fix: use snapshot_download to avoid OOM during docker build

Dockerfile

@@ -46,14 +46,15 @@ RUN pip install --no-cache-dir --user -r requirements.txt
 #   - Whisper base              (~140 MB)
 RUN python -c "\
 from sentence_transformers import SentenceTransformer; \
-from transformers import AutoModelForCausalLM, AutoTokenizer; \
-import whisper, torch; \
+from transformers import AutoTokenizer; \
+from huggingface_hub import snapshot_download; \
+import whisper; \
 print('--- Downloading sentence-transformers (120 MB) ---'); \
 SentenceTransformer('sentence-transformers/paraphrase-MiniLM-L3-v2', cache_folder='/home/user/app/models_cache'); \
 print('--- Downloading TinyLlama tokenizer ---'); \
 AutoTokenizer.from_pretrained('TinyLlama/TinyLlama-1.1B-Chat-v1.0', cache_dir='/home/user/app/models_cache', trust_remote_code=True); \
-print('--- Downloading TinyLlama model weights (2.2 GB, please wait) ---'); \
-AutoModelForCausalLM.from_pretrained('TinyLlama/TinyLlama-1.1B-Chat-v1.0', cache_dir='/home/user/app/models_cache', torch_dtype=torch.float32, trust_remote_code=True); \
+print('--- Downloading TinyLlama model weights to disk only (2.2 GB, no RAM loading) ---'); \
+snapshot_download(repo_id='TinyLlama/TinyLlama-1.1B-Chat-v1.0', cache_dir='/home/user/app/models_cache'); \
 print('--- Downloading Whisper base (140 MB) ---'); \
 whisper.load_model('base', download_root='/home/user/app/models_cache/whisper'); \
 print('=== All models downloaded successfully ==='); \