Reduce HF Spaces build time: pin CPU-only torch/torchvision; drop llama-cpp, accelerate/bitsandbytes; constrain chromadb <1.0

requirements.txt

@@ -1,6 +1,8 @@
 # Core dependencies
 gradio>=4.0.0
-torch>=2.0.0
+# Pin to a CPU-only friendly PyTorch to avoid huge CUDA deps on Spaces
+torch==2.3.1
+torchvision==0.18.1
 transformers>=4.30.0
 Pillow>=9.0.0
 
@@ -26,15 +28,15 @@ requests>=2.31.0
 
 # Llama 4 integration
 # Use Hugging Face accelerate; "accelerator" was a typo and can cause install issues
-accelerate>=0.20.0
-# bitsandbytes is only supported on Linux x86_64; avoid install on macOS/ARM
-bitsandbytes>=0.41.0; platform_system == 'Linux' and platform_machine == 'x86_64'
+# accelerate/bitsandbytes are not required in this Space; omit to reduce build size
+# accelerate>=0.20.0
+# bitsandbytes>=0.41.0; platform_system == 'Linux' and platform_machine == 'x86_64'
 sentencepiece>=0.1.99
 protobuf>=4.23.0
 
 # Vector DB and image similarity search
 chroma-hnswlib>=0.7.3
-chromadb>=0.4.18
+chromadb>=0.4.18,<1.0.0
 scipy>=1.10.0,<1.14.0; python_version < '3.12'
 scipy>=1.14.1,<1.15.0; python_version >= '3.12'
 open-clip-torch>=2.20.0
@@ -50,5 +52,7 @@ langchain-openai>=0.1.16
 langchain-community>=0.2.6
 langchain-experimental>=0.0.60
 
-# llama.cpp bindings for loading local GGUF (quantized Q4) models
-llama-cpp-python>=0.2.90
+# llama.cpp bindings for loading local GGUF (quantized Q4) models — optional.
+# Removed from default install to avoid long native builds on Spaces.
+# llama-cpp-python>=0.2.90
+