add .dockerignore

main.py

@@ -1,16 +1,7 @@
-"""ModelHub XC 适配智能体 —— 单任务提交器
+"""ModelHub XC 适配智能体 —— 低排队算力批次（5 条）
 
-启动后向 ModelHub XC 提交 1 条模型适配验证任务，随后保持 /health 存活等待平台停止。
-
-提交内容（已查重，非重复模型）：
-  模型：https://www.modelscope.cn/models/qnguyen3/qwq-1.5b-grpo  (Qwen2 / 1.5B)
-  算力：Ascend_910-b4      框架：vllm      任务类型：text-generation
-
-平台契约：
-  - 根目录 Dockerfile，监听 8080
-  - GET /health 探活
-  - 读取平台注入的 STRATEGY_ID
-  - 正确处理 SIGTERM
+每条任务分配一张低排队算力卡，并选用该卡历史通过率较高的框架。
+OSP 系列 max_position_embeddings=2048，配置里的 max_model_len 已由 4096 下调为 2048。
 """
 import json
 import os
@@ -22,160 +13,93 @@ from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 
 import requests
 
-# ─────────────────────────── 业务配置 ───────────────────────────
 BASE_URL = os.environ.get("BASE_URL", "https://modelhub.org.cn")
 ADD_TASK_ENDPOINT = "/api/adapt/task/add"
-
 XC_TOKEN = os.environ.get("XC_TOKEN", "6616dfe75c874e61987b1f7c908e6d77")
 STRATEGY_ID = os.environ.get("STRATEGY_ID", "")
-
-MODEL_ADDRESS = "https://www.modelscope.cn/models/qnguyen3/qwq-1.5b-grpo"
 TASK_TYPE = "text-generation"
-TARGET_GPU = "Ascend_910-b4"
-FRAMEWORK = "vllm"
 
-# 与 POST /api/adapt/task/build-config?gpuType=Ascend_910-b4&framework=vllm&taskType=text-generation 输出一致
-CONFIG_PARAMS = """framework: vllm
-api: completion
-max_tokens: 1024
-temperature: 0.7
-repetition_penalty: 1.2
-top_p: 0.9
-lang: zh
-max_model_len: 2048
-sut_config:
-  gpu_num: 1
-  values:
-    command:
-    - vllm
-    - serve
-    - /model
-    - --port
-    - '8000'
-    - --served-model-name
-    - llm
-    - --max-model-len
-    - '2048'
-    - --dtype
-    - auto
-    - --gpu-memory-utilization
-    - '0.95'
-    - -tp
-    - '1'
-    - --enforce-eager
-    - --trust-remote-code
-ref_config:
-  gpu_num: 1
-  values:
-    command:
-    - vllm
-    - serve
-    - /model
-    - --port
-    - '8000'
-    - --served-model-name
-    - llm
-    - --max-model-len
-    - '2048'
-    - --dtype
-    - auto
-    - --gpu-memory-utilization
-    - '0.95'
-    - -tp
-    - '1'
-    - --enforce-eager
-    - --trust-remote-code
-"""
+TASKS = [
+    {
+        "label": "OSP-1.4B-100B-Muon-Only",
+        "model": "https://www.modelscope.cn/models/dmis-lab/OSP-1.4B-100B-Muon-Only",
+        "gpu": "MetaX_c-500",
+        "framework": "vllm_tokenizer_patch",
+        "config": "nv_framework: vllm_fix_tokenizer\nframework: vllm_tokenizer_patch\nstorage: gpfs\nmax_tokens: 1024\nmax_model_len: 2048\napi: completion\nsut_config:\n    gpu_num: 1\n    values:\n        command: ['/opt/conda/bin/vllm', 'serve', '/model', '--port', '8000', '--served-model-name', 'llm',  '--max-model-len', '2048', '--gpu-memory-utilization', '0.9',  '--enforce-eager', '--trust-remote-code' ,'-tp', '1' ]\nref_config:\n    gpu_num: 1\n    values:\n        command: ['/opt/entrypoint.sh', '/model', '--port', '80', '--served-model-name', 'llm', '--max-model-len', '2048', '--enforce-eager', '--trust-remote-code', '-tp', '1']"
+    },
+    {
+        "label": "OSP-1.4B-100B-Muon-SSNorm",
+        "model": "https://www.modelscope.cn/models/dmis-lab/OSP-1.4B-100B-Muon-SSNorm",
+        "gpu": "Kunlunxin_p-800",
+        "framework": "vllm_fix_tokenizer",
+        "config": "framework: vllm_fix_tokenizer\nlang: en\nmax_model_len: 2048\nsut_config:\n  gpu_num: 1\n  values:\n    command:\n    - /opt/entrypoint.sh    - /model\n    - --port\n    - '80'\n    - --served-model-name\n    - llm\n    - --max-model-len\n    - '2048'\n    - --enforce-eager\n    - --trust-remote-code\n    - -tp\n    - '1'\nref_config:\n  gpu_num: 1\n  values:\n    command:\n    - /opt/entrypoint.sh    - /model\n    - --port\n    - '80'\n    - --served-model-name\n    - llm\n    - --max-model-len\n    - '2048'\n    - --enforce-eager\n    - --trust-remote-code\n    - -tp\n    - '1'"
+    },
+    {
+        "label": "OSP-1.4B-100B-Shampoo-SSNorm",
+        "model": "https://www.modelscope.cn/models/dmis-lab/OSP-1.4B-100B-Shampoo-SSNorm",
+        "gpu": "Cambricon_mlu-370-x8",
+        "framework": "vllm-mlu",
+        "config": "framework: vllm-mlu\nnv_framework: vllm_fix_tokenizer\napi: completion\nlang: en\ntemperature: 0.4\nrepetition_penalty: 1.1\ntop_p: 0.9\nsut_config:\n  gpu_num: 1\n  values:\n    env:\n    - name: MAX_MODEL_LEN\n      value: 2048\n    - name: VLLM_ALLOW_LONG_MAX_MODEL_LEN\n      value: '1'\n    command:\n    - vllm\n    - serve\n    - /model\n    - --port\n    - '8000'\n    - --served-model-name\n    - llm\n    - --max-model-len\n    - '2048'\n    - --trust-remote-code\n    - --dtype\n    - float16\nref_config:\n  cpu_num: 2\n  gpu_num: 1\n  values:\n    env:\n    - name: MAX_MODEL_LEN\n      value: 2048\n    - name: VLLM_ALLOW_LONG_MAX_MODEL_LEN\n      value: '1'\n    command:\n    - /opt/entrypoint.sh\n    - /model\n    - --port\n    - '80'\n    - --served-model-name\n    - llm\n    - --max-model-len\n    - '2048'\n    - --trust-remote-code\n    - --dtype\n    - float16\n"
+    },
+    {
+        "label": "OSP-1.4B-100B-Muon-EmbProj",
+        "model": "https://www.modelscope.cn/models/dmis-lab/OSP-1.4B-100B-Muon-EmbProj",
+        "gpu": "Biren_166m",
+        "framework": "vllm_fix_tokenizer",
+        "config": "framework: vllm_fix_tokenizer\nlang: en\nmax_model_len: 2048\nsut_config:\n  gpu_num: 1\n  values:\n    command:\n    - /opt/entrypoint.sh    - /model\n    - --port\n    - '80'\n    - --served-model-name\n    - llm\n    - --max-model-len\n    - '2048'\n    - --enforce-eager\n    - --trust-remote-code\n    - -tp\n    - '1'\nref_config:\n  gpu_num: 1\n  values:\n    command:\n    - /opt/entrypoint.sh    - /model\n    - --port\n    - '80'\n    - --served-model-name\n    - llm\n    - --max-model-len\n    - '2048'\n    - --enforce-eager\n    - --trust-remote-code\n    - -tp\n    - '1'"
+    },
+    {
+        "label": "OSP-1.4B-1T-Muon-SSNorm-EmbProj",
+        "model": "https://www.modelscope.cn/models/dmis-lab/OSP-1.4B-1T-Muon-SSNorm-EmbProj",
+        "gpu": "Sunrise_pt-200-x1",
+        "framework": "vllm_fix_tokenizer",
+        "config": "framework: vllm_fix_tokenizer\napi: completion\nmax_tokens: 1024\ntemperature: 0.7\nrepetition_penalty: 1.2\ntop_p: 0.9\nlang: zh\nmax_model_len: 2048\nsut_config:\n  gpu_num: 1\n  values:\n    command:\n    - /opt/entrypoint.sh\n    - /model\n    - --port\n    - '8000'\n    - --served-model-name\n    - llm\n    - --max-model-len\n    - '2048'\n    - --gpu-memory-utilization\n    - '0.9'\n    - --enforce-eager\n    - --trust-remote-code\nref_config:\n  gpu_num: 1\n  values:\n    command:\n    - /opt/entrypoint.sh\n    - /model\n    - --port\n    - '8000'\n    - --served-model-name\n    - llm\n    - --max-model-len\n    - '2048'\n    - --dtype\n    - float16\n    - --gpu-memory-utilization\n    - '0.95'\n    - -tp\n    - '1'\n    - --enforce-eager\n    - --trust-remote-code\n"
+    }
+]
 
 HEADERS = {"Content-Type": "application/json", "Xc-Token": XC_TOKEN}
-
 HTTP_HOST, HTTP_PORT = "0.0.0.0", 8080
-MAX_ATTEMPTS = 5
-RETRY_GAP = 15  # 秒
-
-_state = {
-    "strategy_id": STRATEGY_ID,
-    "phase": "starting",       # starting | submitting | done | error
-    "model_address": MODEL_ADDRESS,
-    "target_gpu": TARGET_GPU,
-    "framework": FRAMEWORK,
-    "attempts": 0,
-    "attempt_log": [],
-    "task_id": None,
-    "started_at": None,
-    "finished_at": None,
-}
+GAP = 4
 _shutdown = threading.Event()
-_submitting = threading.Lock()
-_submitted = False
+_state = {"phase": "starting", "total": len(TASKS), "submitted": 0, "failed": 0, "results": []}
 
 
-def submit_task():
-    """提交一次；已提交过则直接返回。"""
-    global _submitted
-    with _submitting:
-        if _submitted:
-            return True
-
-    payload = {
-        "configParams": CONFIG_PARAMS,
-        "framework": FRAMEWORK,
-        "modelAddress": MODEL_ADDRESS,
-        "targetGpu": TARGET_GPU,
+def submit_one(t):
+    body = {
+        "configParams": t["config"],
+        "framework": t["framework"],
+        "modelAddress": t["model"],
+        "targetGpu": t["gpu"],
         "taskType": TASK_TYPE,
         "strategyId": STRATEGY_ID,
     }
+    try:
+        r = requests.post(BASE_URL + ADD_TASK_ENDPOINT, headers=HEADERS, json=body, timeout=90)
+        j = r.json()
+    except Exception as e:  # noqa: BLE001
+        j = {"code": -1, "message": "exception: %s" % e}
+    tid = (j.get("data") or {}).get("id")
+    print("[submit] %s gpu=%s fw=%s code=%s msg=%s taskId=%s"
+          % (t["label"], t["gpu"], t["framework"], j.get("code"), j.get("message"), tid), flush=True)
+    return j.get("code"), tid
 
+
+def run():
     _state["phase"] = "submitting"
-    _state["started_at"] = datetime.utcnow().isoformat()
-
-    for i in range(1, MAX_ATTEMPTS + 1):
+    for i, t in enumerate(TASKS, 1):
         if _shutdown.is_set():
             break
-        _state["attempts"] = i
-        try:
-            resp = requests.post(BASE_URL + ADD_TASK_ENDPOINT, headers=HEADERS,
-                                 json=payload, timeout=60)
-            body = resp.json()
-        except Exception as e:  # noqa: BLE001
-            body = {"code": -1, "message": "exception: %s" % e}
-
-        code = body.get("code")
-        msg = body.get("message", "")
-        entry = {"attempt": i, "time": datetime.utcnow().isoformat(),
-                 "http_status": None, "code": code, "message": msg}
-        try:
-            entry["http_status"] = resp.status_code
-        except Exception:  # noqa: BLE001
-            pass
-        _state["attempt_log"].append(entry)
-        print("[submit] #%d code=%s message=%s" % (i, code, msg), flush=True)
-
-        if code == 0:
-            data = body.get("data") or {}
-            _state["task_id"] = data.get("id")
-            _state["phase"] = "done"
-            _state["finished_at"] = datetime.utcnow().isoformat()
-            _submitted = True
-            print("[submit] 提交成功 taskId=%s" % _state["task_id"], flush=True)
-            return True
-
-        # 已存在/重复等不可重试错误，直接退出
-        if code not in (-1, 60014, 50000):
-            _state["phase"] = "error"
-            _state["finished_at"] = datetime.utcnow().isoformat()
-            print("[submit] 终止：不可重试错误", flush=True)
-            return False
-
-        if i < MAX_ATTEMPTS and not _shutdown.is_set():
-            time.sleep(RETRY_GAP)
-
-    _state["phase"] = "error"
-    _state["finished_at"] = datetime.utcnow().isoformat()
-    return False
+        code, tid = submit_one(t)
+        _state["results"].append({"label": t["label"], "gpu": t["gpu"],
+                                  "framework": t["framework"], "code": code, "task_id": tid})
+        _state["submitted" if code == 0 else "failed"] += 1
+        if i < len(TASKS):
+            time.sleep(GAP)
+    _state["phase"] = "done"
+    print("[batch] 完成 submitted=%d failed=%d" % (_state["submitted"], _state["failed"]), flush=True)
 
 
-class Handler(BaseHTTPRequestHandler):
+class H(BaseHTTPRequestHandler):
     def do_GET(self):
         if self.path == "/health":
             self._json({"status": "ok"})
@@ -184,56 +108,29 @@ class Handler(BaseHTTPRequestHandler):
         else:
             self._json({"error": "not found"}, 404)
 
-    def do_POST(self):
-        if self.path == "/task":
-            submit_task()
-            self._json(_state)
-        else:
-            self._json({"error": "not found"}, 404)
-
-    def _json(self, body, code=200):
-        raw = json.dumps(body, ensure_ascii=False, default=str).encode()
-        self.send_response(code)
+    def _json(self, b, c=200):
+        raw = json.dumps(b, ensure_ascii=False, default=str).encode()
+        self.send_response(c)
         self.send_header("Content-Type", "application/json")
         self.send_header("Content-Length", str(len(raw)))
         self.end_headers()
         self.wfile.write(raw)
 
-    def log_message(self, fmt, *args):
-        print(self.address_string() + " - " + (fmt % args), flush=True)
-
-
-def _handle_signal(signum, _frame):
-    print("[main] 收到信号 %s，准备退出" % signum, flush=True)
-    _shutdown.set()
+    def log_message(self, *a):
+        pass
 
 
 def main():
-    signal.signal(signal.SIGTERM, _handle_signal)
-    signal.signal(signal.SIGINT, _handle_signal)
-
-    print("[main] strategy_id=%s" % STRATEGY_ID, flush=True)
-    print("[main] model=%s gpu=%s framework=%s" % (MODEL_ADDRESS, TARGET_GPU, FRAMEWORK), flush=True)
-
-    http_thread = threading.Thread(target=_run_http, daemon=False)
-    http_thread.start()
-
-    worker = threading.Thread(target=submit_task, daemon=True)
-    worker.start()
-
-    _shutdown.wait()
-    print("[main] 退出", flush=True)
-
-
-def _run_http():
-    server = ThreadingHTTPServer((HTTP_HOST, HTTP_PORT), Handler)
-    server.timeout = 1
-    print("[http] 监听 %s:%d" % (HTTP_HOST, HTTP_PORT), flush=True)
+    signal.signal(signal.SIGTERM, lambda *_: _shutdown.set())
+    print("[main] strategy_id=%s tasks=%d" % (STRATEGY_ID, len(TASKS)), flush=True)
+    for t in TASKS:
+        print("[main] 计划: %s @ %s / %s" % (t["label"], t["gpu"], t["framework"]), flush=True)
+    srv = ThreadingHTTPServer((HTTP_HOST, HTTP_PORT), H)
+    srv.timeout = 1
+    threading.Thread(target=run, daemon=True).start()
     while not _shutdown.is_set():
-        server.handle_request()
-    server.server_close()
-    print("[http] 已关闭", flush=True)
+        srv.handle_request()
+    srv.server_close()
 
 
-if __name__ == "__main__":
-    main()
+main()