fix docker

Dockerfile

@@ -2,36 +2,36 @@ FROM nvidia/cuda:12.3.2-cudnn9-runtime-ubuntu22.04
 
 WORKDIR /app
 
-ENV DEBIAN_FRONTEND=noninteractive
-ENV PYTHONUNBUFFERED=1
-ENV PYTHONDONTWRITEBYTECODE=1
+ENV DEBIAN_FRONTEND=noninteractive \
+    PYTHONUNBUFFERED=1 \
+    PYTHONDONTWRITEBYTECODE=1
 
-# ติดตั้ง dependencies
+# system deps
 RUN apt-get update && apt-get install -y \
-    python3.10 \
-    python3.10-venv \
-    python3-pip \
-    ffmpeg \
-    git \
-    libsndfile1 \
-    curl \
-    tzdata \
-    wget \
-    build-essential \
+    python3.10 python3.10-venv python3-pip \
+    ffmpeg git libsndfile1 curl tzdata wget build-essential \
     && ln -sf /usr/bin/python3.10 /usr/bin/python \
     && ln -sf /usr/bin/pip3 /usr/bin/pip \
     && rm -rf /var/lib/apt/lists/*
 
-# ตั้ง timezone
+# timezone
 RUN ln -fs /usr/share/zoneinfo/Asia/Bangkok /etc/localtime && \
     dpkg-reconfigure -f noninteractive tzdata
 
-# สร้าง directory cache ต่าง ๆ
+# caches
 RUN mkdir -p /tmp/hf_cache /tmp/torch_cache /tmp/matplotlib /tmp/xdg_cache /tmp/home /tmp/uploads /tmp/pythainlp_data \
     && chmod -R 777 /tmp/hf_cache /tmp/torch_cache /tmp/matplotlib /tmp/xdg_cache /tmp/home /tmp/uploads /tmp/pythainlp_data
 
-# เพิ่ม PATH สำหรับ cuDNN 9 ให้เจอ .so
-ENV HUGGINGFACE_HUB_CACHE=/tmp/hf_cache \
+# --------- IMPORTANT: limit threads to avoid libgomp spawn fail ---------
+ENV OMP_NUM_THREADS=1 \
+    OPENBLAS_NUM_THREADS=1 \
+    MKL_NUM_THREADS=1 \
+    NUMEXPR_NUM_THREADS=1 \
+    VECLIB_MAXIMUM_THREADS=1 \
+    BLIS_NUM_THREADS=1 \
+    TOKENIZERS_PARALLELISM=false \
+    PL_DISABLE_VERSION_CHECK=1 \
+    HUGGINGFACE_HUB_CACHE=/tmp/hf_cache \
     HF_HOME=/tmp/hf_cache \
     HF_CACHE=/tmp/hf_cache \
     TRANSFORMERS_CACHE=/tmp/hf_cache \
@@ -45,26 +45,25 @@ ENV HUGGINGFACE_HUB_CACHE=/tmp/hf_cache \
     LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH \
     PYTHAINLP_DATA_DIR=/tmp/pythainlp_data
 
-# ติดตั้ง Python dependencies
+# python deps
 COPY requirements.txt .
-
 RUN pip install --upgrade pip
 
-# ติดตั้ง PyTorch ที่ใช้ CUDA 12.1
+# torch cuda12.1
 RUN pip install --no-cache-dir \
-    torch==2.3.0+cu121 \
-    torchvision==0.18.0+cu121 \
-    torchaudio==2.3.0 \
+    torch==2.3.0+cu121 torchvision==0.18.0+cu121 torchaudio==2.3.0 \
     --extra-index-url https://download.pytorch.org/whl/cu121
 
-# ติดตั้ง dependencies อื่น ๆ
+# others
 RUN pip install --no-cache-dir -r requirements.txt
 
-# Copy source code เข้า container
+# app
 COPY . /app
 
-# ระบุ port ที่จะ expose
 EXPOSE 7860
 
-# เริ่มแอป
-CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]
+# run: limit workers & concurrency; set ulimit before start
+CMD ["bash","-lc", "\
+  ulimit -u 4096 || true; \
+  exec uvicorn app:app --host 0.0.0.0 --port 7860 --workers 1 --limit-concurrency 8 \
+"]

app.py

@@ -9,7 +9,11 @@ import torch
 import uvicorn
 import asyncio
 import logging
-
+os.environ.setdefault("OMP_NUM_THREADS", "1")
+os.environ.setdefault("OPENBLAS_NUM_THREADS", "1")
+os.environ.setdefault("MKL_NUM_THREADS", "1")
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
 from config import UPLOAD_FOLDER, SUPABASE_URL, SUPABASE_KEY
 from models import pipelines, models, model_lock, load_model_bundle, overlap_pipeline
 from utils import (

models.py

@@ -54,13 +54,15 @@ async def load_model_bundle():
         pipeline = Pipeline.from_pretrained(
             "pyannote/speaker-diarization-3.1",
             use_auth_token=token,
-            cache_dir=HF_CACHE_DIR
+            cache_dir=HF_CACHE_DIR,
+            revision="v3.1.1",
         ).to(device_torch)
 
         overlap_pipeline = Pipeline.from_pretrained(
                 "pyannote/overlapped-speech-detection",
                 use_auth_token=token,
-                cache_dir=HF_CACHE_DIR  # ใช้ cache เดียวกับโมเดลอื่น
+                cache_dir=HF_CACHE_DIR,
+                revision="v3.1.1",
             )
         model_fallback_chain = [PREFERRED_MODEL] + [m for m in FALLBACK_MODELS if m != PREFERRED_MODEL]