fix docker build

backend/.dockerignore

@@ -24,3 +24,6 @@ algorithm_subprocesses/**/*.xlsx
 algorithm_subprocesses/**/*.xls
 *.md
 .dockerignore
+# Offline bake inputs: keep *.pth / *.task; ignore partials and accidental duplicates.
+weights/**/*.part
+weights/**/*.pth.*

backend/.env.example

@@ -4,11 +4,11 @@
 #
 # 5-6 ActionFormer 实时算法（默认开启）：
 #   - app/resources/actionformer_epoch_045.pth.tar 必须存在（离线下发，~110MB，未入 git）。
-#   - VideoSwin Swin3D-T 权重在 Docker 构建时预下载到 /app/.cache/torch（见 scripts/bake_torch_hub_checkpoint.py）；
-#     运行时不再访问 pytorch.org。「首次运行」指 torch 缓存为空时才会联网下载；现已改为构建时烘焙进镜像。
-#     国内 PyPI 镜像（南大/清华/阿里）不同步 /models/*.pth，构建默认先试 uv.agentsmirror.com 再回退官方源。
-#     离线/弱网：先 wget 权重到 backend/weights/swin3d_t-7615ae03.pth，再 docker compose build api。
-#     或：docker compose build --build-arg PYTORCH_MODELS_URL=https://your-mirror/.../swin3d_t-7615ae03.pth api
+#   - torchvision 预训练 hub 权重在 Docker 构建时烘焙到 /app/.cache/torch（scripts/bake_pretrained_weights.py）：
+#     swin3d_t-7615ae03.pth（VideoSwin）、resnet50-0676ba61.pth（医生 ReID）；运行时不再访问 pytorch.org。
+#     国内 PyPI 镜像不同步 /models/*.pth；构建默认 uv.agentsmirror.com 再回退官方源。
+#     离线/弱网：将上述文件放入 backend/weights/ 后 docker compose build api（本地有则跳过下载）。
+#     或：docker compose build --build-arg PYTORCH_MODELS_MIRROR=https://your-mirror/download.pytorch.org api
 #   - Linux GPU 机：镜像内 torch / torchvision / torchaudio 为 cu130 wheel；
 #     宿主机需 NVIDIA 驱动 + NVIDIA Container Toolkit；`api` 服务已配置 `gpus: all`。
 #     启动后可验证：docker compose exec api python -c "import torch; print(torch.cuda.is_available())"

backend/Dockerfile

@@ -56,11 +56,20 @@ COPY algorithm_subprocesses ./algorithm_subprocesses/
 # Bake runtime patches/assets so non-root api never writes the read-only bundle tree.
 COPY app/algorithm_runner/actionformer_release/libs/utils/nms.py \
      algorithm_subprocesses/5.15/code/actionformer_release/libs/utils/nms.py
-RUN mkdir -p algorithm_subprocesses/5.15/doctor_identity_package/.mediapipe_models && \
-    curl -fsSL --retry 3 \
-      -o algorithm_subprocesses/5.15/doctor_identity_package/.mediapipe_models/pose_landmarker_lite.task \
-      "https://storage.googleapis.com/mediapipe-models/pose_landmarker/pose_landmarker_lite/float16/1/pose_landmarker_lite.task" && \
-    test -s algorithm_subprocesses/5.15/doctor_identity_package/.mediapipe_models/pose_landmarker_lite.task
+# Optional offline assets (torch hub + MediaPipe); see backend/weights/.gitkeep
+COPY weights ./weights/
+RUN set -eux; \
+    MP_DEST=algorithm_subprocesses/5.15/doctor_identity_package/.mediapipe_models/pose_landmarker_lite.task; \
+    mkdir -p "$(dirname "$MP_DEST")"; \
+    if [ -s weights/pose_landmarker_lite.task ]; then \
+      cp weights/pose_landmarker_lite.task "$MP_DEST"; \
+      echo "mediapipe pose model from local weights/"; \
+    else \
+      curl -fsSL --retry 3 -o "$MP_DEST" \
+        "https://storage.googleapis.com/mediapipe-models/pose_landmarker/pose_landmarker_lite/float16/1/pose_landmarker_lite.task"; \
+      echo "mediapipe pose model downloaded"; \
+    fi; \
+    test -s "$MP_DEST"
 
 # uv.lock pins uv.agentsmirror.com artifact URLs. Rewrite to mainland mirrors (same /packages/... paths).
 # PyPI: Tsinghua  |  PyTorch wheel index: 南大 (syncs download.pytorch.org / download-r2)
@@ -73,29 +82,28 @@ RUN sed -i \
 
 ENV UV_DEFAULT_INDEX=https://pypi.tuna.tsinghua.edu.cn/simple
 
-# VideoSwin (Swin3D-T) hub weights: bake at build time so RTSP batch jobs never hit pytorch.org at runtime.
-# Domestic PyPI mirrors (NJU/Tsinghua/Aliyun) only sync pip wheels, not /models/*.pth; default tries
-# uv.agentsmirror.com (same ecosystem as uv.lock) then download.pytorch.org. Optional offline bake:
-#   backend/weights/swin3d_t-7615ae03.pth  (see weights/.gitkeep)
-# Override: --build-arg PYTORCH_MODELS_URL=... or PYTORCH_MODELS_MIRROR=...
+# Torchvision hub weights (Swin3D-T + ResNet50): bake at build so runtime never hits pytorch.org.
+# Domestic PyPI mirrors do not sync /models/*.pth. Optional offline: backend/weights/*.pth (see .gitkeep).
+# Override: --build-arg PYTORCH_MODELS_MIRROR=... or PYTORCH_MODELS_URL=... (swin only).
 ARG PYTORCH_MODELS_MIRROR=
 ARG PYTORCH_MODELS_URL=
 ENV PYTORCH_MODELS_MIRROR=${PYTORCH_MODELS_MIRROR} \
     PYTORCH_MODELS_URL=${PYTORCH_MODELS_URL}
-# Optional offline weight (only .gitkeep by default; add swin3d_t-7615ae03.pth before build if needed).
-COPY weights ./weights/
 
+# Do not cache-mount /app/.cache/torch: BuildKit cache mounts are not exported into the image.
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv sync --frozen --no-dev --no-compile --refresh-package numpy --refresh-package mediapipe && \
     .venv/bin/python -c "import alembic" && \
     .venv/bin/python -c "import numpy; import numpy.lib._index_tricks_impl" && \
     .venv/bin/python -c "import mediapipe as mp; print('mediapipe', mp.__version__)" && \
-    mkdir -p /app/.cache/ultralytics && \
-    PYTORCH_MODELS_LOCAL_PATH=/app/weights/swin3d_t-7615ae03.pth \
-      .venv/bin/python scripts/bake_torch_hub_checkpoint.py && \
+    mkdir -p /app/.cache/ultralytics /app/.cache/torch/hub/checkpoints && \
+    PYTORCH_MODELS_LOCAL_DIR=/app/weights \
+      .venv/bin/python scripts/bake_pretrained_weights.py && \
+    test -s /app/.cache/torch/hub/checkpoints/swin3d_t-7615ae03.pth && \
+    test -s /app/.cache/torch/hub/checkpoints/resnet50-0676ba61.pth && \
     rm -rf /app/weights && \
-    TORCH_HOME=/app/.cache/torch .venv/bin/python -c "from torchvision.models.video import Swin3D_T_Weights, swin3d_t; swin3d_t(weights=Swin3D_T_Weights.KINETICS400_V1); print('swin3d_t cached ok')" && \
-    chmod -R a+rX /app/.venv /app/algorithm_subprocesses /app/.cache/torch /app/.cache/ultralytics
+    chmod -R a+rX /app/.venv /app/algorithm_subprocesses && \
+    chmod -R a+rwX /app/.cache/torch /app/.cache/ultralytics
 
 ENV PATH="/app/.venv/bin:$PATH"
 

backend/scripts/bake_pretrained_weights.py

@@ -0,0 +1,212 @@
+#!/usr/bin/env python3
+"""Bake torchvision / torch.hub checkpoints into TORCH_HOME during Docker image build.
+
+Runtime batch jobs and doctor identity must not download from pytorch.org when the
+api container runs as a non-root compose user (read-only /app/.cache/torch).
+
+Place optional offline copies under backend/weights/ before build:
+  - swin3d_t-7615ae03.pth   (VideoSwin / Swin3D-T, ~110MB)
+  - resnet50-0676ba61.pth   (doctor ReID backbone, ~98MB)
+
+Override mirrors: PYTORCH_MODELS_MIRROR, PYTORCH_MODELS_URL (per-file URL not supported).
+Legacy single-file: PYTORCH_MODELS_LOCAL_PATH (applies only to swin3d_t).
+"""
+
+from __future__ import annotations
+
+import os
+import shutil
+import subprocess
+import sys
+from dataclasses import dataclass
+from pathlib import Path
+
+OFFICIAL_PREFIX = "https://download.pytorch.org/models"
+AGENTSMIRROR_PREFIX = "https://uv.agentsmirror.com/download.pytorch.org/models"
+
+
+@dataclass(frozen=True)
+class HubCheckpoint:
+    filename: str
+    min_bytes: int
+    label: str
+
+
+# All torchvision hub files downloaded at runtime by production paths (as of 5.15 bundle).
+HUB_CHECKPOINTS: tuple[HubCheckpoint, ...] = (
+    HubCheckpoint(
+        filename="swin3d_t-7615ae03.pth",
+        min_bytes=100_000_000,
+        label="VideoSwin Swin3D-T (Kinetics-400)",
+    ),
+    HubCheckpoint(
+        filename="resnet50-0676ba61.pth",
+        min_bytes=90_000_000,
+        label="doctor ReID ResNet50 (ImageNet-1K)",
+    ),
+)
+
+
+def _candidate_urls(filename: str) -> list[str]:
+    explicit = (os.environ.get("PYTORCH_MODELS_URL") or "").strip()
+    # Legacy single-URL override applied only to VideoSwin (other files use mirror list).
+    if explicit and filename == "swin3d_t-7615ae03.pth":
+        return [explicit]
+
+    raw = (os.environ.get("PYTORCH_MODELS_MIRROR") or "").strip().rstrip("/")
+    prefixes = [p for p in (raw, AGENTSMIRROR_PREFIX) if p]
+    urls: list[str] = []
+    for prefix in prefixes:
+        urls.append(f"{prefix}/models/{filename}")
+    urls.append(f"{OFFICIAL_PREFIX}/{filename}")
+
+    seen: set[str] = set()
+    ordered: list[str] = []
+    for url in urls:
+        if url in seen:
+            continue
+        seen.add(url)
+        ordered.append(url)
+    return ordered
+
+
+def _local_source(filename: str) -> Path | None:
+    legacy = (os.environ.get("PYTORCH_MODELS_LOCAL_PATH") or "").strip()
+    if legacy and filename == "swin3d_t-7615ae03.pth":
+        path = Path(legacy)
+        if path.is_file():
+            return path
+
+    local_dir = (os.environ.get("PYTORCH_MODELS_LOCAL_DIR") or "").strip()
+    if local_dir:
+        path = Path(local_dir) / filename
+        if path.is_file():
+            return path
+    return None
+
+
+def _copy_local(src: Path, dest: Path, *, min_bytes: int) -> None:
+    size = src.stat().st_size
+    if size < min_bytes:
+        raise OSError(f"local file too small ({size} bytes): {src}")
+    dest.parent.mkdir(parents=True, exist_ok=True)
+    shutil.copy2(src, dest)
+
+
+def _download_with_curl(url: str, dest: Path, *, min_bytes: int) -> None:
+    curl = shutil.which("curl")
+    if curl is None:
+        raise RuntimeError("curl not found")
+    tmp = dest.with_suffix(dest.suffix + ".part")
+    tmp.unlink(missing_ok=True)
+    proc = subprocess.run(
+        [
+            curl,
+            "-fL",
+            "--retry",
+            "5",
+            "--retry-all-errors",
+            "--retry-delay",
+            "3",
+            "--connect-timeout",
+            "30",
+            "--max-time",
+            "1800",
+            "-o",
+            str(tmp),
+            url,
+        ],
+        check=False,
+        capture_output=True,
+        text=True,
+    )
+    if proc.returncode != 0:
+        tmp.unlink(missing_ok=True)
+        detail = (proc.stderr or proc.stdout or "").strip() or f"exit={proc.returncode}"
+        raise RuntimeError(detail)
+    size = tmp.stat().st_size
+    if size < min_bytes:
+        tmp.unlink(missing_ok=True)
+        raise OSError(f"download too small ({size} bytes)")
+    tmp.replace(dest)
+
+
+def _download_with_torch(url: str, dest: Path, *, min_bytes: int) -> None:
+    import torch
+
+    dest.parent.mkdir(parents=True, exist_ok=True)
+    torch.hub.download_url_to_file(url, str(dest), progress=True)
+    size = dest.stat().st_size
+    if size < min_bytes:
+        dest.unlink(missing_ok=True)
+        raise OSError(f"download too small ({size} bytes)")
+
+
+def bake_hub_checkpoint(
+    spec: HubCheckpoint,
+    *,
+    torch_home: Path,
+) -> None:
+    dest = torch_home / "hub" / "checkpoints" / spec.filename
+    if dest.is_file() and dest.stat().st_size >= spec.min_bytes:
+        print(f"already baked [{spec.label}]: {dest} ({dest.stat().st_size} bytes)")
+        return
+
+    local = _local_source(spec.filename)
+    if local is not None:
+        try:
+            _copy_local(local, dest, min_bytes=spec.min_bytes)
+            print(f"baked [{spec.label}] {dest} ({dest.stat().st_size} bytes) from local {local}")
+            return
+        except OSError as exc:
+            print(f"local copy failed for {spec.filename}: {exc}", file=sys.stderr)
+
+    errors: list[str] = []
+    for url in _candidate_urls(spec.filename):
+        try:
+            print(f"downloading [{spec.label}] {url}")
+            _download_with_curl(url, dest, min_bytes=spec.min_bytes)
+            print(f"baked [{spec.label}] {dest} ({dest.stat().st_size} bytes) from {url}")
+            return
+        except (OSError, RuntimeError) as exc:
+            errors.append(f"curl {url}: {exc}")
+            dest.unlink(missing_ok=True)
+
+    official = f"{OFFICIAL_PREFIX}/{spec.filename}"
+    try:
+        print(f"torch.hub fallback [{spec.label}]: {official}")
+        _download_with_torch(official, dest, min_bytes=spec.min_bytes)
+        print(f"baked [{spec.label}] {dest} ({dest.stat().st_size} bytes) from {official}")
+        return
+    except (OSError, RuntimeError) as exc:
+        errors.append(f"torch {official}: {exc}")
+        dest.unlink(missing_ok=True)
+
+    print(f"failed to bake {spec.label} ({spec.filename}):", file=sys.stderr)
+    for line in errors:
+        print(f"  - {line}", file=sys.stderr)
+    raise SystemExit(1)
+
+
+def warm_torchvision_hub_models() -> None:
+    """Load models so torchvision verifies hub checkpoints (no network if baked)."""
+    from torchvision.models import ResNet50_Weights, resnet50
+    from torchvision.models.video import Swin3D_T_Weights, swin3d_t
+
+    resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
+    print("resnet50 IMAGENET1K_V1 ok")
+    swin3d_t(weights=Swin3D_T_Weights.KINETICS400_V1)
+    print("swin3d_t KINETICS400_V1 ok")
+
+
+def main() -> int:
+    torch_home = Path(os.environ.get("TORCH_HOME", "/app/.cache/torch"))
+    for spec in HUB_CHECKPOINTS:
+        bake_hub_checkpoint(spec, torch_home=torch_home)
+    warm_torchvision_hub_models()
+    print("all pretrained hub weights baked")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

backend/scripts/bake_torch_hub_checkpoint.py

@@ -1,150 +1,11 @@
 #!/usr/bin/env python3
-"""Bake torchvision hub checkpoints into TORCH_HOME during Docker image build."""
+"""Backward-compatible entrypoint; bakes all hub weights (see bake_pretrained_weights.py)."""
 
 from __future__ import annotations
 
-import os
-import shutil
-import subprocess
-import sys
+import runpy
 from pathlib import Path
 
-CHECKPOINT = "swin3d_t-7615ae03.pth"
-OFFICIAL_URL = f"https://download.pytorch.org/models/{CHECKPOINT}"
-AGENTSMIRROR_URL = f"https://uv.agentsmirror.com/download.pytorch.org/models/{CHECKPOINT}"
-MIN_BYTES = 100_000_000
-
-
-def _candidate_urls() -> list[str]:
-    explicit = (os.environ.get("PYTORCH_MODELS_URL") or "").strip()
-    if explicit:
-        return [explicit]
-
-    raw = (os.environ.get("PYTORCH_MODELS_MIRROR") or "").strip().rstrip("/")
-    prefixes = [p for p in (raw, "https://uv.agentsmirror.com/download.pytorch.org") if p]
-    urls: list[str] = []
-    for prefix in prefixes:
-        urls.append(f"{prefix}/models/{CHECKPOINT}")
-    urls.extend([AGENTSMIRROR_URL, OFFICIAL_URL])
-
-    seen: set[str] = set()
-    ordered: list[str] = []
-    for url in urls:
-        if url in seen:
-            continue
-        seen.add(url)
-        ordered.append(url)
-    return ordered
-
-
-def _copy_local(src: Path, dest: Path) -> None:
-    size = src.stat().st_size
-    if size < MIN_BYTES:
-        raise OSError(f"local file too small ({size} bytes): {src}")
-    dest.parent.mkdir(parents=True, exist_ok=True)
-    shutil.copy2(src, dest)
-
-
-def _download_with_curl(url: str, dest: Path) -> None:
-    curl = shutil.which("curl")
-    if curl is None:
-        raise RuntimeError("curl not found")
-    tmp = dest.with_suffix(dest.suffix + ".part")
-    tmp.unlink(missing_ok=True)
-    proc = subprocess.run(
-        [
-            curl,
-            "-fL",
-            "--retry",
-            "5",
-            "--retry-all-errors",
-            "--retry-delay",
-            "3",
-            "--connect-timeout",
-            "30",
-            "--max-time",
-            "1800",
-            "-o",
-            str(tmp),
-            url,
-        ],
-        check=False,
-        capture_output=True,
-        text=True,
-    )
-    if proc.returncode != 0:
-        tmp.unlink(missing_ok=True)
-        detail = (proc.stderr or proc.stdout or "").strip() or f"exit={proc.returncode}"
-        raise RuntimeError(detail)
-    size = tmp.stat().st_size
-    if size < MIN_BYTES:
-        tmp.unlink(missing_ok=True)
-        raise OSError(f"download too small ({size} bytes)")
-    tmp.replace(dest)
-
-
-def _download_with_torch(url: str, dest: Path) -> None:
-    import torch
-
-    dest.parent.mkdir(parents=True, exist_ok=True)
-    torch.hub.download_url_to_file(url, str(dest), progress=True)
-    size = dest.stat().st_size
-    if size < MIN_BYTES:
-        dest.unlink(missing_ok=True)
-        raise OSError(f"download too small ({size} bytes)")
-
-
-def main() -> int:
-    torch_home = Path(os.environ.get("TORCH_HOME", "/app/.cache/torch"))
-    dest = torch_home / "hub" / "checkpoints" / CHECKPOINT
-    if dest.is_file() and dest.stat().st_size >= MIN_BYTES:
-        print(f"already baked: {dest} ({dest.stat().st_size} bytes)")
-        return 0
-
-    local_raw = (os.environ.get("PYTORCH_MODELS_LOCAL_PATH") or "").strip()
-    if local_raw:
-        local = Path(local_raw)
-        if local.is_file():
-            try:
-                _copy_local(local, dest)
-                print(f"baked {dest} ({dest.stat().st_size} bytes) from local {local}")
-                return 0
-            except OSError as exc:
-                print(f"local copy failed: {exc}", file=sys.stderr)
-
-    errors: list[str] = []
-    for url in _candidate_urls():
-        try:
-            print(f"downloading {url}")
-            _download_with_curl(url, dest)
-            print(f"baked {dest} ({dest.stat().st_size} bytes) from {url}")
-            return 0
-        except (OSError, RuntimeError) as exc:
-            errors.append(f"curl {url}: {exc}")
-            dest.unlink(missing_ok=True)
-
-    for url in (OFFICIAL_URL,):
-        try:
-            print(f"torch.hub fallback: {url}")
-            _download_with_torch(url, dest)
-            print(f"baked {dest} ({dest.stat().st_size} bytes) from {url}")
-            return 0
-        except (OSError, RuntimeError) as exc:
-            errors.append(f"torch {url}: {exc}")
-            dest.unlink(missing_ok=True)
-
-    print("failed to bake VideoSwin checkpoint:", file=sys.stderr)
-    for line in errors:
-        print(f"  - {line}", file=sys.stderr)
-    print(
-        "hint: domestic PyPI mirrors (NJU/Tsinghua/Aliyun) do not sync /models/*.pth; "
-        "pre-download once and set PYTORCH_MODELS_LOCAL_PATH, or pass "
-        "PYTORCH_MODELS_URL / --build-arg PYTORCH_MODELS_MIRROR to a mirror that hosts "
-        f"/models/{CHECKPOINT}",
-        file=sys.stderr,
-    )
-    return 1
-
-
 if __name__ == "__main__":
-    raise SystemExit(main())
+    target = Path(__file__).resolve().parent / "bake_pretrained_weights.py"
+    runpy.run_path(str(target), run_name="__main__")

backend/scripts/rebuild-api-image.sh

@@ -1,5 +1,8 @@
 #!/usr/bin/env bash
-# Clear BuildKit cache and rebuild the API image (fixes corrupted layer / unpigz errors).
+# Clear BuildKit cache and rebuild the API image.
+# Fixes export errors such as:
+#   archive/tar: invalid tar header
+#   unpigz: corrupted -- invalid deflate data
 set -euo pipefail
 cd "$(dirname "$0")/.."
 
@@ -8,12 +11,20 @@ docker builder prune -af
 docker buildx prune -af 2>/dev/null || true
 docker rmi -f backend-api:latest 2>/dev/null || true
 
+if [[ "${AGGRESSIVE_PRUNE:-0}" == "1" ]]; then
+  echo "Aggressive prune (dangling images + build cache)..."
+  docker system prune -af 2>/dev/null || true
+fi
+
 if [[ "${RESTART_DOCKER:-0}" == "1" ]]; then
   echo "Restarting Docker..."
   sudo systemctl restart docker
+  sleep 2
 fi
 
-echo "Building api image (--no-cache)..."
-docker compose build api --no-cache
+echo "Building api image (--no-cache, no attestations)..."
+export DOCKER_BUILDKIT=1
+export COMPOSE_BAKE=false
+docker compose build api --no-cache --provenance=false --sbom=false
 
 echo "Done. Recreate container: docker compose up -d --force-recreate api"

backend/tests/test_bake_pretrained_weights.py

@@ -0,0 +1,67 @@
+from __future__ import annotations
+
+import importlib.util
+from pathlib import Path
+
+import pytest
+
+
+def _load_module():
+    import sys
+
+    path = Path(__file__).resolve().parents[1] / "scripts" / "bake_pretrained_weights.py"
+    spec = importlib.util.spec_from_file_location("bake_pretrained_weights", path)
+    mod = importlib.util.module_from_spec(spec)
+    assert spec.loader is not None
+    sys.modules[spec.name] = mod
+    spec.loader.exec_module(mod)
+    return mod
+
+
+def test_candidate_urls_default(monkeypatch: pytest.MonkeyPatch) -> None:
+    mod = _load_module()
+    monkeypatch.delenv("PYTORCH_MODELS_URL", raising=False)
+    monkeypatch.delenv("PYTORCH_MODELS_MIRROR", raising=False)
+    urls = mod._candidate_urls("resnet50-0676ba61.pth")
+    assert urls[0].endswith("/models/resnet50-0676ba61.pth")
+    assert urls[-1] == f"{mod.OFFICIAL_PREFIX}/resnet50-0676ba61.pth"
+
+
+def test_candidate_urls_explicit_override_applies_to_swin_only(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    mod = _load_module()
+    monkeypatch.setenv("PYTORCH_MODELS_URL", "https://example.com/swin3d_t-7615ae03.pth")
+    assert mod._candidate_urls("swin3d_t-7615ae03.pth") == [
+        "https://example.com/swin3d_t-7615ae03.pth"
+    ]
+    assert "example.com" not in mod._candidate_urls("resnet50-0676ba61.pth")[0]
+
+
+def test_local_source_prefers_legacy_swin_path(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> None:
+    mod = _load_module()
+    swin = tmp_path / "swin3d_t-7615ae03.pth"
+    swin.write_bytes(b"x" * mod.HUB_CHECKPOINTS[0].min_bytes)
+    monkeypatch.setenv("PYTORCH_MODELS_LOCAL_PATH", str(swin))
+    assert mod._local_source("swin3d_t-7615ae03.pth") == swin
+    assert mod._local_source("resnet50-0676ba61.pth") is None
+
+
+def test_local_source_uses_local_dir(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> None:
+    mod = _load_module()
+    weights = tmp_path / "weights"
+    weights.mkdir()
+    resnet = weights / "resnet50-0676ba61.pth"
+    resnet.write_bytes(b"x")
+    monkeypatch.setenv("PYTORCH_MODELS_LOCAL_DIR", str(weights))
+    assert mod._local_source("resnet50-0676ba61.pth") == resnet
+
+
+def test_bake_skips_when_dest_already_valid(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    mod = _load_module()
+    monkeypatch.setenv("TORCH_HOME", str(tmp_path / "torch"))
+    spec = mod.HUB_CHECKPOINTS[1]
+    dest = tmp_path / "torch" / "hub" / "checkpoints" / spec.filename
+    dest.parent.mkdir(parents=True)
+    dest.write_bytes(b"x" * spec.min_bytes)
+    mod.bake_hub_checkpoint(spec, torch_home=tmp_path / "torch")

backend/tests/test_bake_torch_hub_checkpoint.py

@@ -1,28 +1,13 @@
+"""Legacy entrypoint must delegate to bake_pretrained_weights."""
+
 from __future__ import annotations
 
-import importlib.util
 from pathlib import Path
 
 
-def _load_module():
-    path = Path(__file__).resolve().parents[1] / "scripts" / "bake_torch_hub_checkpoint.py"
-    spec = importlib.util.spec_from_file_location("bake_torch_hub_checkpoint", path)
-    mod = importlib.util.module_from_spec(spec)
-    assert spec.loader is not None
-    spec.loader.exec_module(mod)
-    return mod
-
-
-def test_candidate_urls_default(monkeypatch):
-    mod = _load_module()
-    monkeypatch.delenv("PYTORCH_MODELS_URL", raising=False)
-    monkeypatch.delenv("PYTORCH_MODELS_MIRROR", raising=False)
-    urls = mod._candidate_urls()
-    assert urls[0].startswith("https://uv.agentsmirror.com/download.pytorch.org/models/")
-    assert urls[-1] == mod.OFFICIAL_URL
-
-
-def test_candidate_urls_explicit_override(monkeypatch):
-    mod = _load_module()
-    monkeypatch.setenv("PYTORCH_MODELS_URL", "https://example.com/swin3d_t-7615ae03.pth")
-    assert mod._candidate_urls() == ["https://example.com/swin3d_t-7615ae03.pth"]
+def test_bake_torch_hub_checkpoint_delegates() -> None:
+    text = (
+        Path(__file__).resolve().parents[1] / "scripts" / "bake_torch_hub_checkpoint.py"
+    ).read_text(encoding="utf-8")
+    assert "bake_pretrained_weights.py" in text
+    assert "runpy.run_path" in text

backend/weights/.gitkeep

@@ -0,0 +1,18 @@
+# Optional offline assets for Docker build (see scripts/bake_pretrained_weights.py).
+# If a file exists here, image build copies it instead of downloading.
+#
+# Torchvision hub (→ /app/.cache/torch/hub/checkpoints/):
+#   swin3d_t-7615ae03.pth   (~110MB)  VideoSwin / batch Phase1
+#   resnet50-0676ba61.pth   (~98MB)   doctor ReID backbone (ImageNet)
+#
+# MediaPipe (→ algorithm_subprocesses/5.15/doctor_identity_package/.mediapipe_models/):
+#   pose_landmarker_lite.task  (~6MB)
+#
+# wget https://download.pytorch.org/models/swin3d_t-7615ae03.pth
+# wget https://download.pytorch.org/models/resnet50-0676ba61.pth
+# curl -fsSL -o pose_landmarker_lite.task \
+#   "https://storage.googleapis.com/mediapipe-models/pose_landmarker/pose_landmarker_lite/float16/1/pose_landmarker_lite.task"
+#
+# Bundled in image via COPY (not downloaded at build):
+#   algorithm_subprocesses/5.15/weights/*.pt, actionformer_epoch_045.pth.tar
+#   doctor_identity_package/doctor_info.pth

docs/Docker部署.md

@@ -28,7 +28,12 @@ operation-room-monitor/
 - 算法子进程包：`backend/algorithm_subprocesses/5.15/`（含 `main.py` 与 `weights/`；镜像构建时会 `COPY` 进容器，勿在 `.dockerignore` 中整目录排除）
 - 标注视频中文字体：镜像内已安装 `fonts-noto-cjk`、`fonts-wqy-microhei`（供 `visualize_result_video.py` 绘制耗材标签）
 - 医生识别（MediaPipe Pose）：镜像内已安装 `libgles2`、`libegl1`、`libegl-mesa0`、`libglx-mesa0`、`libgl1-mesa-dri` 等 Mesa/GLVND 库；构建阶段会 `import mediapipe` 校验 `libGLESv2.so.2` 可用。子进程强制 CPU delegate。若仍见该错误，请 **`docker compose build --no-cache api`** 后重启（勿沿用旧 tarball 镜像）
-- 可选备用权重：`backend/app/resources/actionformer_epoch_045.pth.tar`
+- **构建时预下载的预训练资源**（`scripts/bake_pretrained_weights.py`，本地有则跳过，见 `backend/weights/.gitkeep`）：
+  - `swin3d_t-7615ae03.pth`（VideoSwin / batch Phase1）
+  - `resnet50-0676ba61.pth`（医生 ReID 骨干）
+  - `pose_landmarker_lite.task`（MediaPipe Pose）
+- 业务权重随 `algorithm_subprocesses/5.15/weights/` 打进镜像（YOLO / ActionFormer 等，非 hub 下载）
+- 可选备用：`backend/app/resources/actionformer_epoch_045.pth.tar`
 
 ---
 
@@ -58,31 +63,6 @@ docker compose down
 docker compose down -v   # 删除 PostgreSQL / MinIO 卷
 ```
 
-### 构建 API 镜像失败：`invalid tar header` / `unpigz: corrupted`
-
-`uv sync` 已成功，但在 **exporting / unpacking** 阶段报错时，通常是 **Docker 本地层缓存或存储损坏**，与 Dockerfile 无关。
-
-按顺序处理：
-
-```bash
-cd backend
-chmod +x scripts/rebuild-api-image.sh
-
-# 清缓存并重建（推荐）
-./scripts/rebuild-api-image.sh
-
-# 仍失败时：重启 Docker 后再跑
-RESTART_DOCKER=1 ./scripts/rebuild-api-image.sh
-
-# 再失败：改用旧版构建器（无 BuildKit）
-COMPOSE_DOCKER_CLI_BUILD=0 DOCKER_BUILDKIT=0 docker compose build api --no-cache
-docker compose up -d --force-recreate api
-```
-
-手动等价步骤：`docker builder prune -af` → `docker rmi -f backend-api:latest` → `docker compose build api --no-cache`。
-
-确认根分区剩余空间充足（建议 ≥ 20GB）；空间不足时大层导出也容易损坏。
-
 ### RTSP 切片在宿主机无法用 VLC 打开
 
 默认情况下 API 容器以 **root** 写入 `./logs`，切片属主为 `root:root`。普通用户虽可用 `cat` 读取，但 **Snap 版 VLC** 等沙箱应用常会报 Permission denied。