fix code

Dockerfile

@@ -63,7 +63,4 @@ COPY . /app
 EXPOSE 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 \
-"]
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860", "--workers", "1", "--limit-concurrency", "8"]

app.py

@@ -9,11 +9,6 @@ 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 (

config.py

@@ -1,5 +1,6 @@
 import os
 from pathlib import Path
+import torch
 
 HF_CACHE_DIR = "/tmp/hf_cache"
 WHISPER_CACHE_DIR = "/tmp/whisper_cache"
@@ -7,6 +8,9 @@ TORCH_CACHE_DIR = "/tmp/torch_cache"
 MATPLOTLIB_CONFIG = "/tmp/matplotlib"
 UPLOAD_FOLDER = "/tmp/uploads"
 PYTHAI_NLP = "/tmp/pythainlp_data"
+OMP_NUM_THREADS="1"
+OPENBLAS_NUM_THREADS="1"
+MKL_NUM_THREADS="1"
 # Set env vars ก่อน import ใด ๆ ทั้งหมด
 os.environ["HUGGINGFACE_HUB_CACHE"] = HF_CACHE_DIR
 os.environ["HF_HOME"] = HF_CACHE_DIR
@@ -20,6 +24,12 @@ os.environ["TMP"] = "/tmp"
 os.environ["XDG_CACHE_HOME"] = "/tmp"
 os.environ["HOME"] = "/tmp/home"  # Force HOME to writable location
 os.environ["PYTHAINLP_DATA_DIR"] = PYTHAI_NLP
+os.environ["OMP_NUM_THREADS"] = OMP_NUM_THREADS
+os.environ["OPENBLAS_NUM_THREADS"] = OPENBLAS_NUM_THREADS
+os.environ["MKL_NUM_THREADS"] = MKL_NUM_THREADS
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
 # สร้าง directories ทั้งหมด
 for path in [HF_CACHE_DIR, WHISPER_CACHE_DIR, TORCH_CACHE_DIR, MATPLOTLIB_CONFIG, "/tmp/home", UPLOAD_FOLDER,PYTHAI_NLP]:
     os.makedirs(path, exist_ok=True)

models.py

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