Update gpu-space-cpu/SKILL.md inspired of https://github.com/huggingface/skills/tree/main/skills/huggingface-spaces

gpu-space-cpu/SKILL.md

@@ -5,15 +5,16 @@ description: Convert GPU HuggingFace Spaces to CPU-only for free tier deployment
 
 # GPU → CPU Space demo Conversion
 
-- Free space Tier CPU = 2 CPU cores (vCPUs) + 16GB RAM + 50GB non-persistent disk space 
+- Free space Tier CPU = 2 CPU cores (vCPUs) + 18GB RAM + 50GB non-persistent disk space
+- RAM is the binding limit (not disk): load large models with `low_cpu_mem_usage=True` (needs `accelerate`) so peak RAM stays ~1× the weights instead of ~2× during load.
 
 ## Workflow
 
 1. Grep GPU deps: `@spaces.GPU|bitsandbytes|flash-attn|triton|xformers|auto-gptq|exllama|apex|\.cuda\(|device.*cuda`
 2. Remove GPU packages, replace code: `cuda`→`cpu`, `float16`→`int8`, remove `.half()`, `.cuda()`, `device_map="auto"`
-3. Create: app.py (all logic, prioritize recent built-in gradio [v6+](https://www.gradio.app/guides/gradio-6-migration-guide)), README.md, requirements.txt
-4. Test: `pip install -r requirements.txt && python app.py` and Test API by adding mcp https://www.gradio.app/guides/building-mcp-server-with-gradio
-5. Deploy after local test passes
+3. Create: app.py (all logic, prioritize recent built-in gradio [v6+](https://www.gradio.app/guides/gradio-6-migration-guide)), README.md, requirements.txt. Handlers can run concurrently — write outputs to `tempfile.NamedTemporaryFile(...)`, never a fixed path like `output.png` (concurrent calls clobber it and files leak across users); keep module-level globals read-only.
+4. Test: `pip install -r requirements.txt && python app.py` and test the API by adding MCP (`demo.launch(mcp_server=True)`) https://www.gradio.app/guides/building-mcp-server-with-gradio — every API-exposed handler needs a docstring + type hints, otherwise it surfaces as an opaque, unusable MCP tool.
+5. Deploy after local test passes: `hf upload <id> . --repo-type space` (omit `--repo-type space` and it silently creates a *model* repo instead). On a build failure read `hf spaces logs <id> --build` and fix the **first** error, not the last.
 
 ## requirements.txt
 
@@ -21,7 +22,9 @@ description: Convert GPU HuggingFace Spaces to CPU-only for free tier deployment
 --extra-index-url https://download.pytorch.org/whl/cpu
 torch
 ```
-Never pin transitive deps. No packages.txt (ffmpeg/git/cmake pre-installed).
+- `torchvision` / `torchaudio` are **not** preinstalled — add them (unpinned) if you import them; the index above serves their `+cpu` builds too.
+- Default: never pin transitive deps. Exception: a build-time `ResolutionImpossible` against the Gradio SDK — unpin (or bump) the offender, usually `pydantic` / `uvicorn` / `huggingface_hub` / `jinja2`. `gradio[mcp]` specifically needs `uvicorn>=0.31.1` + `pydantic>=2.11.10`.
+- Use packages.txt (ffmpeg/git/cmake pre-installed) instead of docker first.
 
 ## README.md
 
@@ -32,6 +35,7 @@ emoji: X
 sdk: gradio
 sdk_version: 6.3.0
 app_file: app.py
+short_description: ...   # ≤ 60 chars; the Hub rejects longer (undocumented cap)
 ---
 ```
 
@@ -40,7 +44,7 @@ app_file: app.py
 1. ONNX Runtime (Optimum+OMP_NUM_THREADS=2) INT8 (preferred - onnxruntime fast on CPU)
 2. TorchAO
 3. torch.quantization.quantize_dynamic
-3. Onnx FP32 is faster than FP16 on CPU
+4. Onnx FP32 is faster than FP16 on CPU
 
 ## Stop