fix(deploy): use pre-built CPU wheels to avoid OOM build

Compiling llama-cpp-python from source exceeded 16 GB RAM on CPU Basic
(exit 137 OOMKilled). Use whl/cpu and pytorch cpu indexes instead.

- Pin llama-cpp-python==0.3.28 with abetlen cpu wheel index
- Add pytorch.org/whl/cpu for lighter torch install
- Remove cmake/build-essential from packages.txt
- Update local install docs to force-reinstall cu124 after requirements

Co-authored-by: Cursor <cursoragent@cursor.com>

FIELD_NOTES.md

@@ -55,7 +55,7 @@ The custom frontend is designed for my mother: big buttons, clear language, no t
 
 **VLM attribute extraction isn't perfect.** Gemma 3 4B sometimes confuses "navy" with "black", or calls a blazer a "jacket". At 4B parameters, some accuracy trade-offs are inevitable. The structured JSON parsing with fallback regex helps catch malformed responses.
 
-**Cold start on Spaces.** Downloading a 3GB GGUF model on first launch takes time. CPU Basic compiles `llama-cpp-python` from source and downloads torch — expect 8–15 minutes on first boot. Inference is CPU-only (~30–90 s per garment).
+**Cold start on Spaces.** First build downloads pre-built CPU wheels (no compilation). Expect 3–5 minutes for pip install, then ~2–3 minutes for the first GGUF download. Inference is CPU-only (~30–90 s per garment).
 
 ## Stack
 

README.md

@@ -74,11 +74,14 @@ Set `HF_TOKEN` in Space Secrets before first use (required for model and dataset
 cd packages/wardrobe-us
 python -m venv .venv && source .venv/bin/activate
 
-# Install with CUDA 12.4 GPU acceleration (recommended):
-pip install llama-cpp-python \
-  --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124
+# Base deps (CPU wheels — same as HF Space):
 pip install -r requirements.txt
 
+# Override llama-cpp-python with CUDA 12.4 GPU wheel:
+pip install llama-cpp-python==0.3.28 \
+  --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124 \
+  --force-reinstall --no-deps
+
 # Custom minimal frontend (default — gr.Server + Alpine.js):
 python app.py
 

packages.txt

@@ -1,3 +1 @@
-cmake
-build-essential
 libopenblas-dev

requirements.txt

@@ -1,5 +1,7 @@
+--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
+--extra-index-url https://download.pytorch.org/whl/cpu
+llama-cpp-python==0.3.28
 gradio==6.17.3
-llama-cpp-python>=0.3.28
 huggingface-hub>=1.18.0
 Pillow>=12.0.0
 boto3>=1.35.0