soliciteaqui commited on
Commit
3bc85de
·
verified ·
1 Parent(s): 2c29501

Upload 4 files

Browse files
Files changed (4) hide show
  1. Dockerfile +29 -36
  2. main.py +3 -12
  3. requirements.txt +4 -4
  4. start.sh +28 -0
Dockerfile CHANGED
@@ -1,42 +1,35 @@
1
- FROM nvidia/cuda:12.1.1-cudnn8-runtime-ubuntu22.04
2
 
3
- ENV DEBIAN_FRONTEND=noninteractive \
4
- PIP_NO_CACHE_DIR=1 \
5
- PYTHONUNBUFFERED=1 \
6
- TRANSFORMERS_CACHE=/home/user/.cache/huggingface \
7
- HF_HOME=/home/user/.cache/huggingface \
8
- XDG_CACHE_HOME=/home/user/.cache \
9
- WHISPER_CACHE_DIR=/home/user/.cache/whisper \
10
- WHISPER_MODEL=large-v2 \
11
- MAX_UPLOAD_BYTES=52428800 \
12
- MAX_CONCURRENT=1
13
 
14
- RUN apt-get update && apt-get install -y --no-install-recommends \
15
- ffmpeg git python3 python3-pip && rm -rf /var/lib/apt/lists/*
16
 
17
- # Non-root runtime user
18
- RUN useradd -m -u 1000 user
19
- WORKDIR /app
20
 
21
- # ---- Python deps
22
- COPY requirements.txt /app/requirements.txt
23
- RUN python3 -m pip install --upgrade pip && python3 -m pip install -r /app/requirements.txt
 
 
24
 
25
- # ---- Warm model cache to avoid cold start
26
- RUN mkdir -p /home/user/.cache/huggingface /home/user/.cache/whisper && \
27
- chown -R user:user /home/user/.cache
28
- RUN python3 - <<'PY'
29
- import os, whisper
30
- cache = os.environ.get("WHISPER_CACHE_DIR", "/home/user/.cache/whisper")
31
- model = os.environ.get("WHISPER_MODEL","large-v2")
32
- whisper.load_model(model, download_root=cache)
33
- print(f"Preloaded model {model} into {cache}")
34
- PY
35
 
36
- # ---- App code
37
- COPY main.py /app/main.py
38
- RUN chown -R user:user /app
39
- USER user
40
-
41
- EXPOSE 7860
42
- CMD ["uvicorn","main:app","--host","0.0.0.0","--port","7860","--timeout-keep-alive","75"]
 
1
+ FROM nvidia/cuda:12.1.1-cudnn8-runtime-ubuntu22.04
2
 
3
+ ENV DEBIAN_FRONTEND=noninteractive \
4
+ PIP_NO_CACHE_DIR=1 \
5
+ PYTHONUNBUFFERED=1 \
6
+ TRANSFORMERS_CACHE=/home/user/.cache/huggingface \
7
+ HF_HOME=/home/user/.cache/huggingface \
8
+ XDG_CACHE_HOME=/home/user/.cache \
9
+ WHISPER_CACHE_DIR=/home/user/.cache/whisper \
10
+ WHISPER_MODEL=tiny \
11
+ MAX_UPLOAD_BYTES=52428800 \
12
+ MAX_CONCURRENT=1
13
 
14
+ RUN apt-get update && apt-get install -y --no-install-recommends \
15
+ ffmpeg git python3 python3-pip && rm -rf /var/lib/apt/lists/*
16
 
17
+ # Non-root runtime user
18
+ RUN useradd -m -u 1000 user
19
+ WORKDIR /app
20
 
21
+ # ---- Python deps
22
+ COPY requirements.txt /app/requirements.txt
23
+ RUN python3 -m pip install --upgrade pip \
24
+ && python3 -m pip install --no-cache-dir --index-url https://download.pytorch.org/whl/cu121 torch==2.3.1+cu121 \
25
+ && python3 -m pip install --no-cache-dir -r /app/requirements.txt
26
 
27
+ # App code
28
+ COPY main.py /app/main.py
29
+ COPY start.sh /app/start.sh
30
+ RUN chmod +x /app/start.sh && chown -R user:user /app /home/user
31
+ USER user
 
 
 
 
 
32
 
33
+ EXPOSE 7860
34
+ # Warm model cache at runtime if missing, then launch
35
+ ENTRYPOINT ["/app/start.sh"]
 
 
 
 
main.py CHANGED
@@ -49,7 +49,7 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
49
  fp16 = device == "cuda"
50
  log.info(f"Loading Whisper '%s' on %s (fp16=%s)", MODEL_NAME, device, fp16)
51
 
52
- # Global model (loaded at process start, benefits from Docker preload)
53
  model = whisper.load_model(MODEL_NAME, device=device, download_root=CACHE_DIR)
54
 
55
  # ------------------------
@@ -57,7 +57,7 @@ model = whisper.load_model(MODEL_NAME, device=device, download_root=CACHE_DIR)
57
  # ------------------------
58
  app = FastAPI(
59
  title="Whisper API",
60
- version="1.2.0",
61
  openapi_url="/api/v1/stt/openapi.json",
62
  docs_url="/api/v1/stt/docs"
63
  )
@@ -86,40 +86,32 @@ def home():
86
 
87
  @app.get("/healthz")
88
  def healthz():
89
- # cheap check: returns device and cuda mem if available
90
  info = {"device": device}
91
  if device == "cuda":
92
  try:
93
- info["cuda_mem_allocated"] = torch.cuda.memory_allocated() # bytes
94
  info["cuda_mem_reserved"] = torch.cuda.memory_reserved()
95
  except Exception:
96
  pass
97
  return {"ok": True, **info}
98
 
99
- # ---- helpers ----
100
  def _safe_tmp(suffix: str = ".wav") -> str:
101
  fd, path = tempfile.mkstemp(prefix="aud_", suffix=suffix, dir="/tmp")
102
  os.close(fd)
103
  return path
104
 
105
  def _ffmpeg_norm(src: str, dst: str):
106
- # Force mono/16k PCM WAV for maximum stability
107
  cmd = ["ffmpeg","-y","-i",src,"-ac","1","-ar","16000","-f","wav",dst]
108
  r = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
109
  if r.returncode != 0:
110
  raise RuntimeError(r.stderr.decode(errors="ignore")[:400])
111
 
112
- # ------------------------
113
- # Transcription endpoint
114
- # ------------------------
115
  @app.post("/transcribe")
116
  async def transcribir_audio(file: UploadFile = File(...), _: bool = Depends(verify_token)):
117
- # Read whole body (keeps code simple, you can stream if needed)
118
  raw = await file.read()
119
  if len(raw) > MAX_BYTES:
120
  raise HTTPException(413, f"Archivo excede {MAX_BYTES/1024/1024:.1f} MB")
121
 
122
- # Use safe temp names (avoid path traversal using user filename)
123
  src_suffix = os.path.splitext(file.filename or "")[1] or ".bin"
124
  src = _safe_tmp(src_suffix)
125
  with open(src, "wb") as f:
@@ -160,7 +152,6 @@ async def transcribir_audio(file: UploadFile = File(...), _: bool = Depends(veri
160
  except Exception:
161
  pass
162
 
163
- # Local run: uvicorn main:app --host 0.0.0.0 --port 7860
164
  if __name__ == "__main__":
165
  import uvicorn
166
  uvicorn.run("main:app", host="0.0.0.0", port=int(os.getenv("PORT", "7860")))
 
49
  fp16 = device == "cuda"
50
  log.info(f"Loading Whisper '%s' on %s (fp16=%s)", MODEL_NAME, device, fp16)
51
 
52
+ # Global model (loaded at process start; cache is prewarmed by start.sh)
53
  model = whisper.load_model(MODEL_NAME, device=device, download_root=CACHE_DIR)
54
 
55
  # ------------------------
 
57
  # ------------------------
58
  app = FastAPI(
59
  title="Whisper API",
60
+ version="1.2.1",
61
  openapi_url="/api/v1/stt/openapi.json",
62
  docs_url="/api/v1/stt/docs"
63
  )
 
86
 
87
  @app.get("/healthz")
88
  def healthz():
 
89
  info = {"device": device}
90
  if device == "cuda":
91
  try:
92
+ info["cuda_mem_allocated"] = torch.cuda.memory_allocated()
93
  info["cuda_mem_reserved"] = torch.cuda.memory_reserved()
94
  except Exception:
95
  pass
96
  return {"ok": True, **info}
97
 
 
98
  def _safe_tmp(suffix: str = ".wav") -> str:
99
  fd, path = tempfile.mkstemp(prefix="aud_", suffix=suffix, dir="/tmp")
100
  os.close(fd)
101
  return path
102
 
103
  def _ffmpeg_norm(src: str, dst: str):
 
104
  cmd = ["ffmpeg","-y","-i",src,"-ac","1","-ar","16000","-f","wav",dst]
105
  r = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
106
  if r.returncode != 0:
107
  raise RuntimeError(r.stderr.decode(errors="ignore")[:400])
108
 
 
 
 
109
  @app.post("/transcribe")
110
  async def transcribir_audio(file: UploadFile = File(...), _: bool = Depends(verify_token)):
 
111
  raw = await file.read()
112
  if len(raw) > MAX_BYTES:
113
  raise HTTPException(413, f"Archivo excede {MAX_BYTES/1024/1024:.1f} MB")
114
 
 
115
  src_suffix = os.path.splitext(file.filename or "")[1] or ".bin"
116
  src = _safe_tmp(src_suffix)
117
  with open(src, "wb") as f:
 
152
  except Exception:
153
  pass
154
 
 
155
  if __name__ == "__main__":
156
  import uvicorn
157
  uvicorn.run("main:app", host="0.0.0.0", port=int(os.getenv("PORT", "7860")))
requirements.txt CHANGED
@@ -1,6 +1,6 @@
1
- fastapi==0.110.0
2
- uvicorn==0.27.1
3
  python-multipart==0.0.9
4
- torch>=2.0.0
5
  openai-whisper==20231117
6
- ffmpeg-python==0.2.0
 
 
1
+ fastapi==0.111.0
2
+ uvicorn==0.30.6
3
  python-multipart==0.0.9
 
4
  openai-whisper==20231117
5
+ # torch is installed separately for CUDA 12.1 (cu121) in Dockerfile
6
+ # ffmpeg provided by OS package in Docker image
start.sh ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env bash
2
+ set -euo pipefail
3
+
4
+ export HOME=/home/user
5
+ export TRANSFORMERS_CACHE=${TRANSFORMERS_CACHE:-/home/user/.cache/huggingface}
6
+ export HF_HOME=${HF_HOME:-/home/user/.cache/huggingface}
7
+ export XDG_CACHE_HOME=${XDG_CACHE_HOME:-/home/user/.cache}
8
+ export WHISPER_CACHE_DIR=${WHISPER_CACHE_DIR:-/home/user/.cache/whisper}
9
+ export WHISPER_MODEL=${WHISPER_MODEL:-large-v2}
10
+
11
+ mkdir -p "$TRANSFORMERS_CACHE" "$WHISPER_CACHE_DIR"
12
+
13
+ echo "[startup] Checking Whisper model cache for $WHISPER_MODEL ..."
14
+ python3 - <<'PY'
15
+ import os, glob, whisper
16
+ cache = os.environ.get("WHISPER_CACHE_DIR", "/home/user/.cache/whisper")
17
+ model = os.environ.get("WHISPER_MODEL", "large-v2")
18
+ # Simple presence check (whisper stores model files under cache dir)
19
+ existing = glob.glob(os.path.join(cache, "*"))
20
+ if existing:
21
+ print(f"[startup] Cache has {len(existing)} entries in {cache}")
22
+ else:
23
+ print(f"[startup] Preloading model {model} into {cache} (first run)...")
24
+ whisper.load_model(model, download_root=cache)
25
+ print("[startup] Preload done")
26
+ PY
27
+
28
+ exec uvicorn main:app --host 0.0.0.0 --port 7860 --timeout-keep-alive 75