Dev

.beamignore

@@ -0,0 +1,30 @@
+# Generated by Beam SDK
+.beamignore
+.git
+.idea
+.python-version
+.vscode
+.venv
+venv
+__pycache__
+.DS_Store
+.config
+drive/MyDrive
+.coverage
+.pytest_cache
+.ipynb
+.ruff_cache
+.dockerignore
+.ipynb_checkpoints
+.env.local
+.envrc
+**/__pycache__/
+**/.pytest_cache/
+**/node_modules/
+**/.venv/
+*.pyc
+.next/
+.circleci
+.venv-py39-backup
+models
+

.claude/settings.local.json

@@ -46,7 +46,22 @@
       "Bash(node --check app.js)",
       "Bash(grep -nA8 \"content\" tailwind.config.js)",
       "Bash(grep -vE \"^ ?M \\(bn|fil|hi|id|pt-br|ur|zh|ja|ko|fr|de|es|vi|th|ar|ru|tr|it|pl|nl|fa\\)/|how-to-remove|/index\\\\.html$\")",
-      "Bash(grep -E \"\\\\.\\(py|js|css|md|json\\)$|scripts/\")"
+      "Bash(grep -E \"\\\\.\\(py|js|css|md|json\\)$|scripts/\")",
+      "Bash(python3 /private/tmp/gen_nb.py)",
+      "Bash(python3 -c \"import json,sys; nb=json.load\\(open\\('benchmark/miaocut_gpu_benchmark.ipynb'\\)\\); print\\('valid json, nbformat', nb['nbformat'], 'cells', len\\(nb['cells']\\)\\)\")",
+      "Bash(python3 -)",
+      "Bash(python3 -c \"import json,ast; nb=json.load\\(open\\('benchmark/miaocut_gpu_benchmark.ipynb'\\)\\); [ast.parse\\('\\\\n'.join\\(l for l in ''.join\\(c['source']\\).splitlines\\(\\) if not l.lstrip\\(\\).startswith\\('!'\\)\\)\\) for c in nb['cells'] if c['cell_type']=='code']; print\\('valid + code OK'\\)\")",
+      "Bash(python3 -c \"import json,ast; nb=json.load\\(open\\('benchmark/miaocut_gpu_benchmark.ipynb'\\)\\); [ast.parse\\('\\\\n'.join\\(l for l in ''.join\\(c['source']\\).splitlines\\(\\) if not l.lstrip\\(\\).startswith\\('!'\\)\\)\\) for c in nb['cells'] if c['cell_type']=='code']; print\\('notebook 已更新默认配置为 matting，校验通过'\\)\")",
+      "Bash(python3 -m py_compile pro-gpu/app.py)",
+      "Bash(node --check pro-gateway/src/index.js)",
+      "Bash(node pro-gateway/test/smoke.mjs)",
+      "WebSearch",
+      "Bash(python -m py_compile main.py)",
+      "Bash(python -)",
+      "Bash(git --no-pager diff main.py)",
+      "Bash(node pro-gateway/test/smoke-batch.mjs)",
+      "Bash(node --check batch/batch.js)",
+      "Bash(node --check pro-account.js)"
     ]
   }
 }

.github/workflows/deploy-backend-dev.yml

@@ -0,0 +1,160 @@
+name: Deploy Backend Dev
+
+on:
+  push:
+    branches: [ dev ]
+    paths:
+      - "main.py"
+      - "requirements.txt"
+      - "Dockerfile"
+      - ".github/workflows/deploy-backend-dev.yml"
+  workflow_dispatch:
+
+jobs:
+  deploy_spaces:
+    name: Push to Hugging Face Spaces (Dev)
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      - name: Install huggingface_hub
+        run: pip install --quiet "huggingface_hub>=0.24"
+
+      - name: Rolling deploy to Dev Hugging Face Spaces
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_DEV }}
+          GITHUB_SHA: ${{ github.sha }}
+        run: |
+          set -euo pipefail
+          if [ -z "$HF_TOKEN" ]; then
+            echo "HF_TOKEN_DEV secret is not configured."
+            exit 1
+          fi
+          python - <<'PY'
+          import os
+          import sys
+          import time
+          import json
+          import urllib.request
+
+          from huggingface_hub import HfApi
+
+          api = HfApi(token=os.environ["HF_TOKEN"])
+          sha_short = os.environ.get("GITHUB_SHA", "manual")[:8]
+
+          # Dev Space for deployment
+          SPACES = [
+              ("miao_cut_dev1", "wu408124546-miao-cut-dev1.hf.space"),
+          ]
+          ALLOW = ["main.py", "Dockerfile", "requirements.txt", "README.md"]
+
+          POLL_INTERVAL          = 10
+          BUILD_START_TIMEOUT    = 240
+          BUILD_FINISH_TIMEOUT   = 1500
+          READY_TIMEOUT          = 360
+          READY_STABLE_HITS      = 2
+
+          BUILDING_STAGES = {"BUILDING", "RUNNING_BUILDING", "APP_STARTING", "RUNNING_APP_STARTING"}
+          FATAL_STAGES    = {"BUILD_ERROR", "RUNTIME_ERROR", "CONFIG_ERROR", "NO_APP_FILE"}
+
+          def stage_of(repo_id):
+              try:
+                  return api.get_space_runtime(repo_id).stage or "UNKNOWN"
+              except Exception as exc:
+                  print(f"    (warn) get_space_runtime({repo_id}) failed: {exc}")
+                  return "UNKNOWN"
+
+          def wait_build_start(repo_id):
+              deadline = time.time() + BUILD_START_TIMEOUT
+              while time.time() < deadline:
+                  stage = stage_of(repo_id)
+                  if stage in FATAL_STAGES:
+                      raise RuntimeError(f"{repo_id} entered fatal stage during build start: {stage}")
+                  if stage in BUILDING_STAGES:
+                      print(f"    [{repo_id}] build started (stage={stage})")
+                      return True
+                  time.sleep(POLL_INTERVAL)
+              return False
+
+          def wait_build_finish(repo_id):
+              deadline = time.time() + BUILD_FINISH_TIMEOUT
+              last = None
+              while time.time() < deadline:
+                  stage = stage_of(repo_id)
+                  if stage != last:
+                      print(f"    [{repo_id}] stage={stage}")
+                      last = stage
+                  if stage in FATAL_STAGES:
+                      raise RuntimeError(f"{repo_id} build failed: stage={stage}")
+                  if stage == "RUNNING":
+                      return
+                  time.sleep(POLL_INTERVAL)
+              raise TimeoutError(f"{repo_id} did not return to RUNNING within {BUILD_FINISH_TIMEOUT}s")
+
+          def probe(host):
+              url = f"https://{host}/"
+              try:
+                  req = urllib.request.Request(url, headers={"User-Agent": "miaocut-rollout-dev"})
+                  with urllib.request.urlopen(req, timeout=10) as r:
+                      if r.status != 200:
+                          return (False, False)
+                      data = json.loads(r.read().decode("utf-8"))
+                      return (bool(data.get("ok")), bool(data.get("model_ready")))
+              except Exception:
+                  return (False, False)
+
+          def wait_http_ready(host):
+              deadline = time.time() + READY_TIMEOUT
+              hits = 0
+              while time.time() < deadline:
+                  ok, ready = probe(host)
+                  if ok and ready:
+                      hits += 1
+                      if hits >= READY_STABLE_HITS:
+                          print(f"    [{host}] healthy & model_ready ✅")
+                          return
+                  else:
+                      hits = 0
+                  time.sleep(POLL_INTERVAL)
+              raise TimeoutError(f"{host} not healthy/model_ready within {READY_TIMEOUT}s")
+
+          def deploy_one(space, host):
+              repo_id = f"wu408124546/{space}"
+              print(f"==> [{repo_id}] uploading {ALLOW}")
+              api.upload_folder(
+                  folder_path=".",
+                  repo_id=repo_id,
+                  repo_type="space",
+                  allow_patterns=ALLOW,
+                  commit_message=f"Deploy {sha_short} from GitHub Actions (rolling dev)",
+              )
+              if wait_build_start(repo_id):
+                  wait_build_finish(repo_id)
+              else:
+                  print(f"    [{repo_id}] no rebuild detected (files unchanged); verifying current health")
+              wait_http_ready(host)
+
+          failed = None
+          for i, (space, host) in enumerate(SPACES):
+              try:
+                  deploy_one(space, host)
+                  print(f"    [{space}] done ({i + 1}/{len(SPACES)})\n")
+              except Exception as exc:
+                  failed = (space, exc)
+                  print(f"!! [{space}] deploy failed: {exc}", file=sys.stderr)
+                  print("!! Aborting rollout — remaining Spaces keep serving the previous version.",
+                        file=sys.stderr)
+                  break
+
+          if failed:
+              sys.exit(1)
+          print("Rolling deploy complete: all 3 Dev Spaces updated with zero downtime.")
+          PY

.gitignore

@@ -49,3 +49,14 @@ models/*.bak
 .claude
 .agent
 .playwright-cli
+ads.txt
+b27e253973b44d85af293d4f007fba0c.txt
+googlebbfb4abe1fc3d9df.html
+miaocut-seo-design-doc.md
+old-photo-restoration-design.md
+quantize_birefnet.ipynb
+docs/
+.claude/*
+.agents/
+scripts/
+benchmark

Dockerfile

@@ -50,7 +50,15 @@ download_url_to_file('https://github.com/enesmsahin/simple-lama-inpainting/relea
 # ============================================================
 FROM python:3.11-slim
 
-# 系统依赖：MediaPipe/OpenCV 运行时需要 libGL/libglib。
+# 系统依赖：libgl1 / libglib2.0-0 —— MediaPipe / OpenCV 运行时依赖
+#
+# 【曾用过 jemalloc,已回退】
+# 之前尝试过 apt install libjemalloc2 + LD_PRELOAD 让 jemalloc 替换 glibc malloc,
+# 目标是降低碎片、更快归还内存给 OS。Space 端 A/B 后发现在 2 vCPU CPU-bound FP32
+# 推理场景下(BiRefNet-general-lite-768),jemalloc 的 background_thread 反而抢占
+# vCPU,导致 sharp 推理慢 ~3s(6.5s -> 9.5s median)。当前 Space 内存并不是瓶颈
+# (oom_kills=0,cgroup 用量只到 15%),不值这个速度代价,先撤。
+# 如果将来内存成为瓶颈再回来考虑 jemalloc + background_thread:false 或者 tcmalloc。
 RUN apt-get update && \
     apt-get install -y --no-install-recommends libgl1 libglib2.0-0 && \
     useradd -m -u 1000 user && \
@@ -91,6 +99,10 @@ COPY --chown=user:user models/birefnet-general-lite-768.onnx /opt/miaocut-models
 # 模型已内置到镜像；rembg 模型缺失时会自动退到可写缓存目录。
 # big-lama（去水印 inpainting 模型）由上面 builder 预置到 /opt/miaocut-models/big-lama.pt，
 # 用 LAMA_MODEL 指定后 SimpleLama 直接加载本地文件，容器重启不再去 GitHub 重下 196MB。
+#
+# MALLOC_ARENA_MAX / MALLOC_TRIM_THRESHOLD_ 是 glibc 特有的 tuning，配合 main.py 里的
+# _malloc_trim() 主动归还内存,减少 RSS 累积。之前尝试过用 jemalloc 替换 glibc,
+# 在 2 vCPU Space 上反而拖慢 sharp 推理 ~3s(见 apt install 注释),已撤回,当前继续用 glibc。
 ENV PORT=7860 \
     U2NET_HOME=/opt/miaocut-models \
     LAMA_MODEL=/opt/miaocut-models/big-lama.pt \