Optimize SWE-bench build_images: skip make_test_spec when images are pre-loaded

- make_test_spec() for all 500 samples was the real bottleneck that made
  swe_bench_verified appear stuck at 'Processing records: 0%' even when all
  instance images were already loaded locally.
- Now image names are computed directly from instance IDs, matching the
  swebench naming convention. Local images are listed once and missing ones
  are reported immediately. make_test_spec is only called for images that
  actually need to be built.
- With 500 pre-loaded images, build_images() now completes in ~45s instead of
  hanging indefinitely.

Also add bash/diagnose_swe_images.py for quickly checking which images are
missing locally.

bash/diagnose_swe_images.py

@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+"""
+诊断 swe_bench_verified 镜像准备阶段卡在哪里。
+
+用法:
+    python bash/diagnose_swe_images.py
+
+它会模拟 evalscope 内部 build_images() 的前半段：
+1. 加载数据集
+2. 生成 500 个 instance 的镜像名
+3. 列出本地 Docker 镜像
+4. 对比哪些已存在、哪些缺失
+5. 打印每一步耗时
+"""
+
+import os
+import sys
+import time
+from pathlib import Path
+
+# 让 evalscope 可导入
+PROJECT_ROOT = Path(__file__).parent.parent.resolve()
+sys.path.insert(0, str(PROJECT_ROOT / 'evalscope'))
+
+os.environ.setdefault('USE_MODELSCOPE_HUB', '1')
+os.environ.setdefault('HF_ENDPOINT', 'https://hf-mirror.com')
+os.environ.setdefault('PYTHONUNBUFFERED', '1')
+
+from evalscope.api.dataset import RemoteDataLoader, FieldSpec
+from evalscope.benchmarks.swe_bench.utils import resolve_swebench_arch
+from swebench.harness.test_spec.test_spec import make_test_spec
+
+
+def get_local_images():
+    import subprocess
+    result = subprocess.run(
+        ['docker', 'images', '--format', '{{.Repository}}:{{.Tag}}'],
+        capture_output=True, text=True
+    )
+    if result.returncode != 0:
+        print(f'ERROR: docker images failed: {result.stderr}')
+        return set()
+    return {line.strip() for line in result.stdout.splitlines() if line.strip()}
+
+
+def main():
+    t0 = time.time()
+
+    print('==> 1. 加载 SWE-bench_Verified 数据集', flush=True)
+    loader = RemoteDataLoader(
+        data_id_or_path='princeton-nlp/SWE-bench_Verified',
+        split='test',
+        sample_fields=FieldSpec(
+            input='problem_statement',
+            metadata=[
+                'instance_id', 'repo', 'base_commit', 'patch',
+                'PASS_TO_PASS', 'FAIL_TO_PASS', 'test_patch', 'version',
+                'environment_setup_commit', 'hints_text', 'created_at',
+            ],
+        ),
+    )
+    samples = loader.load()
+    print(f'    样本数: {len(samples)}, 耗时: {time.time()-t0:.2f}s', flush=True)
+
+    t1 = time.time()
+    print('==> 2. 生成 500 个镜像名 (make_test_spec)', flush=True)
+    missing_repo = 0
+    image_names = []
+    last_log = time.time()
+    for idx, s in enumerate(samples):
+        metadata = s.metadata if hasattr(s, 'metadata') else s
+        instance_id = metadata.get('instance_id')
+        try:
+            arch = resolve_swebench_arch(instance_id, '')
+            spec = make_test_spec(metadata, namespace='swebench', arch=arch)
+            image_names.append(spec.instance_image_key)
+        except KeyError as e:
+            missing_repo += 1
+            if missing_repo <= 3:
+                print(f'    WARNING: instance {instance_id} 缺少字段: {e}', flush=True)
+        if time.time() - last_log > 10:
+            print(f'    进度: {idx+1}/{len(samples)} ...', flush=True)
+            last_log = time.time()
+    print(f'    生成镜像名: {len(image_names)}, 缺失字段: {missing_repo}, 耗时: {time.time()-t1:.2f}s', flush=True)
+
+    t2 = time.time()
+    print('==> 3. 扫描本地 Docker 镜像', flush=True)
+    local_images = get_local_images()
+    print(f'    本地镜像总数: {len(local_images)}, 耗时: {time.time()-t2:.2f}s', flush=True)
+
+    t3 = time.time()
+    print('==> 4. 对比缺失镜像', flush=True)
+    needed = set(image_names)
+    present = needed & local_images
+    missing = needed - local_images
+    print(f'    需要: {len(needed)}, 已存在: {len(present)}, 缺失: {len(missing)}, 耗时: {time.time()-t3:.2f}s', flush=True)
+
+    if missing:
+        print('\n前 10 个缺失镜像:', flush=True)
+        for name in list(missing)[:10]:
+            print(f'    {name}', flush=True)
+    else:
+        print('\n所有需要镜像均已存在，evalscope 不应该再拉取/构建。', flush=True)
+
+    print(f'\n总耗时: {time.time()-t0:.2f}s', flush=True)
+
+
+if __name__ == '__main__':
+    main()

evalscope/evalscope/benchmarks/swe_bench/build_images.py

@@ -57,18 +57,43 @@ def build_images(
     # Note that remote images are named eg "sphinx-doc_1776_sphinx-11502"
     namespace = (dockerhub_username or 'swebench') if use_remote_images else None
 
+    def _get_instance_image_name(instance_id: str) -> str:
+        """Fast image name computation that matches make_test_spec output."""
+        arch = resolve_swebench_arch(instance_id, force_arch)
+        if namespace:
+            # Remote / pre-built image format: swebench/sweb.eval.x86_64.repo_1776_name-123:latest
+            updated_id = instance_id.replace('__', '_1776_')
+            return f'{namespace}/sweb.eval.{arch}.{updated_id}:latest'
+        # Local build format: sweb.eval.x86_64.repo__name-123:latest
+        return f'sweb.eval.{arch}.{instance_id}:latest'
+
+    # Fast path: compute image names directly without calling make_test_spec.
+    # make_test_spec is very slow for 500 samples and is only needed when we
+    # actually have to build missing images locally.
+    logger.info('Computing SWE-Bench image names from instance IDs...')
     for swebench_instance in samples_hf:
-        arch = resolve_swebench_arch(swebench_instance['instance_id'], force_arch)
+        instance_id = swebench_instance['instance_id']
-        test_spec = make_test_spec(swebench_instance, namespace=namespace, arch=arch)
+        id_to_docker_image[instance_id] = _get_instance_image_name(instance_id)
-        docker_image_name = test_spec.instance_image_key
-        id_to_docker_image[swebench_instance['instance_id']] = docker_image_name
-        id_to_test_spec[swebench_instance['instance_id']] = test_spec
 
     # Get list of locally available Docker images
+    logger.info('Listing local Docker images...')
     available_docker_images = _get_available_docker_images()
     samples_to_build_images_for = [
         s for s in samples_hf if id_to_docker_image[s['instance_id']] not in available_docker_images
     ]
+    logger.info(
+        f'SWE-Bench images: {len(samples_hf)} needed, '
+        f'{len(samples_hf) - len(samples_to_build_images_for)} already local, '
+        f'{len(samples_to_build_images_for)} missing'
+    )
+
+    # Only build expensive test_specs for images that are actually missing.
+    if samples_to_build_images_for:
+        logger.info('Building test specs for missing images (this may take a while)...')
+        for swebench_instance in samples_to_build_images_for:
+            arch = resolve_swebench_arch(swebench_instance['instance_id'], force_arch)
+            test_spec = make_test_spec(swebench_instance, namespace=namespace, arch=arch)
+            id_to_test_spec[swebench_instance['instance_id']] = test_spec
 
     # Try to pull images from Docker Hub first if requested
     if use_remote_images and len(samples_to_build_images_for) > 0: