Aayan873 commited on
Commit
8750f58
Β·
verified Β·
1 Parent(s): 9889ee0

Update models.py

Browse files
Files changed (1) hide show
  1. models.py +28 -19
models.py CHANGED
@@ -9,49 +9,59 @@ from transformers import CLIPProcessor, CLIPModel
9
  from gpt4all import GPT4All
10
  from config import MODELS_CONFIG, DEPLOYMENT_MODE
11
  from device_setup import setup_device
12
-
13
  @st.cache_resource
14
  def load_models():
15
- '''
16
- Load all required models for multimodal processing
17
- Returns: text_encoder, clip_model, clip_processor, whisper_model, ocr_reader, llm, device, has_gpu
18
- '''
19
  try:
 
 
 
 
 
 
 
 
 
 
 
 
 
20
  # Setup device
21
  device, has_gpu = setup_device()
 
22
 
23
- # Progress tracking for better UX
24
  progress = st.progress(0, text="Initializing models...")
25
 
26
- # Text embedding model
27
  progress.progress(0.1, text="Loading text encoder...")
28
  st.write("πŸ“ Loading text encoder...")
29
- text_encoder = SentenceTransformer(MODELS_CONFIG["text_encoder"], device=device)
 
 
 
 
30
 
31
- # CLIP model
32
  progress.progress(0.3, text="Loading CLIP model...")
33
  st.write("πŸ–ΌοΈ Loading CLIP model...")
34
  clip_model_name = MODELS_CONFIG["clip_model"]
35
- clip_model = CLIPModel.from_pretrained(clip_model_name).to(device)
36
- clip_processor = CLIPProcessor.from_pretrained(clip_model_name)
37
 
38
- # Whisper - use smaller model for deployment
39
  progress.progress(0.5, text="Loading Whisper...")
40
  st.write("🎀 Loading Whisper model...")
41
  whisper_size = MODELS_CONFIG["whisper_base"]
42
- whisper_model = whisper.load_model(whisper_size, device=device)
43
 
44
- # OCR Reader
45
  progress.progress(0.7, text="Loading OCR...")
46
  st.write("πŸ” Loading OCR reader...")
47
- ocr_reader = easyocr.Reader(['en'], gpu=has_gpu)
48
 
49
- # Local LLM - with better error handling
50
  progress.progress(0.9, text="Loading LLM...")
51
  st.write("🧠 Loading local LLM...")
52
-
53
  try:
54
- # Try to load the model
55
  llm = GPT4All(MODELS_CONFIG["llm_model"], allow_download=True)
56
  st.write("βœ… LLM loaded successfully!")
57
  except Exception as llm_error:
@@ -61,7 +71,6 @@ def load_models():
61
 
62
  progress.progress(1.0, text="All models loaded!")
63
  st.success("πŸŽ‰ All models loaded successfully!")
64
-
65
  return text_encoder, clip_model, clip_processor, whisper_model, ocr_reader, llm, device, has_gpu
66
 
67
  except Exception as e:
 
9
  from gpt4all import GPT4All
10
  from config import MODELS_CONFIG, DEPLOYMENT_MODE
11
  from device_setup import setup_device
 
12
  @st.cache_resource
13
  def load_models():
 
 
 
 
14
  try:
15
+ # Writable caches first (prevent /.cache writes)
16
+ import pathlib
17
+ cache_root = os.environ.get("HF_HOME", "/tmp/hf")
18
+ pathlib.Path(cache_root).mkdir(parents=True, exist_ok=True)
19
+
20
+ # Make sure all relevant env vars point to the same place
21
+ os.environ.setdefault("HF_HOME", cache_root)
22
+ os.environ.setdefault("TRANSFORMERS_CACHE", cache_root)
23
+ os.environ.setdefault("HUGGINGFACE_HUB_CACHE", cache_root)
24
+ os.environ.setdefault("HF_DATASETS_CACHE", cache_root)
25
+ os.environ.setdefault("SENTENCE_TRANSFORMERS_HOME", cache_root)
26
+ os.environ.setdefault("XDG_CACHE_HOME", "/tmp")
27
+
28
  # Setup device
29
  device, has_gpu = setup_device()
30
+ device_str = "cuda" if has_gpu else "cpu"
31
 
 
32
  progress = st.progress(0, text="Initializing models...")
33
 
34
+ # Text embeddings (explicit cache folder)
35
  progress.progress(0.1, text="Loading text encoder...")
36
  st.write("πŸ“ Loading text encoder...")
37
+ text_encoder = SentenceTransformer(
38
+ MODELS_CONFIG["text_encoder"],
39
+ device=device,
40
+ cache_folder=cache_root
41
+ )
42
 
43
+ # CLIP (route Hugging Face cache explicitly)
44
  progress.progress(0.3, text="Loading CLIP model...")
45
  st.write("πŸ–ΌοΈ Loading CLIP model...")
46
  clip_model_name = MODELS_CONFIG["clip_model"]
47
+ clip_model = CLIPModel.from_pretrained(clip_model_name, cache_dir=cache_root).to(device)
48
+ clip_processor = CLIPProcessor.from_pretrained(clip_model_name, cache_dir=cache_root)
49
 
50
+ # Whisper (force CPU on free tier unless GPU present)
51
  progress.progress(0.5, text="Loading Whisper...")
52
  st.write("🎀 Loading Whisper model...")
53
  whisper_size = MODELS_CONFIG["whisper_base"]
54
+ whisper_model = whisper.load_model(whisper_size, device=device_str)
55
 
56
+ # EasyOCR (avoid CUDA probing on CPU Spaces)
57
  progress.progress(0.7, text="Loading OCR...")
58
  st.write("πŸ” Loading OCR reader...")
59
+ ocr_reader = easyocr.Reader(['en'], gpu=False if not has_gpu else True)
60
 
61
+ # Local LLM
62
  progress.progress(0.9, text="Loading LLM...")
63
  st.write("🧠 Loading local LLM...")
 
64
  try:
 
65
  llm = GPT4All(MODELS_CONFIG["llm_model"], allow_download=True)
66
  st.write("βœ… LLM loaded successfully!")
67
  except Exception as llm_error:
 
71
 
72
  progress.progress(1.0, text="All models loaded!")
73
  st.success("πŸŽ‰ All models loaded successfully!")
 
74
  return text_encoder, clip_model, clip_processor, whisper_model, ocr_reader, llm, device, has_gpu
75
 
76
  except Exception as e: