides.get("result_path"):
+        out_files.append(overrides["result_path"])
+    out_files = list(set(out_files))  # 去重
+
+    return {
+        "name": name,
+        "projectId": project_id,
+        "bohrJobId": int(job_id),
+        "inputs": [f"{user_id}/job/{job_id}/input/"],
+        "softwareId": software_id,
+        "softwareVersionId": software_version_id,
+        "cmd": cmd,
+        "compatibilityUpgrade": {"enabled": False},
+        "nnode": ov("nnode", 1),
+        "restartPolicy": 0,
+        "restartDelaySeconds": 20,
+        "maxRescheduleTimes": ov("max_reschedule_times", 0),
+        "logFiles": log_files,
+        "outFiles": out_files,
+        "skuId": sku_id,
+        "cpu": cpu,
+        "gpuModel": gpu_model or "cpu",
+        "gpuNum": gpu_num or 0,
+        "memory": memory,
+        "maxRunTime": max_run_time,
+        "maxRunTimeUnit": max_run_time_unit,
+        "inputFileList": [],
+    }
+
+
+def submit_one(client: DmaClient, input_files: list, *, name: str, cmd: str, overrides: dict, check: bool = True) -> dict:
+    """提交单个任务，走完整 7 步。input_files 是 [(filename, bytes), ...] 列表。返回结果 dict。"""
+    start = time.time()
+    user_id = client._user_id()
+    result = {"ok": False, "job_id": None, "bohr_job_id": None, "error": None, "elapsed_ms": 0}
+
+    def fail(step, resp):
+        result["error"] = f"{step}: code={resp.get('code')} msg={resp.get('msg') or resp.get('message')}"
+        result["elapsed_ms"] = int((time.time() - start) * 1000)
+        return result
+
+    try:
+        # 预检（可选）
+        if check:
+            sku_name = overrides.get("sku_name")
+            if sku_name:
+                try:
+                    resp = check_machine_type(client, sku_name, overrides.get("nnode", 1))
+                    if resp.get("code") != 0 or not (resp.get("data") or {}).get("type"):
+                        return fail("check_machine_type", resp)
+                except Exception as e:
+                    print(f"  ⚠️ 预检异常（忽略）: {e}")
+
+        # 步骤1 创建 job
+        resp = client.post_json("/irpapi/szc/v1/job/", {"client": 3, "createSource": 2})
+        if resp.get("code") != 0:
+            return fail("step1 create job", resp)
+        job_id = resp.get("data", {}).get("jobId")
+
+        # 步骤2 获取 input token
+        resp = client.get(f"/irpapi/szc/v1/job/{job_id}/input/token")
+        if resp.get("code") != 0:
+            return fail("step2 get token", resp)
+        storage_token = resp.get("data", {}).get("token")
+
+        # 步骤3-5 上传文件
+        for filename, file_bytes in input_files:
+            # 步骤3 init
+            init_body = {
+                "path": f"{user_id}/job/{job_id}/input/{filename}",
+                "option": {"contentDisposition": f'attachment; filename="{filename}"'}
+            }
+            resp = client.post_json(
+                "/dmaapi/szc/v1/file/upload/multipart/init",
+                init_body,
+                headers={"X-Storage-Token": storage_token}
+            )
+            if resp.get("code") != 0:
+                return fail(f"step3 init ({filename})", resp)
+            initial_key = resp.get("data", {}).get("initialKey")
+
+            # 用 initial_key 构造 X-Storage-Param
+            storage_param = base64.b64encode(
+                json.dumps({"initialKey": initial_key, "number": 1, "partSize": len(file_bytes)}).encode()
+            ).decode()
+
+            # 步骤4 upload
+            resp = client.post_raw(
+                "/dmaapi/szc/v1/file/upload/multipart/upload",
+                file_bytes,
+                headers={
+                    "X-Storage-Token": storage_token,
+                    "X-Storage-Param": storage_param,
+                    "Content-Type": "application/x-www-form-urlencoded",
+                }
+            )
+            if resp.get("code") != 0:
+                return fail(f"step4 upload ({filename})", resp)
+            part_string = resp.get("data", {}).get("partString")
+            if part_string is None:
+                return fail(f"step4 upload (no partString) ({filename})", resp)
+
+            # 步骤5 complete
+            complete_body = {"initialKey": initial_key, "partString": [part_string]}
+            resp = client.post_json(
+                "/dmaapi/szc/v1/file/upload/multipart/complete",
+                complete_body,
+                headers={"X-Storage-Token": storage_token},
+            )
+            if resp.get("code") != 0:
+                return fail(f"step5 complete ({filename})", resp)
+
+        # 步骤6 snapshot
+        snap = build_snapshot_body(client, job_id, user_id, name=name, cmd=cmd, overrides=overrides)
+        resp = client.post_json(f"/irpapi/szc/v1/job/{job_id}/snapshot", snap)
+        if resp.get("code") != 0:
+            return fail("step6 snapshot", resp)
+
+        # 步骤7 commit
+        resp = client.post_empty(f"/irpapi/szc/v1/job/commit/{job_id}",
+                                 headers={"Content-Type": "application/x-www-form-urlencoded"})
+        if resp.get("code") != 0:
+            return fail("step7 commit", resp)
+
+        data = resp.get("data") or {}
+        result["job_id"] = data.get("jobId") or job_id
+        result["bohr_job_id"] = data.get("bohrJobId") or job_id
+        result["ok"] = True
+    except Exception as e:
+        result["error"] = f"exception: {e}"
+    result["elapsed_ms"] = int((time.time() - start) * 1000)
+    return result
+
+
+def submit_batch(client: DmaClient, count: int, concurrency: int, input_files: list,
+                 *, name_prefix: str, cmd: str, overrides: dict, check: bool) -> tuple:
+    """批量并发提交，每线程独立 DmaClient + requests.Session。返回 (results, elapsed_s)。"""
+    start = time.time()
+    results = []
+
+    def task(i):
+        c = DmaClient(http_session=requests.Session())
+        return submit_one(c, input_files,
+                          name=f"{name_prefix}-{i}", cmd=cmd, overrides=overrides, check=check)
+
+    with ThreadPoolExecutor(max_workers=concurrency) as ex:
+        futures = [ex.submit(task, i) for i in range(count)]
+        for fu in as_completed(futures):
+            r = fu.result()
+            results.append(r)
+            mark = "✅" if r["ok"] else "❌"
+            tail = "" if r["ok"] else f" {r['error']}"
+            print(f"  [{len(results)}/{count}] {mark} job={r.get('job_id')} "
+                  f"bohr={r.get('bohr_job_id')} {r['elapsed_ms']}ms{tail}")
+    return results, time.time() - start
+
+
+def _pct(sorted_times, p):
+    if not sorted_times:
+        return 0
+    idx = max(0, min(len(sorted_times) - 1, int(len(sorted_times) * p / 100)))
+    return sorted_times[idx]
+
+
+def print_stats(results, elapsed_s):
+    total = len(results)
+    ok = sum(1 for r in results if r["ok"])
+    fail = total - ok
+    rate = ok / total * 100 if total else 0
+    rps = ok / elapsed_s if elapsed_s else 0
+    times = sorted(r["elapsed_ms"] for r in results if r["ok"])
+    avg = sum(times) / len(times) if times else 0
+    print(f"\n{'=' * 60}")
+    print("批量提交统计")
+    print(f"{'=' * 60}")
+    print(f"总数: {total}  成功: {ok}  失败: {fail}  成功率: {rate:.1f}%")
+    print(f"RPS: {rps:.2f}  总耗时: {elapsed_s:.2f}s")
+    if times:
+        print(f"单任务耗时(ms): avg={avg:.0f} P50={_pct(times, 50)} "
+              f"P95={_pct(times, 95)} P99={_pct(times, 99)}")
+    if fail:
+        steps = Counter()
+        for r in results:
+            if not r["ok"] and r["error"]:
+                steps[r["error"].split(":")[0]] += 1
+        print("\n失败按步骤归类:")
+        for k, v in steps.most_common():
+            print(f"  {k}: {v}")
+    print(f"{'=' * 60}")
+
+
+def poll_task_status(client: DmaClient, job_id: int, timeout: int = 300) -> None:
+    """轮询单个任务的状态直到结束。"""
+    import time
+    start_time = time.time()
+    last_status = None
+
+    while time.time() - start_time < timeout:
+        # 通过任务列表查询指定任务
+        resp = client.get("/irpapi/szc/v1/job/list",
+                         params={"page": 1, "pageSize": 100, "queryType": "all", "createSource": 2})
+        if resp.get("code") == 0:
+            items = resp.get("data", {}).get("items", [])
+            # 查找目标任务
+            job = next((j for j in items if j.get("bohrId") == job_id or j.get("id") == job_id), None)
+            if job:
+                status = job.get("status")
+                status_name = STATUS_MAP.get(status, f"未知({status})")
+
+                if status != last_status:
+                    print(f"  任务状态: {status_name}")
+                    last_status = status
+
+                # 终态: 2=成功, 3=失败, 4=取消, 6=超时
+                if status in [2, 3, 4, 6]:
+                    return
+
+        time.sleep(2)
+
+    print(f"  ⚠️ 轮询超时（{timeout}s），任务仍在运行")
+
+
+def handle_submit(client, args):
+    """处理 submit 命令。"""
+    # 1. 加载配置并合并参数
+    overrides = {}
+
+    # 从 JSON 配置文件读取
+    if args.config_file:
+        json_config = load_json_config(args.config_file)
+        # 如果包装成批量列表，则提取 _batch_data[0] 作为合并参考，避免报错
+        if isinstance(json_config, dict) and json_config.get("_is_batch_list"):
+            json_config = json_config["_batch_data"][0] if json_config["_batch_data"] else {}
+
+        # 映射 JSON 字段到 overrides
+        if json_config.get("project_id"):
+            overrides["project_id"] = json_config["project_id"]
+        if json_config.get("job_name"):
+            args.job_name = json_config["job_name"]
+        if json_config.get("machine_type"):
+            overrides["machine_type"] = json_config["machine_type"]
+        if json_config.get("image_address"):
+            overrides["image_address"] = json_config["image_address"]
+        if json_config.get("command"):
+            args.command = json_config["command"]
+        if json_config.get("input_directory"):
+            args.input_directory = json_config["input_directory"]
+        if json_config.get("max_run_time"):
+            overrides["max_run_time"] = json_config["max_run_time"]
+        if json_config.get("max_reschedule_times"):
+            overrides["max_reschedule_times"] = json_config["max_reschedule_times"]
+        if json_config.get("nnode"):
+            overrides["nnode"] = json_config["nnode"]
+        if json_config.get("log_file"):
+            overrides["log_file"] = json_config["log_file"]
+        if json_config.get("out_files"):
+            overrides["out_files"] = json_config["out_files"]
+        if json_config.get("result_path"):
+            overrides["result_path"] = json_config["result_path"]
+        if json_config.get("dataset_path"):
+            overrides["dataset_path"] = json_config["dataset_path"]
+
+    # CLI 参数覆盖 JSON 配置
+    if args.project_id:
+        overrides["project_id"] = args.project_id
+    if args.machine_type:
+        overrides["machine_type"] = args.machine_type
+    if args.image_address:
+        overrides["image_address"] = args.image_address
+    if args.max_run_time:
+        overrides["max_run_time"] = args.max_run_time
+    if args.max_reschedule_times:
+        overrides["max_reschedule_times"] = args.max_reschedule_times
+    if args.nnode:
+        overrides["nnode"] = args.nnode
+    if args.log_file:
+        overrides["log_file"] = args.log_file
+    if args.out_files:
+        overrides["out_files"] = args.out_files
+    if args.result_path:
+        overrides["result_path"] = args.result_path
+    if args.dataset_path:
+        overrides["dataset_path"] = args.dataset_path
+
+    # 2. 自动查找 machine_type 和 image_address
+    if overrides.get("machine_type"):
+        print(f"🔍 查找机型: {overrides['machine_type']}")
+        sku_info = find_sku_by_machine_type(client, overrides["machine_type"])
+        overrides["sku_id"] = sku_info["skuId"]
+        overrides["sku_name"] = sku_info["skuName"]
+        overrides["cpu"] = sku_info["cpu"]
+        overrides["memory"] = sku_info["memory"]
+        overrides["gpu_model"] = sku_info["gpuModel"]
+        overrides["gpu_num"] = sku_info["gpuNum"]
+        print(f"✅ 找到机型: {sku_info['skuId']} | {sku_info['skuName']} "
+              f"(cpu={sku_info['cpu']} mem={sku_info['memory']} gpu={sku_info['gpuModel']}x{sku_info['gpuNum']})")
+
+    if overrides.get("image_address"):
+        print(f"🔍 查找镜像: {overrides['image_address']}")
+        sw_info = find_software_by_image_address(client, overrides["image_address"])
+        overrides["software_id"] = sw_info["softwareId"]
+        overrides["software_version_id"] = sw_info["softwareVersionId"]
+        print(f"✅ 找到镜像: {sw_info['softwareId']} | {sw_info['softwareName']} "
+              f"-> {sw_info['softwareVersionId']} | {sw_info['softwareVersion']}")
+        # 如果未指定命令，使用默认命令
+        if not args.command and sw_info.get("default_cmd"):
+            args.command = sw_info["default_cmd"]
+            print(f"✅ 使用默认命令: {args.command}")
+
+    # 3. 确定命令
+    cmd = args.command or client.defaults.get("cmd", "python calc.py")
+
+    # 4. 确定任务名
+    name_prefix = args.name_prefix or client.defaults.get("name_prefix", "cli-job")
+    name = args.job_name or f"{name_prefix}-1"
+
+    # 5. 准备输入文件
+    input_dir = args.input_directory
+    input_files = []
+    input_dir_specified = False
+
+    # 检查 JSON 配置文件中是否指定了 input_directory
+    if args.config_file:
+        json_config = load_json_config(args.config_file)
+        if json_config.get("input_directory"):
+            input_dir_specified = True
+
+    # 只有明确指定了 input_directory 才上传目录，否则使用默认文件
+    if input_dir and os.path.isdir(input_dir) and input_dir_specified:
+        # 上传目录下所有文件
+        for filepath in Path(input_dir).iterdir():
+            if filepath.is_file():
+                with open(filepath, "rb") as f:
+                    input_files.append((filepath.name, f.read()))
+        print(f"📁 上传目录 {input_dir}: {len(input_files)} 个文件")
+    else:
+        # 使用默认文件
+        with open(DEFAULT_INPUT, "rb") as f:
+            input_files.append((DEFAULT_INPUT.name, f.read()))
+
+    if not input_files:
+        raise RuntimeError("无输入文件")
+
+    # 6. 确定并发数
+    concurrency = args.concurrency or client.defaults.get("concurrency", 5)
+    check = not args.no_check
+
+    # 7. 处理批量提交
+    if args.batch_file:
+        # 方式1: 从 JSON 数组文件批量提交
+        batch_res = load_json_config(args.batch_file)
+        if isinstance(batch_res, dict) and batch_res.get("_is_batch_list"):
+            batch_configs = batch_res["_batch_data"]
+        else:
+            batch_configs = batch_res
+
+        if not isinstance(batch_configs, list):
+            raise RuntimeError("batch_file 必须是 JSON 数组")
+
+        print(f"🚀 批量提交 {len(batch_configs)} 个任务（来自 JSON 数组），并发数 {concurrency}")
+        results = []
+        start_time = time.time()
+
+        def submit_batch_item(config: dict, idx: int):
+            """提交单个批量任务"""
+            # 合并参数：config 覆盖 overrides
+            item_overrides = overrides.copy()
+            for key in ["project_id", "job_name", "machine_type", "image_address", "command",
+                       "input_directory", "max_run_time", "max_reschedule_times", "nnode",
+                       "log_file", "out_files", "result_path", "dataset_path"]:
+                if config.get(key):
+                    item_overrides[key] = config[key]
+
+            # 自动查找
+            if item_overrides.get("machine_type"):
+                sku_info = find_sku_by_machine_type(client, item_overrides["machine_type"])
+                item_overrides["sku_id"] = sku_info["skuId"]
+                item_overrides["sku_name"] = sku_info["skuName"]
+                item_overrides["cpu"] = sku_info["cpu"]
+                item_overrides["memory"] = sku_info["memory"]
+                item_overrides["gpu_model"] = sku_info["gpuModel"]
+                item_overrides["gpu_num"] = sku_info["gpuNum"]
+
+            if item_overrides.get("image_address"):
+                sw_info = find_software_by_image_address(client, item_overrides["image_address"])
+                item_overrides["software_id"] = sw_info["softwareId"]
+                item_overrides["software_version_id"] = sw_info["softwareVersionId"]
+
+            # 准备输入文件
+            item_input_dir = config.get("input_directory", "./")
+            item_files = []
+            if item_input_dir and os.path.isdir(item_input_dir) and config.get("input_directory"):
+                for filepath in Path(item_input_dir).iterdir():
+                    if filepath.is_file():
+                        with open(filepath, "rb") as f:
+                            item_files.append((filepath.name, f.read()))
+            else:
+                with open(DEFAULT_INPUT, "rb") as f:
+                    item_files.append((DEFAULT_INPUT.name, f.read()))
+
+            # 确定参数
+            item_name = config.get("job_name") or f"{name_prefix}-{idx}"
+            item_cmd = config.get("command") or cmd
+
+            # 提交
+            c = DmaClient(http_session=requests.Session())
+            return submit_one(c, item_files, name=item_name, cmd=item_cmd,
+                            overrides=item_overrides, check=check)
+
+        with ThreadPoolExecutor(max_workers=concurrency) as ex:
+            futures = [ex.submit(submit_batch_item, cfg, i) for i, cfg in enumerate(batch_configs, 1)]
+            for fu in as_completed(futures):
+                r = fu.result()
+                results.append(r)
+                mark = "✅" if r["ok"] else "❌"
+                tail = "" if r["ok"] else f" {r['error']}"
+                print(f"  [{len(results)}/{len(batch_configs)}] {mark} job={r.get('job_id')} ({r['elapsed_ms']}ms){tail}")
+
+        elapsed = time.time() - start_time
+        print_stats(results, elapsed)
+        return
+
+    if args.batch_input_dir:
+        # 方式2: 扫描目录批量提交
+        if not os.path.isdir(args.batch_input_dir):
+            raise RuntimeError(f"目录不存在: {args.batch_input_dir}")
+
+        subdirs = [d for d in Path(args.batch_input_dir).iterdir() if d.is_dir()]
+        if not subdirs:
+            raise RuntimeError(f"目录 {args.batch_input_dir} 下无子目录")
+
+        print(f"🚀 批量提交 {len(subdirs)} 个任务（扫描目录），并发数 {concurrency}")
+        results = []
+        start_time = time.time()
+
+        def submit_dir_task(subdir: Path, idx: int):
+            """提交目录任务"""
+            # 使用全局参数 + 该目录的 input_directory
+            item_overrides = overrides.copy()
+            item_overrides["input_directory"] = str(subdir)
+
+            # 准备输入文件
+            item_files = []
+            for filepath in subdir.iterdir():
+                if filepath.is_file():
+                    with open(filepath, "rb") as f:
+                        item_files.append((filepath.name, f.read()))
+
+            if not item_files:
+                with open(DEFAULT_INPUT, "rb") as f:
+                    item_files.append((DEFAULT_INPUT.name, f.read()))
+
+            # 任务名使用子目录名
+            item_name = f"{name_prefix}-{subdir.name}"
+
+            # 提交
+            c = DmaClient(http_session=requests.Session())
+            return submit_one(c, item_files, name=item_name, cmd=cmd,
+                            overrides=item_overrides, check=check)
+
+        with ThreadPoolExecutor(max_workers=concurrency) as ex:
+            futures = [ex.submit(submit_dir_task, d, i) for i, d in enumerate(subdirs, 1)]
+            for fu in as_completed(futures):
+                r = fu.result()
+                results.append(r)
+                mark = "✅" if r["ok"] else "❌"
+                tail = "" if r["ok"] else f" {r['error']}"
+                print(f"  [{len(results)}/{len(subdirs)}] {mark} job={r.get('job_id')} ({r['elapsed_ms']}ms){tail}")
+
+        elapsed = time.time() - start_time
+        print_stats(results, elapsed)
+        return
+
+    # 8. dry-run 模式
+    if args.dry_run:
+        snap_body = build_snapshot_body(client, 0, client._user_id(), name=name, cmd=cmd, overrides=overrides)
+        print("【dry-run】将提交的 snapshot body（jobId/inputs 为占位）：")
+        print(json.dumps(snap_body, indent=2, ensure_ascii=False))
+        return
+
+    # 9. 单个任务或 --count 批量
+    count = args.count
+
+    if count > 1:
+        print(f"🚀 批量提交 {count} 个任务，并发数 {concurrency}")
+        results, elapsed = submit_batch(client, count, concurrency, input_files,
+                                        name_prefix=name_prefix, cmd=cmd, overrides=overrides, check=check)
+        print_stats(results, elapsed)
+    else:
+        print(f"🚀 提交单个任务: {name}")
+        result = submit_one(client, input_files, name=name, cmd=cmd, overrides=overrides, check=check)
+        if result["ok"]:
+            print(f"✅ 提交成功 jobId={result['job_id']} ({result['elapsed_ms']}ms)")
+            # --poll: 轮询当前任务状态
+            if args.poll:
+                print("📊 轮询任务状态...")
+                poll_task_status(client, result['job_id'])
+        else:
+            print(f"❌ 提交失败: {result['error']}")