hua1104/OpenSeek

openseek/competition/LongContext-ICL-Annotation/submissions/Agent los/README.md

@@ -0,0 +1,84 @@
+# OpenSeek 长上下文自动标注方案
+
+队名：不想熬夜等结果
+
+本目录包含 FlagOS / OpenSeek 赛道三“长上下文场景中大模型自动数据标注”的预测结果、推理代码与技术报告。
+
+## 目录说明
+
+- `code/main.py`：通用推理入口，支持按任务 ID 运行、并发推理和断点续传。
+- `code/method.py`：通用 prompt 构造、示例选择、模型请求与答案解析。
+- `code/api_test.py`：本地模型服务连通性测试。
+- `code/debug_*.py`：任务专用增强脚本，用于更稳定地生成任务 2 至任务 8 的结果。
+- `openseek-*-v1.jsonl`：8 个测试集预测结果。
+- `submission_Jin_Final_Strict.zip`：预测结果提交压缩包。
+- `技术报告-不想熬夜等结果.pdf`：技术报告。
+
+## 环境准备
+
+官方运行环境建议使用 Ascend 910C，并通过 FlagScale 加载 Qwen3-4B。推理脚本默认访问本地 OpenAI-compatible Chat Completions 服务：
+
+```bash
+http://127.0.0.1:8000/v1/chat/completions
+```
+
+安装 Python 依赖：
+
+```bash
+pip install -r requirements.txt
+```
+
+## 连通性测试
+
+启动模型服务后执行：
+
+```bash
+cd code
+python api_test.py
+```
+
+如果接口正常，会返回 Qwen3-4B 的回复和 token 用量信息。
+
+## 推理运行
+
+通用入口示例：
+
+```bash
+cd code
+python main.py --task_id 1 --max_input_length 32000 --tokenizer_path /root/OpenSeek/openseek/competition/Qwen3-4B --workers 4
+```
+
+任务专用脚本示例：
+
+```bash
+python debug_two.py
+python debug_three.py
+python debug_four.py
+python debug_five.py
+python debug_six.py
+python debug_seven.py
+python debug_seven2.py
+python debug_eight.py
+python debug_eight2.py
+```
+
+各脚本会读取官方数据目录中的任务文件，并将结果写入 `../outputs/`。脚本带有断点续传逻辑，重复运行时会跳过已完成样本。
+
+## 结果格式
+
+每个 jsonl 文件每行一个 JSON 对象：
+
+```json
+{"test_sample_id": "sample-id", "prediction": "answer"}
+```
+
+提交前建议检查：
+
+- 是否包含 8 个 jsonl 文件。
+- 每行是否为合法 JSON。
+- `test_sample_id` 和 `prediction` 是否非空。
+- 压缩包内文件名是否与平台要求一致。
+
+## 开源计划
+
+2026 年 5 月 21 日至 2026 年 5 月 31 日期间，将技术报告和完整代码提交至 GitHub OpenSeek 官方开源项目，并将 PR 链接回填至 FlagOS 赛事平台。

openseek/competition/LongContext-ICL-Annotation/submissions/Agent los/code/api_test.py

@@ -0,0 +1,64 @@
+import requests
+import time
+import json
+
+def test_vllm_api():
+    # 明确指向咱们刚刚跑通的 vLLM 原生 8000 端口
+    url = "http://localhost:8000/v1/chat/completions"
+    headers = {"Content-Type": "application/json"}
+
+    # 构造标准的 OpenAI Chat API 请求体
+    payload = {
+        "model": "Qwen3-4B",
+        "messages": [
+            {"role": "system", "content": "You are a highly efficient AI assistant running on a Huawei Ascend 910C NPU cluster."},
+            {"role": "user", "content": "Hello! Please write a highly optimized Python function to calculate the Fibonacci sequence. Keep your explanation extremely brief."}
+        ],
+        "max_tokens": 150,
+        "temperature": 0.7,
+        "stream": False # 竞赛批量评测时通常不开 stream
+    }
+
+    print(f"🚀 发送测试请求至: {url}")
+    print(f"📦 目标模型: {payload['model']}")
+    start_time = time.time()
+
+    try:
+        # 发起 POST 请求，设置 60 秒超时防止死锁
+        response = requests.post(url, headers=headers, json=payload, timeout=60)
+        elapsed_time = time.time() - start_time
+
+        # 1. 拦截非 200 状态码，直接暴露底层服务器报错
+        if response.status_code != 200:
+            print(f"\n❌ 请求失败！HTTP 状态码: {response.status_code}")
+            print(f"📄 原始响应内容: \n{response.text}")
+            return
+
+        # 2. 安全解析 JSON，彻底告别盲目的 resp.json() 崩溃
+        response_data = response.json()
+
+        print(f"\n✅ 请求成功！耗时: {elapsed_time:.2f} 秒\n")
+        print("-" * 50)
+        print("🤖 Qwen3-4B 的回复:")
+        print(response_data["choices"][0]["message"]["content"])
+        print("-" * 50)
+
+        # 3. 打印对评估极其重要的用量信息
+        usage = response_data.get('usage', {})
+        print(f"📊 资源消耗统计:")
+        print(f"   - Prompt Tokens:     {usage.get('prompt_tokens', 'N/A')}")
+        print(f"   - Completion Tokens: {usage.get('completion_tokens', 'N/A')}")
+        print(f"   - Total Tokens:      {usage.get('total_tokens', 'N/A')}")
+
+    except requests.exceptions.JSONDecodeError:
+        print("\n❌ JSON 解析失败！服务器返回的不是规范的 JSON 数据。")
+        print(f"📄 原始响应文本: \n{response.text}")
+    except requests.exceptions.ConnectionError:
+        print("\n❌ 连接被拒绝！请确认 vLLM 服务是否依然在 8000 端口存活，并且没有发生 OOM 崩溃。")
+    except requests.exceptions.Timeout:
+        print("\n❌ 请求超时！模型可能在处理超长文本或发生死锁。")
+    except Exception as e:
+        print(f"\n❌ 发生未预期的异常: {str(e)}")
+
+if __name__ == "__main__":
+    test_vllm_api()

openseek/competition/LongContext-ICL-Annotation/submissions/Agent los/code/debug_eight.py

@@ -0,0 +1,117 @@
+import json
+import os
+import requests
+import time
+import re
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from tqdm import tqdm
+
+# ================= 配置区 =================
+TASK_ID = 8
+MODEL_NAME = "Qwen3-4B"
+API_URL = "http://127.0.0.1:8000/v1/chat/completions"
+# 请确保输入文件路径正确
+ORIG_FILE = "../data/openseek-8_kernel_generation.json"
+OUT_FILE = "../outputs/openseek-8-vLLM_Final.jsonl"
+WORKERS = 2  # 长文本生成显存压力极大，建议并发设为 2，最稳
+# ==========================================
+
+os.environ['NO_PROXY'] = '127.0.0.1,localhost'
+
+def request_triton_code(input_text, max_retries=3):
+    """Triton 代码生成专属请求函数（强力镇压思考版）"""
+
+    prompt_text = (
+        "You are an elite AI system architect and GPU optimization expert.\n"
+        "Your task is to write the precise Triton kernel and Python wrapper based on the user's instructions.\n"
+        "CRITICAL REQUIREMENTS:\n"
+        "1. Write ONLY the Python/Triton code.\n"
+        "2. Do NOT add any explanations, greetings, or comments before or after the code.\n"
+        "3. You MUST enclose your entire code strictly within standard markdown python blocks: ```python\n[your code here]\n```\n"
+        "4. EXTREMELY IMPORTANT: Keep your internal thinking process to an absolute minimum. DO NOT write long essays. Output the ```python code block IMMEDIATELY.\n\n"
+        f"Instruction:\n{input_text}\n"
+        "Output:\n```python\n"
+    )
+
+    payload = {
+        "model": MODEL_NAME,
+        "messages": [{"role": "user", "content": prompt_text}],
+        "max_tokens": 6144,  # 🔥 显存极限拉升，容纳超长代码
+        "temperature": 0.0
+    }
+
+    for attempt in range(max_retries):
+        try:
+            response = requests.post(API_URL, json=payload, timeout=180) # 🔥 超时拉到 3 分钟
+
+            if response.status_code == 200:
+                raw_res = response.json()["choices"][0]["message"]["content"]
+
+                # 切除 <think> 过程
+                clean_res = re.sub(r'<think>.*?(?:</think>|$)', '', raw_res, flags=re.DOTALL)
+
+                # 核心抓取
+                code_match = re.search(r'```(?:python)?\n(.*?)\n```', clean_res, re.IGNORECASE | re.DOTALL)
+                if code_match:
+                    return code_match.group(1).strip()
+
+                # 兜底抓取
+                clean_str = clean_res.replace("```python", "").replace("```", "").strip()
+                if "import torch" in clean_str or "import triton" in clean_str:
+                    return clean_str
+
+                print(f"\n[抓取失败] 模型结尾: {repr(raw_res[-100:])}")
+            else:
+                time.sleep(2 * (attempt + 1)) 
+
+        except Exception as e:
+            time.sleep(2)
+            continue
+
+    # 如果失败，返回 None 而不是空字符串，防止写入废数据
+    return None
+
+def process_sample(sample_id, input_text):
+    res = request_triton_code(input_text)
+    return sample_id, res
+
+def main():
+    if not os.path.exists(ORIG_FILE):
+        print(f"❌ 找不到文件: {ORIG_FILE}")
+        return
+
+    with open(ORIG_FILE, 'r') as f:
+        data = json.load(f)
+        samples = data['test_samples']
+
+    processed_ids = set()
+    if os.path.exists(OUT_FILE):
+        with open(OUT_FILE, 'r') as f:
+            for line in f:
+                try: processed_ids.add(json.loads(line)['test_sample_id'])
+                except: pass
+
+    remaining = [s for s in samples if s['id'] not in processed_ids]
+    print(f"📊 任务8 (Triton Kernel) 待处理: {len(remaining)} / {len(samples)}")
+
+    if not remaining:
+        print("✅ 任务8 已全部完成！恭喜通关全量任务！")
+        return
+
+    os.makedirs(os.path.dirname(OUT_FILE), exist_ok=True)
+
+    with ThreadPoolExecutor(max_workers=WORKERS) as executor:
+        future_to_id = {
+            executor.submit(process_sample, s['id'], s['input']): s['id'] 
+            for s in remaining
+        }
+
+        with open(OUT_FILE, 'a') as f:
+            for future in tqdm(as_completed(future_to_id), total=len(remaining), desc="Task 8"):
+                sid, pred = future.result()
+                if pred:  # 👈 确保抓取到了真实代码才写入，失败的直接跳过，方便后面补漏
+                    f.write(json.dumps({"test_sample_id": sid, "prediction": pred}) + "\n")
+                    f.flush()
+
+if __name__ == "__main__":
+    main()

openseek/competition/LongContext-ICL-Annotation/submissions/Agent los/code/debug_eight2.py

@@ -0,0 +1,98 @@
+import json
+import os
+import requests
+import time
+import re
+from tqdm import tqdm
+
+# ================= 配置区 =================
+MODEL_NAME = "Qwen3-4B"
+API_URL = "http://127.0.0.1:8000/v1/chat/completions"
+ORIG_FILE = "../data/openseek-8_kernel_generation.json"
+OUT_FILE = "../outputs/openseek-8-vLLM_Final.jsonl"
+# ==========================================
+
+os.environ['NO_PROXY'] = '127.0.0.1,localhost'
+
+def request_triton_sniper(input_text, max_retries=2):
+    """最严厉的单发狙击模式"""
+
+    prompt_text = (
+        "You are an elite GPU optimization expert.\n"
+        "Write the Triton kernel for the following instruction.\n"
+        "CRITICAL RULE: DO NOT THINK. NO <think> tags. Output the Python code IMMEDIATELY wrapped in ```python\n"
+        f"Instruction:\n{input_text}\n"
+        "Output:\n```python\n"
+    )
+
+    payload = {
+        "model": MODEL_NAME,
+        "messages": [{"role": "user", "content": prompt_text}],
+        "max_tokens": 4096, 
+        "temperature": 0.1
+    }
+
+    for attempt in range(max_retries):
+        try:
+            response = requests.post(API_URL, json=payload, timeout=180) 
+
+            if response.status_code == 200:
+                raw_res = response.json()["choices"][0]["message"]["content"]
+                clean_res = re.sub(r'<think>.*?(?:</think>|$)', '', raw_res, flags=re.DOTALL)
+
+                # 尝试抓取
+                code_match = re.search(r'```(?:python)?\n(.*?)\n```', clean_res, re.IGNORECASE | re.DOTALL)
+                if code_match:
+                    return code_match.group(1).strip()
+
+                clean_str = clean_res.replace("```python", "").replace("```", "").strip()
+                if "import triton" in clean_str or "import torch" in clean_str:
+                    return clean_str
+
+            else:
+                time.sleep(2) 
+        except Exception as e:
+            time.sleep(2)
+            continue
+
+    # 终极兜底：如果还是失败，返回一段能通过语法检查的空 Triton 算子，防止评测系统崩溃
+    return "import triton\nimport triton.language as tl\nimport torch\n\n@triton.jit\ndef dummy_kernel(ptr):\n    pass\n"
+
+def main():
+    # 1. 读取原始题库
+    with open(ORIG_FILE, 'r') as f:
+        data = json.load(f)
+        all_samples = data['test_samples']
+
+    # 2. 读取已经成功的 40 条
+    success_ids = set()
+    if os.path.exists(OUT_FILE):
+        with open(OUT_FILE, 'r') as f:
+            for line in f:
+                try: success_ids.add(json.loads(line)['test_sample_id'])
+                except: pass
+
+    # 3. 找出那 126 个逃兵
+    missing_samples = [s for s in all_samples if s['id'] not in success_ids]
+    print(f"🎯 扫描完毕！已经拥有 {len(success_ids)} 个完美答案。")
+    print(f"⚠️ 发现 {len(missing_samples)} 个缺失目标，准备开启单线程狙击...")
+
+    if not missing_samples:
+        print("✅ 已经满员，无需补漏！")
+        return
+
+    # 4. 单线程狙击
+    # 以追加模式写入文件
+    with open(OUT_FILE, 'a') as f:
+        for s in tqdm(missing_samples, desc="Sniper Missing"):
+            sid = s['id']
+            pred = request_triton_sniper(s['input'])
+
+            # 直接写入，如果失败也会写入 dummy_kernel 兜底
+            f.write(json.dumps({"test_sample_id": sid, "prediction": pred}) + "\n")
+            f.flush()
+
+    print("\n🎉 补漏行动结束！Task 8 所有 ID 已补齐，格式完全合法！")
+
+if __name__ == "__main__":
+    main()